Best of Ian Nuttall's posts
Natter is a dictation app for agentsinsightMonologue has always been my go-to dictation app over Wispr Flow (I'm in the top 100 users with 660,000 total words)
There were a few things missing though, so I built Natter, my own version, to add them. Design is nowhere near as good, but it works much better for me
https://t.co/sbxsHS4mX5
A few differences:
1. It shows you the transcript as you are talking on the screen. I sometimes lose my train of thought so seeing it keeps me on track
2. It can stream the text into terminal windows so you don't get [Pasted Content 1161 chars]
3. If it makes a mistake I can say "Hey Natter, add a dictionary rule for clawed code to become Claude Code spelled C-L-A-U-D-E in title case"
4. It's faster than Monologue and has the technical rewrites I needed when working with AI agents and giving coding instructions
5. I can type a bit and then decide to switch to dictation and start with "lowercase" for it to continue without a capital letter mid-sentence
It's not perfect but I've done 100k words in the first 11 days! Only supports English right now but I have a draft PR for more languages if anybody wants to test it out.
Ian NuttallAug 26151188Cursor adds parallel agents and cloud workguideCursor v2.0 released
Features
- Run up to eight agents in parallel
- Add Composer agentic coding model (4x faster)
- Promote Browser to GA and embed in-editor
- Make sandboxed terminals GA and default on macOS
- Offer Cloud Agents with 99.9% reliability and instant startup
- Add Team Commands for centrally managed team rules
- Allow building plans in background with another model
- Improve prompt UI: show files as inline pills and let Agent self-gather context
Fixes
- Fix memory leaks and reduce overall memory usage
- Improve LSP performance and increase Python/TypeScript memory limits
Subscribe to Telegram updates:
https://t.co/ZgROhPjqN4Ian NuttallOct 2517861What Factory and Codex should learnguideI've been hammering Factory CLI and Codex CLI for the last 2 weeks. Here's where I think they can learn from each other:
## Factory
- Condense tool calls down. Codex tool calls just stack out of the way. It's calm, doesn't stream the terminal as much, and doesn't overwhelm you.
- Shell command response indicator. When Codex runs a command, it has a red/green indicator to know if succeeded or not. With Droid, I don't know at a glance what the outcome was.
- Option to show the full patch. Codex shows it all so it's clear to see what changed. I don't like truncated output even though there is a shortcut to show it. Saves me thinking or having to view in VS Code.
- Show remaining context. I know Factory's mission is that you don't worry about the context window. But I have my own process for logging sessions and auto-compact can mess with that.
- Send queued prompts after agent is finished. With Codex, I can stack prompt after prompt and the agent won't see it until it has finished it's run. With Factory, it sends the prompt after each tool call and often the agent gets diverted from what it was doing.
- Rewrite it in Rust. ;)
## Codex
- Always use plan tool. Even when I ask it to plan things out it very often doesn't create a proper task list with the plan tool. With Factory I always know exactly what work has been completed and what's next (needed when GPT-5.1 is such a quiet worker!)
- Add a shell denylist option. Generally I use --yolo to get around constant permission checks for things. It works well until it doesn't. In Factory I can use Auto (low) and then set an allowlist for extra commands to keep it safe but not noisy.
- Built in background commands. Factory just added this and Claude Code has it nailed too. Just send long running tasks to the background, let me view/kill if needed, and give the agent access to the logs. Saves me having multiple tabs open for dev, queue, etc.
- Add MCPs from the UI. I always have to remember where to go to manually add an MCP. Factory just does it for me in the UI, even lets me search and install popular ones.
- "Agent is done" notification. My ADHD brain needs to be pulled out of writing this X post when Codex has finished so I can continue shipping my app!
What features would you want from Factory/Codex CLI?Ian NuttallNov 2510359Gemini CLI adds policy and streaming featuresguideGemini CLI v0.11.0 released
Features
- Implement message bus and policy engine.
- Wire up UI for ASK_USER policy decisions.
- Prevent queuing of slash and shell commands.
- Add ability to run MCP prompt commands non-interactively.
- Add output-format stream-json flag for headless mode.
- Include extension name in gemini mcp list.
- Add markdown toggle (Alt+M) to switch rendered/raw view.
- Allow editing queued messages with up-arrow key.
- Remove deprecated --all-files and telemetry flags.
Fixes
- Fix --allowed-tools substring matching in non-interactive mode.
- Fix web_fetch fallback to respect Content-Type header.
- Fix Windows ripgrep detection.
- Fix shell auto-approval parsing for chained commands.
- Fix PTY spawn error handling in macOS sandbox.
Subscribe to Telegram updates:
https://t.co/ZgROhPjqN4Ian NuttallOct 2526157Automatically symlink CLAUDE.md to AGENTS.mdguideUntil Claude supports this just use the prompt below to automatically create a symlink for CLAUDE.md to AGENTS.md in your dev folder(s)
—
Set up an auto-symlink system on my Mac so I only ever maintain AGENTS.md, never CLAUDE.md.
First, ask me which directory (or directories) contain my code projects. Use that as ROOT throughout. Do not assume a path.
Requirements:
1. Create ~/bin/link-agents.sh — a one-shot script that recursively finds every AGENTS.md (case-insensitive) under ROOT (accept it as an argument, default to the path I gave you), skipping node_modules, .git, and vendor, and creates a relative symlink CLAUDE.md -> AGENTS.md in the same folder if no CLAUDE.md already exists. Run it once against ROOT.
2. Create ~/bin/watch-agents.sh — uses fswatch (install via Homebrew if missing) to recursively watch ROOT and create the symlink whenever a new AGENTS.md appears.
3. Create a launchd agent at ~/Library/LaunchAgents/com.linkagents.plist that runs the watcher at login with KeepAlive and ROOT passed as an argument, PATH including /opt/homebrew/bin. Load it.
4. Add CLAUDE.md to my global gitignore (~/.gitignore_global, set core.excludesfile if not already set) so the symlinks are never committed.
5. Make both scripts executable, then verify: create a temp folder with an AGENTS.md under ROOT, confirm CLAUDE.md appears within a few seconds, then clean up the temp folder.
Report what you created and the output of the verification step.Ian NuttallAug 262140Ilo becomes a creator business dashboardguideI just launched the new version of @ilodotso
check it out: https://t.co/ylYJYe44RX
Now fully focused on being a dashboard for business creators + small teams to track:
- Social media audience
- Web analytics
- Sales & MRR
Basically a daily biz summary for busy biz owners!
