decent-skills

Agent skills I use myself. Public, MIT.

/council-review runs a code review through Codex, Gemini and Grok at once. The reviewers get the same evidence packet, cannot inspect the repo, and never see each other's answers. Only the host model reads the real files, and it checks every finding against them before anything is reported to you. Nothing is changed until you say fix.

A real run is in examples/review.md, including the finding the host rejected.

/wawa answers "where are we at, and what's next". It reads the repo first and STATUS.md second. When the two disagree, that disagreement is the answer, because notes are hypotheses with a timestamp on them.

/init-repo-docs writes STATUS.md, TODO.md and DECISIONS.md from git history, planning docs and your own notes, with the evidence traced. It does not commit.

None of the three start on their own. You type them.

Install

Claude Code

claude plugin marketplace add cochinescu/decent-skills
claude plugin install decent-skills@decent-skills

Codex, ChatGPT, Grok

git clone https://github.com/cochinescu/decent-skills.git
cd decent-skills
./install.sh

That links skills/* into ~/.claude/skills/ and ~/.agents/skills/, and the blind reviewer into ~/.claude/agents/. It symlinks when it can and copies when it cannot, and it leaves an existing blind-reviewer.md alone rather than overwriting yours.

/council-review wants codex, agy and grok on PATH for a full run. Default pins: Codex gpt-5.6-sol, Antigravity Gemini 3.1 Pro (High), Grok grok-4.6, and Claude fable when you add with claude. All four sit at their vendor's top tier. Change the pins if your account differs.

What this sends, and what it cannot see

/council-review sends the code you selected to OpenAI, Google and xAI, through their own CLIs under your own accounts. It builds a packet: the diff, the relevant file excerpts, and the check results. That packet leaves your machine. Read it before you point this at anything you cannot share.

Isolation is the design, and it has a cost. Reviewers know only what is in the packet, so they will sometimes be confidently wrong about the code around it. In the example run the reviewer declared a shell command syntactically broken; it was not, and it could not find out, because it cannot run anything. That is the trade: you get an opinion uncontaminated by the session that wrote the code, and you pay for it in context. The host model exists to catch exactly this, which is why findings are verified against the real files and not passed through.

The Claude reviewer stays off by default for the same reason. This command is normally typed in Claude Code, so Claude built the packet and usually wrote the code under review. That makes it the least independent of the four. Add with claude when you want it anyway.

Where it came from

The method predates the skill. Codex, Antigravity and Grok went over the plan, three rounds of it, for the skill we took to the GTM Skillathon in August, Europe's first agent skill hackathon. It won first place out of the 34 that shipped that evening.

Also, not copied here

I also publish anxiety-reset from Anima Felix, and agentmarkup. I do not copy them into this repo. Copies go stale, and the licences are not mine to flatten.

License

MIT · Sebastian Cochinescu