Less AI slop. More engineering judgment and more correct code.
Most coding-agent mistakes are judgment failures, not syntax failures: guessing instead of asking, building abstractions before the second case exists, patching symptoms, accepting review comments blindly, touching unrelated code, and saying "done" without evidence.
greybeard is one unified operating discipline for that:
- Think from facts - surface load-bearing assumptions, ask only questions that change the implementation, and push back when the request is unsafe or overbuilt.
- Build the minimum - reuse what exists, prefer stdlib and native platform features, avoid speculative dependencies and abstractions.
- Debug systematically - reproduce, read the full error, trace bad data to its source, form one hypothesis, and fix the root cause.
- Cut surgically - every changed line traces to the request, verified root cause, or cleanup created by the change.
- Use agents like tools - delegate only independent or review-worthy work, with complete task-specific context.
- Verify before done - run the narrowest check that proves the behavior, then report the result plainly.
npx skills add nassiharel/greybeardThis installs the reusable skill from skills/greybeard/SKILL.md.
/plugin marketplace add nassiharel/greybeard
/plugin install greybeard@greybeard
The plugin manifest points at the same skills/ directory.
Codex can use the native plugin manifest in .codex-plugin/plugin.json, or the portable
instructions file:
curl -o AGENTS.md \
https://raw.githubusercontent.com/nassiharel/greybeard/main/AGENTS.mdUse ~/.codex/AGENTS.md if you want the compact rules globally.
This repo ships gemini-extension.json with contextFileName set to AGENTS.md, so the compact
rules load as extension context.
gemini extensions install https://github.com/nassiharel/greybeardRestart Gemini CLI after installing or updating the extension.
OpenCode can read AGENTS.md from the project root. This repo also ships a thin plugin that
registers skills/ and injects AGENTS.md only when it has not already been loaded.
Add the plugin to your opencode.json:
{
"plugin": ["./.opencode/plugins/greybeard.mjs"]
}If you keep one shared checkout, point that entry at the plugin's absolute path.
Copilot reads .github/copilot-instructions.md from a repository. Install the compact rules with:
curl --create-dirs -o .github/copilot-instructions.md \
https://raw.githubusercontent.com/nassiharel/greybeard/main/.github/copilot-instructions.md| File | Purpose |
|---|---|
skills/greybeard/SKILL.md |
Canonical long-form skill |
AGENTS.md |
Compact portable rules for agents that read project instructions |
.github/copilot-instructions.md |
Copilot adapter generated from AGENTS.md |
.claude-plugin/plugin.json |
Claude Code plugin metadata |
.codex-plugin/plugin.json |
Codex plugin metadata |
gemini-extension.json |
Gemini CLI extension metadata |
opencode.json and .opencode/plugins/greybeard.mjs |
OpenCode plugin adapter |
You should see smaller diffs, fewer invented abstractions, earlier clarifying questions, more root-cause debugging, review feedback checked before implementation, and final answers that lead with what changed and what was verified.
See examples/ for small before/after proof cases.
- Truth over momentum - name uncertainty instead of hiding it.
- Simplicity over cleverness - the best code is still the code never written.
- Root cause over symptoms - patch once where the bug begins.
- Evidence over confidence - done means checked.
- Brevity over filler - cut prose, not precision.
MIT.