Tech stuff:
- Fully rebuilt integrations from Django → Next.js using Ralph loops
- Hosted on @Railway
- Using a @tailwindui theme modified by Opus 4.5 with @droid
- Entire backend built by @OpenAIDevs Codex CLI and Codex 5.2 high
Right now it's got the same integrations as before, but I did add a native Stripe integration with MRR calculations 👀
Lots of plans to add more integrations and features but waiting to see if anybody actually wants to pay for this first ;)Ian NuttallJan 264925A deep comparison of Codex and FactoryinsightI spent the last couple of days using Codex CLI exclusively to see how it compares to Claude Code and Factory
here is a brain dump of everything I learned:
TLDR;
I still prefer Factory because it just has more of the "must have" features I need, it's faster, the plan mode is very good and I don't really want to keep swapping my CLI tool to toggle between claude, gemini, and gpt-5.
Actual code output is probably on a par, mostly down to the fact that GPT-5 is a great model!
I rarely use Claude Code now (but do use Claude desktop app a lot)
Dump:
- codex is just an absolute workhorse. it gets shit done with no fanfare, no chattiness and just ruthlessly efficient. It really does go on and on (and on) churning through tasks
- that lack of chattiness can sometimes be a problem if you don't keep tabs on what it's doing and why. I like to understand exactly what is being built!
- the UI really speaks to my minimalist nature. it's very clean, very compact, it stacks tool calls in a neat little list and just gets out of the way
- because it's built in Rust, unlike other CLIs it NEVER has that annoying screen flicker than eventually crashes your terminal/vs code
- it sometimes has trouble running commands in the sandbox and will come back to say "i couldn't do that" and when pushed, it can ask for permission to do just that - this can slow your progress
- the ability to queue prompts as it's working is genuinely addictive and easy to get into a flow state, flicking back and forth with the app and queuing up more changes - no other CLI feels like this with queues messages, I dunno why
- the limits seem very, very generous. I did two days of solid, heavy coding using GPT-5 High and Codex High, and on the $20 only just hit the limit at ~6k lines of code over a bunch of sessions
- it is still not great at UI and maybe it is just me but it struggled with shadcn components even when giving it the llms.txt url and many docs URLs
- i don't mind too much that it struggled with UI because I do like to be able to grab links to components and pass it with some instructions on how to use it, GPT-5 really excels when you tell it exactly what to do
- in some ways it feels "too good" because there is always a little tweak or change or enhancement it suggests to improve the product. It never claims to be "production ready"
- the context window honestly feels like it goes on FOREVER. you can throw so much at it, iterate over and over and it barely goes down
- I would like the ability to ask it to use subagents to go and do a bunch of stuff and report back. I know i could do more tabs or use codex exec or whatever but so much easier to just ask the agents
- I really want background commands so it can run stuff like `npm run dev` in the background, see all outputs, and let me view/kill them from a shortcut later
- mcp servers are added to config.toml, which I actually quite like the syntax of, but it means I can't symlink or use existing mcp.json files from claude code or factory without porting them (although you can just ask the agent)
- lack of plan mode is a big thing to me. I use the factory cli plan mode a lot with gpt-5 high to plan and then codex medium/auto to implement it. I've seen PRs for this on github so it's prob coming soon.
- I'd also love to be able to quickly switch between models and thinking levels without going through the whole /model thing (just simple keyboard shortcuts)
- finally, when I hit my Plus plan limit, I upgraded to Pro, paid the $200 but it is still showing that I've hit my weekly limit and need to wait two days before I can cook again :(
(note: it just took ~10 mins and I was back in business!)Ian NuttallOct 25333282AI migrated old PHP sites to CloudflareguideBoring but practical real-life example of using AI to "actually ship things"
I have a bunch of old PHP sites that still get 200k+ visitors a month
they cost me $1300 a year on digitalocean!
I opened codex cli in 5 ghostty tabs and fired off essentially the same context to each of them:
- here’s the old project
- rebuild this as static site I can host on cloudflare
- make sure old content/seo/sitemaps are identical
- make sure you check every route and file
I added about 10 “continue” messages to keep 5.2 codex going (pro tip!) and hung out with the kids for a few hours
I had to do another few prompt cycles for each of them to solve different issues but after 2 days I was able to migrate all of them to cloudflare and cancel the vps servers
which is a $1k/yr pay rise for me! 🎉Ian NuttallJan 26591247Pair Opus planning with Codex reviewguideA really nice, simple workflow i'm enjoying right now:
1. Plan and code the feature with opus 4.5
2. Run /review with gpt 5.2 codex
Opus is so fast at iterating, and Codex swoops in and makes sure the code is rock solid before shipping!Ian NuttallJan 26416220Warp Code combines a CLI with IDE featuresguidewarp code feels like a combination of a cli agent and cursor-style ux design
it's a cli that looks like an ide because it gives you:
- editor code view
- project explorer
- one-click to view command output
- switch between agent/cli
- context/credit spend tracking
- task lists
- shared context with warp drive
there is a learning curve because it's a different workflow, but the agent was top of terminal bench until recently and i can see why
would love to see them add:
- subagents
- an agent sdk
- sidebar fonts increasing with cmd +/-
not being paid to post this, btw (feel like i have to add that these days 😉)
i have been using warp for a long while as a terminal and sometimes agent on the $15/mo planIan NuttallOct 25260161The Codex degradation report explainedguideCodex degradation report TLDR:
- The investigation found no single root cause; rather a mix of behavior shifts and small bugs, with several fixes already shipped and more on the way.
- Older hardware underperformed in evals and was removed, and improved load balancing is rolling out to reduce latency.
- Compaction was used more often, and repeated compactions hurt accuracy; recursive summaries were blocked and a warning was added to keep chats focused.
- The `apply_patch` flow could delete and then recreate files after failed diffs, which is risky if interrupted; mitigations are in progress and a longer‑term model fix is planned.
- There was no fleet‑wide latency regression, but some sessions escalated retries and timeouts inefficiently; handling of long‑running and interactive tasks is being improved.
- A constrained‑sampling bug produced out‑of‑distribution tokens, including occasional mid‑sentence language switches in under 0.25% of sessions; a fix is rolling out.
- The Responses API added extra newlines around tool definitions; this had no measurable performance impact.
- Evals show CLI 0.45 maintained quality while reducing tokens by roughly 10%; neither `--search` nor recent prompt changes caused regressions.
- Infra analysis found low auth‑cache hit rates adding about 50 ms to tail latency; working‑directory error rates were unchanged.
- The guidance is to prefer minimalist toolsets and shorter, targeted chats, and a dedicated real‑world performance team is being formed and hiring.Ian NuttallNov 25458158Claude Code should support .agents/skillsguideThis is a petition to get @AnthropicAI / @claudeai code to support the .agents/skills standard, just as these have:
- amp
- codex cli
- factory/droid
- gemini cli
- gh copilot
- kimi cli
- opencode
- vscode
- etc
Like this post if you want them to support .agents/skills ✊Ian NuttallFeb 262.1K116Why public backlink buying carries serious riskinsightThis blew up a bit more than I expected so I thought I'd add a bit of context and nuance to my point.
