| name | evaluate-learnings |
| description | Analyze memory files for generally applicable learnings and extract them to AGENTS.md, skills, or other permanent locations. Use when reviewing accumulated corrections and decisions for broader applicability. |
Evaluate Learnings
Use this skill to review accumulated memory entries (corrections and decisions) and determine which are project-specific and which represent general patterns that should be promoted to permanent, structural locations.
When to use
- After a project has accumulated several memory entries.
- Before backporting improvements to a boilerplate or starter template.
- Periodically, to keep memory files lean and ensure valuable patterns aren't buried.
Steps
1. Read all memory files
Read every file in .agents/memory/. For each entry, note whether it describes:
- Project-specific knowledge — tied to a particular library, API, or architectural choice unique to this project.
- Generic agent behavior — a pattern any agent in any project would benefit from.
- Generic tooling/convention — a pattern tied to the tech stack (e.g., Bun, Vitest, React) but not to this specific project.
2. Classify each entry
For each entry, assign one of these dispositions:
| Disposition | Meaning | Action |
|---|
| Stay | Project-specific; belongs in memory | No action |
| Promote to AGENTS.md | Generic behavioral rule | Add/refine a rule in AGENTS.md |
| Promote to skill | Generic pattern best expressed as skill guidance | Update the relevant skill's SKILL.md |
| Promote to context.md | Architectural insight that applies project-wide | Add to .agents/context.md |
3. Present the classification
Show the user a table with:
| # | Entry title | Source file | Disposition | Target location | Rationale |
|---|
Wait for the user to confirm or adjust before making changes.
4. Apply approved promotions
For each approved promotion:
- Add the learning to its target location. Follow existing formatting and conventions in that file.
- Delete the original memory entry. Active memory files should only contain entries not yet captured elsewhere. After promoting a learning, remove it from the source file (
corrections.md or decisions.md). This keeps active memory lean and avoids stale duplicates competing with the promoted version. Git history preserves the full timeline if needed.
5. Verify coherence
After all promotions:
- Re-read
AGENTS.md to confirm new rules don't contradict existing ones.
- Re-read any updated skills to confirm additions fit the skill's scope.
- Check that no learning was promoted to multiple overlapping locations.
Classification guidelines
Goes in AGENTS.md when…
- It's a behavioral rule that constrains how agents should act (e.g., "ask before guessing", "guard CLI entry points").
- It refines or strengthens an existing rule that proved too vague.
- It clarifies the boundary between agent files (e.g., what goes in
context.md vs AGENTS.md).
Goes in a skill when…
- It's a procedural best practice within a specific workflow (e.g., "read the file tail before appending" fits the
remember skill).
- It's a technique or recipe for a specific tool (e.g., "check DOM errors, not just console" fits
review or playwright-cli).
Stays in memory when…
- It's about a specific library, API, or framework behavior unique to this project.
- It's a workaround for a project-specific integration issue.
- It would be meaningless without the project's specific context.
Anti-patterns
- Don't over-promote. Not every learning is a rule. If a correction happened once and is unlikely to recur in other projects, leave it in memory.
- Don't duplicate. If
AGENTS.md already says "prefer immutability", don't add a memory-derived rule that says the same thing differently.
- Don't generalize prematurely. A pattern observed once in one project is an anecdote, not a rule. Promote with confidence only when the learning clearly applies beyond this project's specific context.