@liran_tal @mattpocockuk gpt-tokenizer is just too slow for a lot of use cases, it's obviously a lot more accurate for openai models, but I found that accuracy wasn't worth the performance penalty for my needs
English
Charles Pick
1.9K posts
Charles Pick
@c_pick
Founder of @codemixdotcom - codemix makes sure humans and coding agents always build the right thing.
York, United Kingdom Katılım Ağustos 2010
711 Takip Edilen760 Takipçiler
@c_pick @mattpocockuk but the ergonomics for that is painful although likely claude can figure out the bash for that by piping sed and awk
if you want to try something cleaner though I built tokenu github.com/lirantal/tokenu
English
Has anyone made du, but for tokens?
I want to make a rule for my agent where if a file is more than 5K tokens, split it up.
English
> it's essentially the same model as postinstall scripts
these are gradually being phased out of package managers because they're so dangerous. Honestly it feels like this needs rethinking. I doubt hardly anyone reads their SKILL.md files thoroughly, and markdown is not a well-known threat vector - I guess with things like this it will become well-known soon.
English
@itechnologynet Ah this fails by default, it only runs if the skill's frontmatter declares something like allowed-tools: Bash/Bash(rm *) etc. Just make sure to check what the skill allows if you're downloading anything external, it's essentially the same model as postinstall scripts
English
if your skill depends on dynamic content, you can embed !`command` in your SKILL.md to inject shell output directly into the prompt
Claude Code runs it when the skill is invoked and swaps the placeholder inline, the model only sees the result!
English
@kentcdodds *British spy. FWIW I don't usually run into this problem either, the dev tools got quite good at compaction (though it is slow and sometimes i start fresh just to avoid that). It's a real thing when using LLMs in apps though and you do need good compaction strategies there too
English
Maybe this is the "German spy raises three fingers meme" but I'm telling you, it's just not been a significant issue for me in those tools. I don't use Claude Code or Open Code so my exposure to this problem might be limited?
English
Everyone knows that the last 40% of the context window of AI models start to get pretty unreliable...
Except I don't experience this at all 🤔
My two primary AI tools are Cursor (mostly GPT 5.4) and ChatGPT with very long conversations. Cursor compacts and it's still does fine.
English
@threepointone very good, next step is give it a language server for tsc
English
full rewritten from scratch, now executed (and secured!) inside a dynamic worker, a lot lighter and powerful.
much, MUCH more to come.
(sharing here early because you know how we do)
sunil pai@threepointone
what if we gave every cloudflare agent a file system (sqlite/r2) tools to operate on it (shell) powered (and secured!) by codemode all powered by workers ai (or byom)
English
hey @supabase @kiwicopple I think Tyler's account's been compromised
English
Do you think you are expected to keep yourself up to date on your field of work outside of work hours?
E.g. learn new technologies, ai-tools etc if you are a developer?
English
@sebastienlorber most of what we do is not speccing programming languages though. The thing that people seem to be missing is the idea of "gradual specification" - you can specify things in close detail when you want fine grained control, and broad strokes the rest of the time
English
Here's a sufficiently detailed spec from a JS std method
(referencing many other specified things)
Wouldn't you rather write the code directly? 😆
gabby@GabriellaG439
New blog post: "A sufficiently detailed spec is code" I wrote this because I was tired of people claiming that the future of agentic coding is thoughtful specification work. As I show in the post, the reality devolves into slop pseudocode haskellforall.com/2026/03/a-suff…
English
Hate MCP all you want, but the sheer potential for malicious SKILL.md files should keep you up at night. That is a wolf in sheep's clothing.
English
@dillon_mulroy i had one of my worst days with agents in ages, the variance is the worst aspect. Consistently good would be ideal, but I could also live with consistently bad, this middle ground feels like a casino
English
@mattzcarey @garrytan like, no one is reading that carefully before they accept it
English
so now we chill executing untrusted bash through a md file
Lydia Hallie ✨@lydiahallie
if your skill depends on dynamic content, you can embed !`command` in your SKILL.md to inject shell output directly into the prompt Claude Code runs it when the skill is invoked and swaps the placeholder inline, the model only sees the result!
English
@DavidKPiano @combdn also, also, curious how this plays with the cache
English
@DavidKPiano @combdn the ui is a lot simpler and easier to understand if you omit deep threading, also how do you know when to "return" to the parent thread (if you're thinking about replacing sub-agents this seems important)
English
biggest LLM-UI tell - <Card><Card><Card><Card></Card></Card></Card></Card>
English
what if this is only part of the story
Think about it
sunil pai@threepointone
what if we gave every cloudflare agent a file system (sqlite/r2) tools to operate on it (shell) powered (and secured!) by codemode all powered by workers ai (or byom)
English
seems like there's 2 different approaches emerging:
1. Maximum freedom for the agents, letting them run with little oversight - the claws, and now hermits of this world
2. Extensive guardrails, tight controls and restrictions that keep agents on track towards specific goals
As models improve it's going to be interesting to see which wins
English
Over the weekend, I used Hermit, the autonomous application framework, to:
- have Hermit manage its own website and come up with a CRM demo
- tasked it to keep track and manage all the loose ends around the house
- created a librarian running my book collection
- set up a personal coach to help me train for this summer's big swimrun event
- onboarded a couple of people to help them use it as assistants for their small business
It is quite amazing that such a small codebase can design and manage its own datastore, create the app, write skills, ...
The most fun is that you design it by describing the job that Hermit has. It is not just code writing or question answering, it is an agent that creates its own environment to be able to do its job.
You should give it is spin, it runs by default in a secure sandbox and I added an easy to get started MacOS walkthrough.
hermit-ai.com
English