There is a real divide in the replies between people who have worked in SEO for a long time and those who haven't.
Those who have never experienced your entire business being removed/invisible from Google maybe don't see the risk vs reward the same way as those who have.
If you do enough outreach for backlinks you'll find out that many sites sell them. There are spreadsheets where you can literally pick and choose from a variety of niches and prices.
So you have a service like Outrank that creates a menu like this and sells those links to its customers. The demand for this is clearly very high based on the responses from Tibo's launch.
When demand is high and you already have a network of sites willing to sell, it's fair to assume that those selling sites will link out to more and more Outrank customers.
Easy money for the seller to just add another post or insert a link into an existing one. Pretty soon you have a very clear pattern that this site is selling links.
And so you look at the other sites that link to the same URLs. It's trivial to correlate these and find all of the sellers and all of the buyers for me - for Google 1000x so.
This is what you call a PBN (private blog network) and it's one of the fastest ways to get manual penalties - for both seller and buyer.
At best: the links will do nothing and Google will simply ignore them.
At worst: your site will disappear from Google overnight.
The argument against me is that you could just submit your competitors and take them out of Google. It trivialises the very real risk that you take whenever you buy links. The risk is significantly higher when it's promoted in a public forum like X.
Manual actions for unnatural links aren't as common as they once were because the spam algorithm detects them and ignores them but that risk profile changes when you start using a very public service (ironically getting more visibility because I raised concerns with it).
@Marie_Haynes has been the expert in this field for many years and she said this in 2024, which is still relevant today:
--
"I could just point links at my competitor then and get them penalized."
The thing is, it's not the links themselves that get sites penalized. It's the patterns that are consistent with building links to manipulate rankings. Google has many years of experience in identifying patterns of Link manipulation.
--
Many SEOs build their own PBNs to control the volume of links, content quality and relevance of them to avoid patterns but it is still a risk assessment they make. It's just done by people who fully understand the consequences because they have experienced it before.
Tibo replied and said that I was hating, but that was not my intention, although I can see how the original post was a little sensationalist and absolute.
Is it a 100% guarantee that your site specifically will get penalized? No. There is significant risk to doing it and my bet is that eventually (and maybe sooner than you think) you will lose Google traffic.
It may not be all-or-nothing. It may be that every algorithm update brings a new drop in traffic over time until the site no longer ranks for any of the things that matter - maybe even your own brand name.
This is all just my opinion based on my experiences of selling and buying many links through many eras of Google algorithm updates.
Use Outrank, don't use Outrank, but at least know and accept the risks if you do. Their landing page does not explain the risks outside of a tiny FAQ that barely scratches the surface.
Finally: make sure you are building an email list, a reddit/Facebook/Discord community and experimenting with paid ads as well.
Nobody should rely on search engines.Ian NuttallSep 2617188Why I hate being a founderinsightConfession: I hate being "a founder"
I love building stuff and learning by doing but when it becomes features for others, complicated devops, feature creep or competing with dozens of alternatives I start looking for the kill switch!
I've talked before about struggling with knowing what to work on despite spending so much time using conductor, codex, claude code, cursor/grok bot etc
https://t.co/7q2prf4CyX
I took a bunch of time off and went back to building small products, sites and tools outside of tech and AI.
those consumer apps make 5x more than keep.md, with almost no maintenance, no costs, and a fraction of the dev build time.
Maybe that's enough? Just do more of that? I can't really share those here because they aren't interesting to a tech audience and there's no moat now - I struck gold with SEO and the market so sharing would only harm me
They're also not long term revenue streams so you need a constant feed of new apps
Feels a bit isolating to just do that too which is why I keep coming back to "higher purpose" and trying to build and do things in public to help/share with others who love this "holy shit you can build anything" era we're in
I just want to do things my way (which is usually the opposite of what others do), build cool stuff, help others, and somehow make good money from it.
I have a few options that are not mutually exclusive and ordered by likelihood to succeed:
1. Keep quietly building consumer apps and grow them with SEO/AEO
2. Take my love for small, single utility dev/productivity tools and sell them at <$50 one time purchase
3. Persist with my newsletter idea and try to grow a sustainable audience
4. Partner up with a dev or team on a good product who want to scale with SEO and AI search.
5. Get sponsored by somebody to just build free stuff in public using their tool (devrel by proxy)
I'm unemployable I think so getting a proper job would not be appealing unless it was fully async, autonomous, and remote and most big cos will never want to do that.
I'm looking for the tough love @Shpigford got when he posted recently about being on the cusp of something but it never quite becoming viable.
https://t.co/7eHpNUH7Y8Ian NuttallAug 268777Clone repositories to run agents in parallelguidemy adsense accounts have made over $1m dollars (seriously) in lifetime earnings
most are dying/dead because of ai. so how do you stay sustainable in an agent-first economy?
- people love free shit
- but they hate (most) ads
- and they rarely pay for info
a standard for inserting ads into mcp, clis, markdown, anything agents use to get context COULD work
but they need to be aware of it up front. we’re used to media sites plastered with dozens of ads but it’s so early with context for agents and very much “pre-adsense” era
also it has to be a proper standard otherwise you run into prompt injection territory and lose the users trust
@AmpCode works because they control the platform and very clearly launched it as a free option with curated ads ❤️
been thinking about this for a long time now, and really hope we find a way for content creators to get paid and for people getting value out of free stuff to not kick off about very focused, relevant, and non-intrusive ads
(also I would really like to replace my adsense earnings 😂)
https://t.co/nnJjxnyvVvIan NuttallJan 264854How Cursor changed my building careerguideCursor changed everything for me.
I was copy-pasting from Claude 3.5 on the web back and forth to VS code like a madman and then this tool came along and instantly gave me the "holy shit" moment...
https://t.co/0qklYOZL18
I started sharing tips on how I was using it, built a video library (RIP Cursorcasts) and grew from 15k to 60k followers, had 3M views on Cursor posts and 20K likes
With it, I shipped 2 software products in public to $10k MRR (sold them for $280k), a paid course, dozens of SEO-powered sites and apps, actually learned how to code and not just vibe code, and made a bunch of internet friends along the way.
I'm addicted to building with AI and it's all thanks to Cursor.
Congrats team. You deserve it. 🫡Ian NuttallAug 2612521Keep MCP documents a Mac Mini setupguideOne of my favourite things about the Keep MCP server is when I am doing something interesting, I tell the agent:
"Create and update a Keep note to document this so I can write a blog post about it later"
Right now it's syncing my MacBook Pro and Mac Mini apps, documents, dotfiles, CLIs, skills, repos etc to run coding agents and GitHub action runners remotely 24/7.
