improve-agent-harness

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Improve an agent harness

## Inputs

- Collect the failed artifact, command output, review feedback, or incident record.
- Read the repository instructions and the skill, card, hook, test, or runtime path that actually
  reaches the failed decision.
- State the desired terminal behavior and the boundary that must not change.

## Workflow

1. Reproduce or mechanically confirm the failure before changing the harness.
2. Route the fix to the earliest layer that can prevent it:
   - task-specific procedure → skill;
   - role responsibility or context boundary → agent card;
   - deterministic invariant → code, schema, hook, or test;
   - cross-task operating rule → repository instruction;
   - personal history or preference → private memory, never this public repository.
3. Explain the failure mechanism. Do not add prose inside an artifact that was never loaded at the
   failing decision point.
4. Apply the smallest upstream change and remove any stale or duplicate instruction it supersedes.
5. Add an empirical regression check. Prefer a black-and-white mechanism signal over asking an
   agent whether its own output looks correct.
6. Re-run the original failure path and the full repository health command.
7. Record what signal would justify another refinement or deletion. Do not create a permanent rule
   for a one-off preference with no recurrence risk.

## Safety boundary

- Keep self-improvement reviewable and version-controlled.
- Do not let a scheduled maintainer change credentials, permissions, external systems, budgets, or
  release boundaries without explicit authority.
- Treat failures as data, not as permission to expand scope.

## Output

Return the incident evidence, root cause, upstream change, regression evidence, and the next
measurable signal. The task is complete only when the original path passes with the new mechanism.

From qilinxiaoxiang/agent-skills-and-cards, MIT.

More from qilinxiaoxiang/agent-skills-and-cards