Skip to content

ai-feedback-learning-loop.md

Latest commit

 

History

History
50 lines (31 loc) · 2.4 KB

ai-feedback-learning-loop.md

File metadata and controls

50 lines (31 loc) · 2.4 KB

AI Feedback Learning Loop

Documents the process for AI agents to propose rule changes and await human approval. Prevents silent rule drift between sessions.

When to Propose a Change

Propose a rule change when any of the following happen:

  • A human corrects your approach during development and the correction is worth preserving project-wide
  • You discover a pattern, constraint, or convention that recurs across multiple tasks and is not yet captured in .agents/rules/
  • A tool, command, or workflow behaves differently than what the rules describe
  • An accepted convention turns out to be incorrect or outdated

Do not propose a change for one-off situations specific to a single task.

Proposal Format

When you identify a change worth proposing, output the following block in your response:

## Rule Change Proposal

**Triggered by**: [What correction or learning surfaced this]
**Proposed rule**: [Exact text to add/change/remove in base.md]
**Section**: [Which section of base.md this belongs in]
**Rationale**: [Why this is worth preserving project-wide]

Awaiting human approval before applying.

Include this block in-line in your response — do not create a file or modify any .agents/rules/ file until the human explicitly approves.

One-at-a-Time Constraint

Only propose one rule change per session. If multiple corrections surface, log them in specs/<feature>/lessons-learned.md and propose them one at a time in future sessions after the current proposal is resolved.

Human Approval Required

AI agents MUST NOT self-apply rule changes. The rule files in .agents/rules/ are governance documents — changes require human review and explicit approval. This prevents agents from silently shifting project conventions based on in-session learning that may be wrong or context-specific.

Approval means the human explicitly says something like "approved", "apply it", or "go ahead and update base.md". Implicit agreement or lack of objection is not approval.

How to Apply After Approval

Once a human explicitly approves a rule change:

  1. Open .agents/rules/base.md (or the relevant rule file)
  2. Apply the exact change described in the proposal — no scope creep
  3. Open a PR with the change; include Rule-Change-Approval: <ref> in the PR description (reference the conversation, issue, or comment where approval was given)
  4. CI will verify the approval reference is present before allowing the merge