It's matching my wallpaper, dock icons, installing Chrome extensions, Tailscale, Jump Desktop.
The Mini is upstairs and I'm just chatting it out with Codex. 🤯Ian NuttallSep 261414Codex struggled while Opus fixed the workguideI thought I was going mad, but 5.2 codex isn't running well today
I had to switch to opus 4.5,had issues like:
- writing a lot of wrong code to create nextjs sitemaps even after being given the exact docs on how to do it (opus deleted it all and fixed it in one shot)
- changing things not to do with what I asked (like text/content)
- spending 10 minutes trying to figure out how to copy a tailwind 4 typography examplr from another project (opus did it in <1m)Ian NuttallJan 266313Comparing Opus Codex and Claude CodeguideI'm refactoring my Hono app that powers the email list at https://t.co/sC6hbJxqJ9 so I can open source it
Gave the same prompt to Codex, Claude Code, and Droid (CC and Droid both using Opus 4.5)
Droid
had the best overall refactor report by far. It planned out a nice content structure for drafts, scheduled, sent
Proposed a full "shape" for a CLI to make everything agent-first and how to use the Hetzner CLI to set up a VPS
Codex (5.2 codex high)
focused much more on file-by-file analysis and found 4 security issues Droid didn't, and 3 reliability/perf issues
also found a bunch of missing files, hardcoded personal things that I need to fix to make it OSS
no code examples at all!
Claude Code
found a couple of potential security issues the others didn't find and had a lot more code examples (Droid had some, Codex had none)
didn't go as deep into the agent-first angle to let the agent set up a VPS and the CLI it recommended wasn't as good as the Droid version imo
boring tldr; I need both Opus 4.5 and 5.2 codex 😰Ian NuttallJan 262811A creator sponsorship pivot for Iloinsighta b2b idea for a full rebuild of @ilodotso from the ground up...
back before the X API destroyed a bunch of businesses, ilo was around $4k MRR, it's now somewhere ~$300
but doing some churn/data with claude code shows me that 70% of monthly revenue and 92% of annual is at risk of churn!
98% of previous lifetime users are not active on the platform any more.
I personally have used it to track my growth on X but I vibe coded my own link-in-bio tool instead of using ilo.
claude reckons it isn't currently solving a real pain point and is more of a "nice to have" and I kinda agree.
so here's my idea of a pivot (let me know below what you think):
The content creator platform to land sponsorship offers for your audience
Launch a professional media kit backed by real-time stats. Get paid directly. No middleman fees.
(ilo = i land offers 😉)
by focusing on content creators who can streamline their business and make money from the product, my theory is it will be a lot more "sticky" and used more frequently
v1: $15/mo or $150/yr for a landing page with real-time stats and stripe/polar/lemonsqueezy/paddle payouts
v2: add a marketplace where brands can find your products and maybe charge a small <5% fee for connecting sponsor to creator
with programmatic seo and aeo/geo i could probably rank pretty well for both brand pages and marketplace pages too 🤔
feel free to pick this apart and tell me why it's a bad idea. maybe time to vibe code an mvp to see how possible this is?Ian NuttallSep 252110Slow down to avoid coding-agent slopguideMy hot take:
Slow down.
- Chat in plan mode.
- Build one feature at a time.
- Review the code and give feedback.
- Log session summaries.
- Give previous summary to /new agent.
- Be very selective with MCPs.
- Max of 2 agents (1 frontend, 1 backend) at a time.
Slow > slop.Ian NuttallNov 253.3K1.6KA cdx shortcut configures Codex CLIguideCodex CLI pro tip
Add this updated function to bashrc/zshrc & run cdx instead of codex to get:
- cdx update to update codex
- gpt-5 codex model
- web search enabled
- full auto mode (read, edit, run commands)
- summaries of model thinking
(function in image alt text) https://t.co/qLlZGajY7BIan NuttallSep 251.1K1.4KAn SEO skill for coding agentsinsightOne of my fave apps I worked on recently is https://t.co/yYlWqhPg9d
"The only SEO skill your agent needs."
# Run this
npm i -g seo
# Then this
seo start
My agents use it with GSC + Ahrefs data to fix issues, find keywords, check AI readiness and write better ranking pages.
I built this over a few months while ranking a bunch of brand new programmatic SEO sites (the pSEO reports are very good for this) and eventually tidied it up into a nice CLI.
That `seo` package name... 🤤
Give it a try and lmk what you think. I don't accept PRs but if you find a problem you can create an issue on GH (it's fully OSS) with a prompt and I'll give it to an agent to fix!Ian NuttallAug 26423854Use Claude Skills with any AI agenttipThis tool lets you use Claude Skills with any AI agent (Codex, Factory, Cursor etc)
- Same format and folders, just CLI instead of tools
- Access marketplace skills through GitHub
- Uses progressive disclosure (skills on demand)
github.com/numman-ali/ope… https://t.co/Hvp2IElxmDIan NuttallOct 25645730Session logs preserve context between agentsguideI log my sessions with /end-session to log:
- overview of what we did
- files we created/deleted/modified
- why we made decisions
- what worked/didn’t work
- steps for next session
then I run /start-session with new agent and it reads the last log to get the context it needsIan NuttallDec 25489509Serve Markdown versions to AI clientsguidealso:
1. add <link rel="alternate" type="text/markdown"> in the html specifying the exact markdown version
2. add a link header for your llms.txt file: Link: </llms.txt>; rel="describedby"; type="text/markdown"Ian NuttallAug 26157315What Claude Code needs to win users backguideA list of things I think Claude Code could do to win back people switching to Codex CLI:
- open source Claude Code
- reduce sycophancy/make it less verbose (or add option for that)
- more transparency about how/why the model degrades
- fix tui flashing bug! PLEASE
- improve model hallucinations like GPT-5 has
- better thinking for removing files/lines of code to prevent accidental deletions
- less boilerplate or pseudo implementations (break it into working chunks if needed)
- ability to change/remove/reduce the system reminder prompts
- file based session auto-compact with much more detail on the conversation for future reference
what would you want to see improve to make CC work better for you?Ian NuttallSep 25793291Gemini designs better while Codex codes betterguideEarly thoughts on Gemini 3 Pro:
- Very fast, very smart
- Does not have the attention to detail of GPT-5.1
- Incredible frontend designer (which I need!)
