improve

Part of the ACED plugin — see that page for install instructions.

Trigger: ACED evals are failing and the user wants to diagnose why and get specific improvement proposals

Diagnoses failing eval cases and proposes targeted edits to the target agent configuration.

What it does

1. Loads the node’s eval.md (found through the SDD spec tree), the target agent configuration, and the most recent results file under .agents/aced/results/<target-slug>/ (runs run first if none exists).
2. Collects all scenarios where pass: false and reads each failing scenario in the frozen .feature.
3. Groups failures by pattern: trigger false-positive/negative, missing step, ambiguous rule, conflicting instruction, scope creep, or description mismatch.
4. Proposes a concrete before/after diff per pattern (e.g. rewrite description: for trigger issues, make a step more prominent for missing steps, add precedence rules for conflicts).
5. Shows all proposed edits and asks for approval before writing anything.
6. After approval, applies the edits, then automatically runs compare (previous git revision vs. current working tree) to confirm improvement without regression.

Role as ACED impl-producer

When dispatched by the SDD conductor (with define-agent) against a frozen .feature, this skill is the impl-producer for agent-config domains — it also authors or refreshes the eval suite (the frozen .feature’s scenarios plus the node’s eval.md run policy) the impl-judge runs, since the judge never authors evals itself.

If no clear fix exists

For inherent non-determinism (high score variance across similar cases), it recommends adding more specific examples, lowering the threshold for that layer, or splitting the agent configuration into two narrower ones — never removing test cases to force a pass.