# Global Agents.md

> Source: <https://gist.github.com/MelbourneDeveloper/4e3d96b98ac345d0a45d1808973e709e>
> Published: 2026-11-18 20:48:05+00:00

**Recommendation:** Adopt a repo-level Agents.md that fixes autonomy rules, model routing per task type, and a build-analyze-test CI order so coding agents act without asking and commit after each step.
More in the playbook "Working with AI coding agents": <https://wpnews.pro/playbooks/coding-agents>

[Agents.md](https://gist.github.com/MelbourneDeveloper/4e3d96b98ac345d0a45d1808973e709e#file-agents-md)
- 1 feature branch, 1 working tree: merge all loose branches/stashes/uncommitted files into it & push. Delete branches already merged into it
- ❌ ask user questions. Decide, act, report. Act **autonomously**
- Duplication = bug. Use [deslop](https://deslop.live/docs/for-ai/)
- Important lints → error. Run static analysis routinely
- 🧹 After work: delete safe system caches & unneeded build files, kill your owned processes that eat CPU or memory and check disk space
- Refactors: violent, immediate, no back-compat. Legacy code banned → delete now
- Minimize code; decide toward less. Delete dead code now
- 📦 Minimize 3rd-party deps: routinely delete unused packages/crates, add none without need, consolidate overlapping deps into one, avoid those with a vulnerability history
- Reply ≤2 sentences, emojis for emphasis. Don't stream responses
- ✅ Tick mandatory plan checklist items as you work
- Website docs, blogs = prose classes, CSS BUDGET 2K LOC default
- ❌ self as git commit coauthor
- PR CI Phases = Build/Analyze/Cache Build -> Run tests parallel with coverage (on cached build) -> Coverage checks
- Declare all models with [typeDiagram](https://typediagram.dev/docs/language-reference.html) , generate code from them, consider[tdbin for wire protocol](https://typediagram.dev/docs/tdbin.html)
- Specs: hierarchical non-numeric IDs, cross-ref'd across code/tests/plans/artifacts
- Repo structure: [AgentPMO](https://www.agentpmo.dev/get-started/) . Deploy structure:[Shipwright](https://github.com/Nimblesite/Shipwright)
- Aim 💯 coverage via use-case E2E tests. Unit tests banned unless isolation needed
- 💰 Watch spend (GH Actions, agents, cloud: GCP/Supabase…). Minimize & prevent unneeded future spend
- Tests: many input permutations & assertions. Prefer long tests: many user interactions, many assertions each
- Offload simple work to pi; pick models per `/Users/christianfindlay/Documents/Code/PiNotes/MODEL-COMPARISON.md` , always each family's newest id in`pi --list-models` (today`zai/glm-5.3-flash` repetitive/fixtures/tests,`zai/glm-5.3` scoped code/reviews,`deepseek/deepseek-flash` code/oracles,`kimi-coding/k3.1-pro-highspeed` UI,`deepseek/deepseek-v4-pro` narrow audits only); bounded tasks, verify results, record model/date/results in PiNotes
- ❌ Claude artifacts: reports/docs → markdown in repo
- Commit after each step; don't let uncommitted files pile up
- 🔒 Close issues only after testing the LATEST release's own binaries 📦. ❌ Never close via PRs