- Not as good at code as Codex 5.1
- FUCKING LOVES TO COMMIT SHIT
My TLDR;
Backend: Codex 5.1
Frontend: Gemini 3Ian NuttallNov 251.3K285Find and install any agent skillguideUse this command to find and add any skill to your AI agents:
npx playbooks add skill
- Enter for fast search, Tab for semantic
- Returns official, trusted sources first https://t.co/pavqvZCE7wIan NuttallJan 26175230Build local tools around your personal agentguideseeing a lot of people say their openclaw dies/doesn’t respond often
PRO TIP: build your own local-first api/cli tools for it and keep your personal agent “dumber”
I have:
- my own cron system
- an event system to wake agent up from scripts
- https://t.co/JiFU50aLoJ to fetch x bookmarks
- my own agent-first email marketing cli - a trained blog/email draft writer
- a google search console cli for improving my seo
- etc
build your own tools don’t ask the agent to do everything!Ian NuttallFeb 26134222
Ian NuttallAug 26151188Cursor adds parallel agents and cloud workguideCursor v2.0 released
Features
- Run up to eight agents in parallel
- Add Composer agentic coding model (4x faster)
- Promote Browser to GA and embed in-editor
- Make sandboxed terminals GA and default on macOS
- Offer Cloud Agents with 99.9% reliability and instant startup
- Add Team Commands for centrally managed team rules
- Allow building plans in background with another model
- Improve prompt UI: show files as inline pills and let Agent self-gather context
Fixes
- Fix memory leaks and reduce overall memory usage
- Improve LSP performance and increase Python/TypeScript memory limits
Subscribe to Telegram updates:
https://t.co/ZgROhPjqN4Ian NuttallOct 2517861What Factory and Codex should learnguideI've been hammering Factory CLI and Codex CLI for the last 2 weeks. Here's where I think they can learn from each other:
## Factory
- Condense tool calls down. Codex tool calls just stack out of the way. It's calm, doesn't stream the terminal as much, and doesn't overwhelm you.
- Shell command response indicator. When Codex runs a command, it has a red/green indicator to know if succeeded or not. With Droid, I don't know at a glance what the outcome was.
- Option to show the full patch. Codex shows it all so it's clear to see what changed. I don't like truncated output even though there is a shortcut to show it. Saves me thinking or having to view in VS Code.
- Show remaining context. I know Factory's mission is that you don't worry about the context window. But I have my own process for logging sessions and auto-compact can mess with that.
- Send queued prompts after agent is finished. With Codex, I can stack prompt after prompt and the agent won't see it until it has finished it's run. With Factory, it sends the prompt after each tool call and often the agent gets diverted from what it was doing.
- Rewrite it in Rust. ;)
## Codex
- Always use plan tool. Even when I ask it to plan things out it very often doesn't create a proper task list with the plan tool. With Factory I always know exactly what work has been completed and what's next (needed when GPT-5.1 is such a quiet worker!)
- Add a shell denylist option. Generally I use --yolo to get around constant permission checks for things. It works well until it doesn't. In Factory I can use Auto (low) and then set an allowlist for extra commands to keep it safe but not noisy.
- Built in background commands. Factory just added this and Claude Code has it nailed too. Just send long running tasks to the background, let me view/kill if needed, and give the agent access to the logs. Saves me having multiple tabs open for dev, queue, etc.
- Add MCPs from the UI. I always have to remember where to go to manually add an MCP. Factory just does it for me in the UI, even lets me search and install popular ones.
- "Agent is done" notification. My ADHD brain needs to be pulled out of writing this X post when Codex has finished so I can continue shipping my app!
What features would you want from Factory/Codex CLI?Ian NuttallNov 2510359Gemini CLI adds policy and streaming featuresguideGemini CLI v0.11.0 released
Features
- Implement message bus and policy engine.
- Wire up UI for ASK_USER policy decisions.
- Prevent queuing of slash and shell commands.
- Add ability to run MCP prompt commands non-interactively.
- Add output-format stream-json flag for headless mode.
- Include extension name in gemini mcp list.
- Add markdown toggle (Alt+M) to switch rendered/raw view.
- Allow editing queued messages with up-arrow key.
- Remove deprecated --all-files and telemetry flags.
Fixes
- Fix --allowed-tools substring matching in non-interactive mode.
- Fix web_fetch fallback to respect Content-Type header.
- Fix Windows ripgrep detection.
- Fix shell auto-approval parsing for chained commands.
- Fix PTY spawn error handling in macOS sandbox.
Subscribe to Telegram updates:
https://t.co/ZgROhPjqN4Ian NuttallOct 2526157Automatically symlink CLAUDE.md to AGENTS.mdguideUntil Claude supports this just use the prompt below to automatically create a symlink for CLAUDE.md to AGENTS.md in your dev folder(s)
—
Set up an auto-symlink system on my Mac so I only ever maintain AGENTS.md, never CLAUDE.md.
First, ask me which directory (or directories) contain my code projects. Use that as ROOT throughout. Do not assume a path.
Requirements:
1. Create ~/bin/link-agents.sh — a one-shot script that recursively finds every AGENTS.md (case-insensitive) under ROOT (accept it as an argument, default to the path I gave you), skipping node_modules, .git, and vendor, and creates a relative symlink CLAUDE.md -> AGENTS.md in the same folder if no CLAUDE.md already exists. Run it once against ROOT.
2. Create ~/bin/watch-agents.sh — uses fswatch (install via Homebrew if missing) to recursively watch ROOT and create the symlink whenever a new AGENTS.md appears.
3. Create a launchd agent at ~/Library/LaunchAgents/com.linkagents.plist that runs the watcher at login with KeepAlive and ROOT passed as an argument, PATH including /opt/homebrew/bin. Load it.
4. Add CLAUDE.md to my global gitignore (~/.gitignore_global, set core.excludesfile if not already set) so the symlinks are never committed.
5. Make both scripts executable, then verify: create a temp folder with an AGENTS.md under ROOT, confirm CLAUDE.md appears within a few seconds, then clean up the temp folder.
Report what you created and the output of the verification step.Ian NuttallAug 262140Ilo becomes a creator business dashboardguideI just launched the new version of @ilodotso
check it out: https://t.co/ylYJYe44RX
Now fully focused on being a dashboard for business creators + small teams to track:
- Social media audience
- Web analytics
- Sales & MRR
Basically a daily biz summary for busy biz owners!
Tech stuff:
- Fully rebuilt integrations from Django → Next.js using Ralph loops
- Hosted on @Railway
- Using a @tailwindui theme modified by Opus 4.5 with @droid
- Entire backend built by @OpenAIDevs Codex CLI and Codex 5.2 high
Right now it's got the same integrations as before, but I did add a native Stripe integration with MRR calculations 👀
Lots of plans to add more integrations and features but waiting to see if anybody actually wants to pay for this first ;)Ian NuttallJan 264925A deep comparison of Codex and FactoryinsightI spent the last couple of days using Codex CLI exclusively to see how it compares to Claude Code and Factory
here is a brain dump of everything I learned:
TLDR;
I still prefer Factory because it just has more of the "must have" features I need, it's faster, the plan mode is very good and I don't really want to keep swapping my CLI tool to toggle between claude, gemini, and gpt-5.
