Self Evolve

ultrathink.

Wire Mode: Apply Self-Improvement Proposals

When scope is (none), or when pending proposals exist and user says "ok":

Fetching proposals:

python3 .claude/hooks/self-improve/self_improve_db.py resolve list

Step 0: Consolidate before showing

User attention is finite — showing 12 proposals when 4 are duplicates, 3 already exist in rules, and 2 are noise wastes their focus on what matters. Before presenting anything, read the existing config (.claude/rules/, CLAUDE.md, CLAUDE.local.md, active skills) to understand what's already there. Then filter: proposals already covered by existing rules, duplicates of each other, and low-value noise that wouldn't improve agent behavior. Merge proposals addressing the same pattern into one entry with the clearest rationale.

Tell the user what you filtered: "Showing X of Y proposals (Z filtered)."

After user finishes reviewing visible proposals, batch-reject the filtered ones:

python3 .claude/hooks/self-improve/self_improve_db.py resolve <id1>,<id2>,... rejected

For each shown proposal, ask the user to approve or reject.

Step 1: Classify — where does it belong?

~80% of proposals become rules — the pipeline detects behavioral corrections, and corrections are rules by definition.

Step 2: Read existing content, check for overlap

Read all files in the target location. Scan for semantic overlap — does an existing rule/section already cover this?

Step 3: Place — merge or add

• If overlap exists: merge into existing entry — expand/refine/strengthen, don't create near-duplicates
• If distinct: add under the most relevant section header

Writing quality — proposals describe specific incidents, but rules must work for the general case:

1. Generalize from the incident. The proposal says "Claude changed the pass dot when user meant N/A dot." The rule should address the class: "When multiple UI elements match an ambiguous reference, confirm which one before changing any." Don't encode the specific incident — encode the pattern.
2. Ground in reasoning, not commands. "ALWAYS confirm ambiguous references" is brittle — the agent doesn't know when it applies. "Because acting on the wrong target wastes a correction cycle, confirm which element when multiple candidates match" gives the agent judgment for edge cases.
3. Trim while merging. When adding to an existing rule file, read what's already there. If existing entries are redundant with the new content or with each other, consolidate. The goal after merging is a tighter file, not a longer one.
4. Watch for accumulation. If 3+ proposals have landed in the same file or section, that's a restructuring signal — the section may need splitting, or the underlying skill needs fixing instead of piling on more rules.
5. Think from the agent's perspective. Before finalizing, ask: "If I were an agent reading this rule for the first time with no context about the incident, would I understand when and why to apply it?" If the answer requires knowledge of the original proposal, rewrite.
6. Re-read with fresh eyes. After merging, re-read the entire section (not just your addition). Does the new content flow with the existing entries, or does it feel bolted on? Revise for coherence.