ai-workbench
AGENTS.md0 starsChanged 47 days agoMITView on GitHub
The public reference for the file-first AI workbench pattern: markdown as state, deterministic scripts, a session loop with an open and a close, and one scripts/check. This repo holds the template, the lint, the scaffold and the two session skills that anyone can copy. Nothing here depends on any other repository. 1. Public repo. No private hostnames, LAN addresses, credentials, personal paths or references to any specific deployment — in code, docs, comments or commit messages. 2. Standard library only.…
In this repository6 other files it gives its agents.
- templates/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- templates/CLAUDE.mdCLAUDE.md
- session-closeSkill
- session-openSkill
- workbench-scaffoldSkill
The file
Files
AGENTS.md — ai-workbench
Purpose
The public reference for the file-first AI workbench pattern: markdown as state,
deterministic scripts, a session loop with an open and a close, and one scripts/check.
This repo holds the template, the lint, the scaffold and the two session skills that
anyone can copy. Nothing here depends on any other repository.
Rules
- Public repo. No private hostnames, LAN addresses, credentials, personal paths or references to any specific deployment — in code, docs, comments or commit messages.
- Standard library only. The lint and the tests run on a bare
python3(3.10+); the check script and the hook are plain bash. No package manager, no third-party tools. templates/andskills/must be copyable verbatim into a downstream repo. They may name each other by their downstream paths; they never assume this repo's layout.AGENTS.mdis capped at 100 lines and the desk file at 120, here and in every copy. The lint enforces it; overflow moves to a linked document, not to a bigger cap.- Pattern text carries no history: no dates of past decisions, no changelog prose, no version metadata in skills. Git history is the changelog.
- Tests use
unittest, live intests/, and cover the lint's observable behaviour (findings, exit codes, flags). No tests for glue. - The lint may only ever report on files; it never edits, formats or deletes anything.
Layout
templates/— the downstream shape:CLAUDE.md,AGENTS.md,STATUS.md,DECISIONS.md. The scaffold fills the placeholders.skills/—session-open,session-close(copied into a downstream.claude/skills/) andworkbench-scaffold(installed globally; it needs a clone of this repo to copy from).scripts/—workbench_lint.pyandcheck; both are copied downstream unchanged..githooks/pre-commit— runsscripts/check; also copied downstream unchanged..github/workflows/check.yml— runsscripts/checkon every push and pull request here; not part of the downstream shape.tests/— the lint's unit tests.docs/why.md— the assumptions and evidence.
Commands
scripts/check— lint plus tests; the whole CI. Read-only.python3 scripts/workbench_lint.py --help— the lint's flags. Read-only.python3 -m unittest discover -s tests -v— tests alone. Read-only.
Working here
Cut a branch, keep the commit small, run scripts/check, push, open a pull request that
references its issue. The maintainer merges. Issues are the only backlog (labels: bug,
task, decision, blocked). Activate the hook once per clone with
git config core.hooksPath .githooks. This repo is the pattern's source, not a
workbench: it has no desk file and does not run the session loop on itself.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

