Imported from filipenegrao/personal-library (
AGENTS.md). Install upstream withnpx skills add filipenegrao/personal-library. Copyright stays with the author.
AGENTS.md — Operational Manual for Agents
This file is mandatory for any agent before acting in the repository. Read this file BEFORE any action. This is not optional.
1. Project context
See CLAUDE.md for product description, stack, folder structure, and commands.
This template is generic: adapt the product placeholders when creating a new project.
2. Repository map
| Path | Purpose |
|---|---|
CLAUDE.md |
Quick project context, stack and commands |
AGENTS.md |
Global operational contract for all agents |
HANDOFF.md |
Memory between sessions; must be updated at the end |
STATUS.json |
Domain and feature states; source of truth for tracking |
docs/architecture.md |
Layers, dependencies and guardrails |
docs/dashboard-integration.md |
Optional contract for integration with Agents Dashboard |
docs/decisions/ |
ADRs of the template and project |
docs/progress.md |
Optional human checklist |
docs/session-log.md |
Append-only history of sessions |
harness/prompts/builder.md |
Base prompt for the implementation agent |
harness/prompts/qa.md |
Base prompt for the validation agent |
harness/prompts/orchestrator.md |
Base prompt for the Builder -> QA cycle |
.github/workflows/harness-ci.yml |
Mandatory quality sensors |
.harness/pipeline_run_id |
Optional local state of dashboard integration; do not version |
~/.claude/hooks/pipeline-event.sh |
Optional helper to report Agents Dashboard stages |
3. Absolute rules
- Never declare a task complete without running the mandatory sensors of the current stack.
- Always prioritize dry-run mode when introducing new automations.
- Always keep a log of decisions and actions applied when the flow requires auditing.
- Always update
HANDOFF.mdat the end of the session. - Always update
STATUS.jsonwhen a feature status changes. - If the project uses
docs/session-log.md, always add an entry at the end of the session. - Never hardcode credentials, tokens, account IDs, or sensitive URLs.
4. Architecture
Refer to docs/architecture.md for layers, allowed dependencies and guardrails.
General rule: dependencies flow from left to right between layers. Never import from a layer to the right of yours.
5. Mandatory workflow
- Read
CLAUDE.md,AGENTS.md,HANDOFF.mdandSTATUS.json. - Select the target feature and confirm an atomic scope.
- Implement within the correct layers.
- Run sensors compatible with the stack.
- Fix failures and run again.
- Update
HANDOFF.mdandSTATUS.json. - Update
docs/session-log.mdif the project uses it. - Deliver objective summary with remaining risks.
6. Minimum sensors before completing
Use only commands compatible with the current stack.
# Python
ruff check .
mypy .
pytest
# Node / React / Next.js
npm run lint
npm run build
# run npm run test only if the repository actually has this script
If a sensor fails, do not ignore it. Fix or explicitly document the block in HANDOFF.md.
The template does not include by default a reusable structural test for architecture.
7. Definition of Done (DoD)
A feature can only be marked as done when:
- the implementation respects architecture and guardrails
- the mandatory sensors are green
- the residual risk has been documented
HANDOFF.mdhas been updatedSTATUS.jsonreflects the real state
8. Escalation to human
Escalate immediately when:
- conflict between security and automation
- product uncertainty affecting agent behavior
- need for new scope outside the current feature
- third failure cycle without resolution
9. Optional Dashboard
If the project uses Agents Dashboard, the integration contract is the versioned set in agents-dashboard/scripts/.
Supported variables:
PIPELINE_DASHBOARD_URLto define the dashboard base URLPIPELINE_RUN_ID_FILEto override the run state file.harness/pipeline_run_idas default per repository when there is no override
Use the dashboard only as observability of the flow; it does not replace HANDOFF.md nor STATUS.json as source of truth for the work.