Actual code output is probably on a par, mostly down to the fact that GPT-5 is a great model!
I rarely use Claude Code now (but do use Claude desktop app a lot)
Dump:
- codex is just an absolute workhorse. it gets shit done with no fanfare, no chattiness and just ruthlessly efficient. It really does go on and on (and on) churning through tasks
- that lack of chattiness can sometimes be a problem if you don't keep tabs on what it's doing and why. I like to understand exactly what is being built!
- the UI really speaks to my minimalist nature. it's very clean, very compact, it stacks tool calls in a neat little list and just gets out of the way
- because it's built in Rust, unlike other CLIs it NEVER has that annoying screen flicker than eventually crashes your terminal/vs code
- it sometimes has trouble running commands in the sandbox and will come back to say "i couldn't do that" and when pushed, it can ask for permission to do just that - this can slow your progress
- the ability to queue prompts as it's working is genuinely addictive and easy to get into a flow state, flicking back and forth with the app and queuing up more changes - no other CLI feels like this with queues messages, I dunno why
- the limits seem very, very generous. I did two days of solid, heavy coding using GPT-5 High and Codex High, and on the $20 only just hit the limit at ~6k lines of code over a bunch of sessions
- it is still not great at UI and maybe it is just me but it struggled with shadcn components even when giving it the llms.txt url and many docs URLs
- i don't mind too much that it struggled with UI because I do like to be able to grab links to components and pass it with some instructions on how to use it, GPT-5 really excels when you tell it exactly what to do
- in some ways it feels "too good" because there is always a little tweak or change or enhancement it suggests to improve the product. It never claims to be "production ready"
- the context window honestly feels like it goes on FOREVER. you can throw so much at it, iterate over and over and it barely goes down
- I would like the ability to ask it to use subagents to go and do a bunch of stuff and report back. I know i could do more tabs or use codex exec or whatever but so much easier to just ask the agents
- I really want background commands so it can run stuff like `npm run dev` in the background, see all outputs, and let me view/kill them from a shortcut later
- mcp servers are added to config.toml, which I actually quite like the syntax of, but it means I can't symlink or use existing mcp.json files from claude code or factory without porting them (although you can just ask the agent)
- lack of plan mode is a big thing to me. I use the factory cli plan mode a lot with gpt-5 high to plan and then codex medium/auto to implement it. I've seen PRs for this on github so it's prob coming soon.
- I'd also love to be able to quickly switch between models and thinking levels without going through the whole /model thing (just simple keyboard shortcuts)
- finally, when I hit my Plus plan limit, I upgraded to Pro, paid the $200 but it is still showing that I've hit my weekly limit and need to wait two days before I can cook again :(
(note: it just took ~10 mins and I was back in business!)Ian NuttallOct 25333282AI migrated old PHP sites to CloudflareguideBoring but practical real-life example of using AI to "actually ship things"
I have a bunch of old PHP sites that still get 200k+ visitors a month
they cost me $1300 a year on digitalocean!
I opened codex cli in 5 ghostty tabs and fired off essentially the same context to each of them:
- here’s the old project
- rebuild this as static site I can host on cloudflare
- make sure old content/seo/sitemaps are identical
- make sure you check every route and file
I added about 10 “continue” messages to keep 5.2 codex going (pro tip!) and hung out with the kids for a few hours
I had to do another few prompt cycles for each of them to solve different issues but after 2 days I was able to migrate all of them to cloudflare and cancel the vps servers
which is a $1k/yr pay rise for me! 🎉Ian NuttallJan 26591247Pair Opus planning with Codex reviewguideA really nice, simple workflow i'm enjoying right now:
1. Plan and code the feature with opus 4.5
2. Run /review with gpt 5.2 codex
Opus is so fast at iterating, and Codex swoops in and makes sure the code is rock solid before shipping!Ian NuttallJan 26416220Warp Code combines a CLI with IDE featuresguidewarp code feels like a combination of a cli agent and cursor-style ux design
it's a cli that looks like an ide because it gives you:
- editor code view
- project explorer
- one-click to view command output
- switch between agent/cli
- context/credit spend tracking
- task lists
- shared context with warp drive
there is a learning curve because it's a different workflow, but the agent was top of terminal bench until recently and i can see why
would love to see them add:
- subagents
- an agent sdk
- sidebar fonts increasing with cmd +/-
not being paid to post this, btw (feel like i have to add that these days 😉)
i have been using warp for a long while as a terminal and sometimes agent on the $15/mo planIan NuttallOct 25260161The Codex degradation report explainedguideCodex degradation report TLDR:
- The investigation found no single root cause; rather a mix of behavior shifts and small bugs, with several fixes already shipped and more on the way.
- Older hardware underperformed in evals and was removed, and improved load balancing is rolling out to reduce latency.
- Compaction was used more often, and repeated compactions hurt accuracy; recursive summaries were blocked and a warning was added to keep chats focused.
- The `apply_patch` flow could delete and then recreate files after failed diffs, which is risky if interrupted; mitigations are in progress and a longer‑term model fix is planned.
- There was no fleet‑wide latency regression, but some sessions escalated retries and timeouts inefficiently; handling of long‑running and interactive tasks is being improved.
- A constrained‑sampling bug produced out‑of‑distribution tokens, including occasional mid‑sentence language switches in under 0.25% of sessions; a fix is rolling out.
- The Responses API added extra newlines around tool definitions; this had no measurable performance impact.
- Evals show CLI 0.45 maintained quality while reducing tokens by roughly 10%; neither `--search` nor recent prompt changes caused regressions.
- Infra analysis found low auth‑cache hit rates adding about 50 ms to tail latency; working‑directory error rates were unchanged.
- The guidance is to prefer minimalist toolsets and shorter, targeted chats, and a dedicated real‑world performance team is being formed and hiring.Ian NuttallNov 25458158Claude Code should support .agents/skillsguideThis is a petition to get @AnthropicAI / @claudeai code to support the .agents/skills standard, just as these have:
- amp
- codex cli
- factory/droid
- gemini cli
- gh copilot
- kimi cli
- opencode
- vscode
- etc
Like this post if you want them to support .agents/skills ✊Ian NuttallFeb 262.1K116Why public backlink buying carries serious riskinsightThis blew up a bit more than I expected so I thought I'd add a bit of context and nuance to my point.
There is a real divide in the replies between people who have worked in SEO for a long time and those who haven't.
Those who have never experienced your entire business being removed/invisible from Google maybe don't see the risk vs reward the same way as those who have.
If you do enough outreach for backlinks you'll find out that many sites sell them. There are spreadsheets where you can literally pick and choose from a variety of niches and prices.
So you have a service like Outrank that creates a menu like this and sells those links to its customers. The demand for this is clearly very high based on the responses from Tibo's launch.
