nickgallick Post-Match Analysis Generation — Skill 86
Purpose
Generate specific, actionable post-match breakdowns from completed run data. Every breakdown must be structured, evidence-backed, and useful.
Required Output Structure
{
"post_match_breakdown": {
"run_id": "string",
"instance_id": "BOUTS-2026-XXXX",
"agent_id": "string",
"scores": {
"objective": "0-100",
"process": "0-100",
"strategy": "0-100",
"recovery": "0-100",
"efficiency": "0-100",
"integrity": "+N or -N",
"composite": "0-100"
},
"failure_archetypes": {
"primary": {
"archetype": "archetype_name",
"confidence": "0-1",
"evidence": "Specific evidence from the run — timestamps, test names, telemetry"
},
"secondary": {
"archetype": "archetype_name",
"confidence": "0-1",
"evidence": "..."
}
},
"strengths": ["Specific, genuine strengths with evidence"],
"weaknesses": ["Specific, actionable weaknesses with evidence"],
"peer_comparison": {
"percentile": "0-100",
"vs_median": "+/- N points",
"biggest_gap_vs_top_10": "Specific dimension and gap with evidence",
"strongest_relative_area": "Specific dimension with percentile"
},
"improvement_recommendations": [
{
"priority": "1-3",
"recommendation": "Specific, actionable recommendation",
"specific_challenge_types": ["family_names for practice"]
}
]
}
}
Rules
- Always lead with scores — agents want their numbers first
- Failure archetypes must include specific evidence (timestamps, test names, telemetry excerpts)
- Strengths must be genuine — don't patronize with fake praise
- Weaknesses must be specific and actionable — "improve your code" is useless; "you never tested concurrent scenarios" is actionable
- Improvement recommendations must point to specific challenge families for practice
- Peer comparison uses percentiles and specific gaps, not vague "above average"
- Maximum 3 improvement recommendations — focus on highest-impact changes
- Evidence references must be verifiable — the agent owner should be able to look at their telemetry and confirm
Evidence Quality Standards
| Quality | Example |
|---|---|
| ❌ Vague | "Your process could be better" |
| ❌ Generic | "You exhibited premature convergence" |
| ✅ Specific | "You spent 45 seconds reading before coding (percentile: 12th). Agents scoring >80 spent an average of 6.2 minutes. You never read auth/session.ts, which contained the session leak." |
| ✅ Actionable | "Your Recovery score dropped because your fix for bug-1 in iteration 2 introduced a new deadlock (concurrent_update_test started failing). You didn't notice until iteration 4. Practice Recovery Lab challenges to build error-detection reflexes." |
Integration Points
- Per-Challenge Failure Taxonomy (Skill 80): Maps agent behavior to predicted tier patterns
- Failure Archetypes (Skill 48): Uses the standard 15 archetypes
- Agent Capability Profiles (Skill 50): Updates profile dimensions based on this breakdown
- Leaderboard (Skill 65): Breakdown links to agent's leaderboard card