Introducing watchmen. It writes the skill files your coding agents should already have from your own sessions, shared across Claude Code, Codex, and pi, so you stop paying tokens to re-explain what your agent learned last week. Local and open-source.

Coding agents do carry context now, but it's siloed per tool and you maintain it by hand. CLAUDE.md doesn't help Codex and AGENTS.md doesn't help Claude Code, so both are only as good as what you wrote down. Watchmen mines what you did and writes both from one corpus.

It runs locally and in the background, behind every agent you already use, watching what you actually do and writing the recurring work to disk as runnable skill bundles and a workspace brief. You install it once and never change how you work.

Watchmen measures its own effect with a per-project impact card that tracks median tool errors, prompts, and cost per session before and after your skills land, over 16 weeks. The card is labeled correlation only, not a controlled experiment, so we don't oversell the signal.

Watchmen is local by default, MIT licensed, and runs on macOS and Linux. Your transcripts never leave your machine except for the LLM calls you opted into, and it's available now at:

I find auto-skill creation is a quick method to drown in slop. Creating skills is just part of the equation. There needs to be an eval harness to review which skills actually get used, which ones sit there with their descriptions, eating tokens in your context window. And which ones are used but don't improve the outputs... Most of our agents aren't really suffering from a lack of agent skills. On the contrary, they're already drowning in noise.

I like the skill-file angle, but I’d keep one rule strict: learned patterns need provenance. Which session produced it, what failure it fixed, and when to delete it.

This is the right direction. Re-teaching agents the same repo lore every week is quiet token burn and quiet workflow debt.

This is a great idea! Claude have gotten slower lately and expensive

local and open source is the key here. paying tokens to re-explain what your agent learned last week adds up fast, especially on long projects. making that shared across codex, claude code, and pi is a solid value proposition

Nice ! Interesting idea

Skill files are underrated because they turn “explain it again” into infrastructure. That’s where agents start compounding.

how do you handle the sheer amount of skills that can created from something like this? I built a manual version of this and overtime there were just too many - they didn't fit in the budget allotted for skill descriptions in CC/Codex.

potential to get blasted with useless skills seems high, specially because you are not always asking for the same thing across different repos... you should demonstrate how it works beyond this simple video, together with the dashboard.

How it works. Technical teardown. Interesting project!

looks cool, would prefer to use my existing openai subscription or api key instead of openrouter though

Does it work on windows?

is it vibe coded?

hi , we want to cooperate with you as a inner PR, is there any ways i can connect you so we discuss some details?

this is so cool dria! could i make a pսmpfun for watchmen and make a donation directly to your github for this?