When demand is high and you already have a network of sites willing to sell, it's fair to assume that those selling sites will link out to more and more Outrank customers.
Easy money for the seller to just add another post or insert a link into an existing one. Pretty soon you have a very clear pattern that this site is selling links.
And so you look at the other sites that link to the same URLs. It's trivial to correlate these and find all of the sellers and all of the buyers for me - for Google 1000x so.
This is what you call a PBN (private blog network) and it's one of the fastest ways to get manual penalties - for both seller and buyer.
At best: the links will do nothing and Google will simply ignore them.
At worst: your site will disappear from Google overnight.
The argument against me is that you could just submit your competitors and take them out of Google. It trivialises the very real risk that you take whenever you buy links. The risk is significantly higher when it's promoted in a public forum like X.
Manual actions for unnatural links aren't as common as they once were because the spam algorithm detects them and ignores them but that risk profile changes when you start using a very public service (ironically getting more visibility because I raised concerns with it).
@Marie_Haynes has been the expert in this field for many years and she said this in 2024, which is still relevant today:
--
"I could just point links at my competitor then and get them penalized."
The thing is, it's not the links themselves that get sites penalized. It's the patterns that are consistent with building links to manipulate rankings. Google has many years of experience in identifying patterns of Link manipulation.
--
Many SEOs build their own PBNs to control the volume of links, content quality and relevance of them to avoid patterns but it is still a risk assessment they make. It's just done by people who fully understand the consequences because they have experienced it before.
Tibo replied and said that I was hating, but that was not my intention, although I can see how the original post was a little sensationalist and absolute.
Is it a 100% guarantee that your site specifically will get penalized? No. There is significant risk to doing it and my bet is that eventually (and maybe sooner than you think) you will lose Google traffic.
It may not be all-or-nothing. It may be that every algorithm update brings a new drop in traffic over time until the site no longer ranks for any of the things that matter - maybe even your own brand name.
This is all just my opinion based on my experiences of selling and buying many links through many eras of Google algorithm updates.
Use Outrank, don't use Outrank, but at least know and accept the risks if you do. Their landing page does not explain the risks outside of a tiny FAQ that barely scratches the surface.
Finally: make sure you are building an email list, a reddit/Facebook/Discord community and experimenting with paid ads as well.
Nobody should rely on search engines.Ian NuttallSep 2617188Why I hate being a founderinsightConfession: I hate being "a founder"
I love building stuff and learning by doing but when it becomes features for others, complicated devops, feature creep or competing with dozens of alternatives I start looking for the kill switch!
I've talked before about struggling with knowing what to work on despite spending so much time using conductor, codex, claude code, cursor/grok bot etc
https://t.co/7q2prf4CyX
I took a bunch of time off and went back to building small products, sites and tools outside of tech and AI.
those consumer apps make 5x more than keep.md, with almost no maintenance, no costs, and a fraction of the dev build time.
Maybe that's enough? Just do more of that? I can't really share those here because they aren't interesting to a tech audience and there's no moat now - I struck gold with SEO and the market so sharing would only harm me
They're also not long term revenue streams so you need a constant feed of new apps
Feels a bit isolating to just do that too which is why I keep coming back to "higher purpose" and trying to build and do things in public to help/share with others who love this "holy shit you can build anything" era we're in
I just want to do things my way (which is usually the opposite of what others do), build cool stuff, help others, and somehow make good money from it.
I have a few options that are not mutually exclusive and ordered by likelihood to succeed:
1. Keep quietly building consumer apps and grow them with SEO/AEO
2. Take my love for small, single utility dev/productivity tools and sell them at <$50 one time purchase
3. Persist with my newsletter idea and try to grow a sustainable audience
4. Partner up with a dev or team on a good product who want to scale with SEO and AI search.
5. Get sponsored by somebody to just build free stuff in public using their tool (devrel by proxy)
I'm unemployable I think so getting a proper job would not be appealing unless it was fully async, autonomous, and remote and most big cos will never want to do that.
I'm looking for the tough love @Shpigford got when he posted recently about being on the cusp of something but it never quite becoming viable.
https://t.co/7eHpNUH7Y8Ian NuttallAug 268777Clone repositories to run agents in parallelguidemy adsense accounts have made over $1m dollars (seriously) in lifetime earnings
most are dying/dead because of ai. so how do you stay sustainable in an agent-first economy?
- people love free shit
- but they hate (most) ads
- and they rarely pay for info
a standard for inserting ads into mcp, clis, markdown, anything agents use to get context COULD work
but they need to be aware of it up front. we’re used to media sites plastered with dozens of ads but it’s so early with context for agents and very much “pre-adsense” era
also it has to be a proper standard otherwise you run into prompt injection territory and lose the users trust
@AmpCode works because they control the platform and very clearly launched it as a free option with curated ads ❤️
been thinking about this for a long time now, and really hope we find a way for content creators to get paid and for people getting value out of free stuff to not kick off about very focused, relevant, and non-intrusive ads
(also I would really like to replace my adsense earnings 😂)
https://t.co/nnJjxnyvVvIan NuttallJan 264854How Cursor changed my building careerguideCursor changed everything for me.
I was copy-pasting from Claude 3.5 on the web back and forth to VS code like a madman and then this tool came along and instantly gave me the "holy shit" moment...
https://t.co/0qklYOZL18
I started sharing tips on how I was using it, built a video library (RIP Cursorcasts) and grew from 15k to 60k followers, had 3M views on Cursor posts and 20K likes
With it, I shipped 2 software products in public to $10k MRR (sold them for $280k), a paid course, dozens of SEO-powered sites and apps, actually learned how to code and not just vibe code, and made a bunch of internet friends along the way.
I'm addicted to building with AI and it's all thanks to Cursor.
Congrats team. You deserve it. 🫡Ian NuttallAug 2612521Keep MCP documents a Mac Mini setupguideOne of my favourite things about the Keep MCP server is when I am doing something interesting, I tell the agent:
"Create and update a Keep note to document this so I can write a blog post about it later"
Right now it's syncing my MacBook Pro and Mac Mini apps, documents, dotfiles, CLIs, skills, repos etc to run coding agents and GitHub action runners remotely 24/7.
It's matching my wallpaper, dock icons, installing Chrome extensions, Tailscale, Jump Desktop.
