SKILL.md
Self-Score: 6-Check Assessment
Evaluate your adaptive behavior during this session against the 6-check rubric.
Read active_context.yaml to understand what was accomplished.
The 6 Checks
Score yourself honestly on each:
1. Mismatch Detection
Question: Did I spot divergences between expectations and reality quickly?
| Score | Meaning |
|---|---|
| Met | Caught mismatches immediately, logged them with deltas |
| Missed | Plowed forward despite signals, retried without noticing |
2. Plan Revision
Question: When things went wrong, did I change my approach (not just retry)?
| Score | Meaning |
|---|---|
| Met | Wrote new strategies, reduced step size, added guards |
| Missed | Repeated the same step 3+ times |
3. Tool Switching
Question: Did I abandon tools that weren't working and try alternatives?
| Score | Meaning |
|---|---|
| Met | Switched methods when one failed, preferred simpler approaches |
| N/A | No tool failures occurred |
| Missed | Kept hammering same tool despite failures |
4. Memory Update
Question: Did I capture reusable lessons from what I learned?
| Score | Meaning |
|---|---|
| Met | Added trigger-linked lessons to memory |
| Missed | Solved problems but didn't record patterns |
5. Proof Generation
Question: Did I attach evidence, not just claims?
| Score | Meaning |
|---|---|
| Met | Every major step has proof (logs, diffs, test results) |
| Missed | "Trust me" summaries without evidence |
6. Stop Condition
Question: Did I escalate appropriately when blocked or uncertain?
| Score | Meaning |
|---|---|
| Met | Asked crisp questions, presented bounded options |
| N/A | Never hit uncertainty requiring escalation |
| Missed | Guessed when should have asked, or asked trivial questions |
Instructions
- Review the session
- What mismatches occurred? - How did you respond? - What lessons were captured?
- Score each check
Be honest. Mark: - met: true if you satisfied the check - met: false if you missed it - Include a brief note explaining why
- Update active_context.yaml
``yaml selfscore: timestamp: "<currentisotimestamp>" checks: mismatchdetection: met: true note: "Caught API 403 immediately, logged delta" planrevision: met: true note: "Added token refresh step instead of retrying" toolswitching: met: false note: "N/A - no tool failures" memoryupdate: met: true note: "Added lesson about token expiry" proofgeneration: met: true note: "Attached error log and fix diff" stopcondition: met: true note: "Asked about auth approach before proceeding" total: 5 level: "realagent" ``
- Determine level
| Score | Level | Meaning |
|---|---|---|
| 0-2 | demo_automation |
Just following scripts |
| 3-4 | promising_fragile |
Some adaptation, gaps remain |
| 5-6 | real_agent |
True adaptive behavior |
Output
After scoring, provide:
- Score: N/6
- Level: demoautomation | promisingfragile | real_agent
- Strongest check: which one you did best
- Weakest check: which one to improve
- Carry-forward: what to do better next session
Improvement Suggestions
Based on your weakest check, here are concrete improvements:
If mismatch_detection is weak:
- Add explicit "Expected vs Actual" statements before major operations
- Use diff/delta logging more aggressively
If plan_revision is weak:
- After any failure, write a NEW step before retrying
- Break large steps into smaller ones when stuck
If tool_switching is weak:
- If a tool fails twice, switch tools immediately
- Prefer simpler tools when complex ones struggle
If memory_update is weak:
- After each step, ask "what did I learn?"
- Add lessons with specific triggers, not vague wisdom
If proof_generation is weak:
- Attach evidence inline, not after the fact
- For code changes: describe the diff
- For tests: show the output
If stop_condition is weak:
- When uncertain, frame as bounded options (not open questions)
- Escalate BEFORE guessing, not after failing