The Mini is upstairs and I'm just chatting it out with Codex. 🤯Ian NuttallSep 261414Codex struggled while Opus fixed the workguideI thought I was going mad, but 5.2 codex isn't running well today
I had to switch to opus 4.5,had issues like:
- writing a lot of wrong code to create nextjs sitemaps even after being given the exact docs on how to do it (opus deleted it all and fixed it in one shot)
- changing things not to do with what I asked (like text/content)
- spending 10 minutes trying to figure out how to copy a tailwind 4 typography examplr from another project (opus did it in <1m)Ian NuttallJan 266313Comparing Opus Codex and Claude CodeguideI'm refactoring my Hono app that powers the email list at https://t.co/sC6hbJxqJ9 so I can open source it
Gave the same prompt to Codex, Claude Code, and Droid (CC and Droid both using Opus 4.5)
Droid
had the best overall refactor report by far. It planned out a nice content structure for drafts, scheduled, sent
Proposed a full "shape" for a CLI to make everything agent-first and how to use the Hetzner CLI to set up a VPS
Codex (5.2 codex high)
focused much more on file-by-file analysis and found 4 security issues Droid didn't, and 3 reliability/perf issues
also found a bunch of missing files, hardcoded personal things that I need to fix to make it OSS
no code examples at all!
Claude Code
found a couple of potential security issues the others didn't find and had a lot more code examples (Droid had some, Codex had none)
didn't go as deep into the agent-first angle to let the agent set up a VPS and the CLI it recommended wasn't as good as the Droid version imo
boring tldr; I need both Opus 4.5 and 5.2 codex 😰Ian NuttallJan 262811A creator sponsorship pivot for Iloinsighta b2b idea for a full rebuild of @ilodotso from the ground up...
back before the X API destroyed a bunch of businesses, ilo was around $4k MRR, it's now somewhere ~$300
but doing some churn/data with claude code shows me that 70% of monthly revenue and 92% of annual is at risk of churn!
98% of previous lifetime users are not active on the platform any more.
I personally have used it to track my growth on X but I vibe coded my own link-in-bio tool instead of using ilo.
claude reckons it isn't currently solving a real pain point and is more of a "nice to have" and I kinda agree.
so here's my idea of a pivot (let me know below what you think):
The content creator platform to land sponsorship offers for your audience
Launch a professional media kit backed by real-time stats. Get paid directly. No middleman fees.
(ilo = i land offers 😉)
by focusing on content creators who can streamline their business and make money from the product, my theory is it will be a lot more "sticky" and used more frequently
v1: $15/mo or $150/yr for a landing page with real-time stats and stripe/polar/lemonsqueezy/paddle payouts
v2: add a marketplace where brands can find your products and maybe charge a small <5% fee for connecting sponsor to creator
with programmatic seo and aeo/geo i could probably rank pretty well for both brand pages and marketplace pages too 🤔
feel free to pick this apart and tell me why it's a bad idea. maybe time to vibe code an mvp to see how possible this is?Ian NuttallSep 252110Slow down to avoid coding-agent slopguideMy hot take:
Slow down.
- Chat in plan mode.
- Build one feature at a time.
- Review the code and give feedback.
- Log session summaries.
- Give previous summary to /new agent.
- Be very selective with MCPs.
- Max of 2 agents (1 frontend, 1 backend) at a time.
Slow > slop.Ian NuttallNov 253.3K1.6KA cdx shortcut configures Codex CLIguideCodex CLI pro tip
Add this updated function to bashrc/zshrc & run cdx instead of codex to get:
- cdx update to update codex
- gpt-5 codex model
- web search enabled
- full auto mode (read, edit, run commands)
- summaries of model thinking
(function in image alt text) https://t.co/qLlZGajY7BIan NuttallSep 251.1K1.4KAn SEO skill for coding agentsinsightOne of my fave apps I worked on recently is https://t.co/yYlWqhPg9d
"The only SEO skill your agent needs."
# Run this
npm i -g seo
# Then this
seo start
My agents use it with GSC + Ahrefs data to fix issues, find keywords, check AI readiness and write better ranking pages.
I built this over a few months while ranking a bunch of brand new programmatic SEO sites (the pSEO reports are very good for this) and eventually tidied it up into a nice CLI.
That `seo` package name... 🤤
Give it a try and lmk what you think. I don't accept PRs but if you find a problem you can create an issue on GH (it's fully OSS) with a prompt and I'll give it to an agent to fix!Ian NuttallAug 26423854Use Claude Skills with any AI agenttipThis tool lets you use Claude Skills with any AI agent (Codex, Factory, Cursor etc)
- Same format and folders, just CLI instead of tools
- Access marketplace skills through GitHub
- Uses progressive disclosure (skills on demand)
github.com/numman-ali/ope… https://t.co/Hvp2IElxmDIan NuttallOct 25645730Session logs preserve context between agentsguideI log my sessions with /end-session to log:
- overview of what we did
- files we created/deleted/modified
- why we made decisions
- what worked/didn’t work
- steps for next session
then I run /start-session with new agent and it reads the last log to get the context it needsIan NuttallDec 25489509Serve Markdown versions to AI clientsguidealso:
1. add <link rel="alternate" type="text/markdown"> in the html specifying the exact markdown version
2. add a link header for your llms.txt file: Link: </llms.txt>; rel="describedby"; type="text/markdown"Ian NuttallAug 26157315What Claude Code needs to win users backguideA list of things I think Claude Code could do to win back people switching to Codex CLI:
- open source Claude Code
- reduce sycophancy/make it less verbose (or add option for that)
- more transparency about how/why the model degrades
- fix tui flashing bug! PLEASE
- improve model hallucinations like GPT-5 has
- better thinking for removing files/lines of code to prevent accidental deletions
- less boilerplate or pseudo implementations (break it into working chunks if needed)
- ability to change/remove/reduce the system reminder prompts
- file based session auto-compact with much more detail on the conversation for future reference
what would you want to see improve to make CC work better for you?Ian NuttallSep 25793291Gemini designs better while Codex codes betterguideEarly thoughts on Gemini 3 Pro:
- Very fast, very smart
- Does not have the attention to detail of GPT-5.1
- Incredible frontend designer (which I need!)
- Not as good at code as Codex 5.1
- FUCKING LOVES TO COMMIT SHIT
My TLDR;
Backend: Codex 5.1
Frontend: Gemini 3Ian NuttallNov 251.3K285Find and install any agent skillguideUse this command to find and add any skill to your AI agents:
npx playbooks add skill
- Enter for fast search, Tab for semantic
- Returns official, trusted sources first https://t.co/pavqvZCE7wIan NuttallJan 26175230Build local tools around your personal agentguideseeing a lot of people say their openclaw dies/doesn’t respond often
PRO TIP: build your own local-first api/cli tools for it and keep your personal agent “dumber”
I have:
- my own cron system
- an event system to wake agent up from scripts
- https://t.co/JiFU50aLoJ to fetch x bookmarks
- my own agent-first email marketing cli - a trained blog/email draft writer
- a google search console cli for improving my seo
- etc
build your own tools don’t ask the agent to do everything!Ian NuttallFeb 26134222No posts found
Try a different phrase or clear your filters.
Couldn't load the archive
Check your connection and try again.