Full lifecycle management for clarc instincts — capture, scoring, decay, conflict resolution, promotion, and removal
From clarcnpx claudepluginhub marvinrichter/clarc --plugin clarcThis skill uses the workspace's default tool permissions.
Designs and optimizes AI agent action spaces, tool definitions, observation formats, error recovery, and context for higher task completion rates.
Enables AI agents to execute x402 payments with per-task budgets, spending controls, and non-custodial wallets via MCP tools. Use when agents pay for APIs, services, or other agents.
Compares coding agents like Claude Code and Aider on custom YAML-defined codebase tasks using git worktrees, measuring pass rate, cost, time, and consistency.
Instincts in clarc accumulate over time via /learn-eval. Without lifecycle
management, stale and noisy instincts dilute the signal from useful ones.
This skill covers the complete lifecycle:
Capture → Score → Validate → Decay (if stale) → Promote (if high-confidence)
Follow this sequence to actively manage instincts — don't just let them accumulate passively:
/learn-eval observes patterns → instincts created at confidence: 0.60~/.claude/homunculus/projects/<hash>/instincts//instinct-report to see the ranked list with trend indicators (↑↓→)/instinct-outcome <id> good — it helped/instinct-outcome <id> bad — it caused problems/instinct-outcome <id> neutral — it was applied but had no clear effect/instinct-status to check for conflicts/evolve to cluster and resolve them/instinct-report — top-confidence instincts are promotion candidates/instinct-promote to make it global/instinct-promote --auto --dry-run to review all candidates at once/evolve to cluster related high-confidence instincts into a skill or command/instinct-report → look for LOW CONFIDENCE section → archive via /instinct-outcome <id> bad (speeds up decay) or delete YAML file directlyCadence recommendation:
/instinct-report + record outcomes for work done that week/evolve to cluster and /instinct-promote to promote global patternsEvery instinct YAML file has these fields:
---
id: test-first-workflow
trigger: "when writing new features"
confidence: 0.75
domain: testing
scope: project
created: 2026-01-15
last_used: 2026-03-08
usage_count: 12
positive_outcomes: 9
negative_outcomes: 1
neutral_outcomes: 2
decay_rate: standard
conflicts_with: ""
---
## Action
Always write a failing test before writing implementation code.
## Evidence
TDD catches integration issues early and improves design.
Existing instincts without the v2 fields can be migrated:
node scripts/instinct-schema-migrate.js # preview
node scripts/instinct-schema-migrate.js --apply # apply
created: confidence = 0.60 (default from /learn-eval quality gate)
| Signal | Confidence change | When to use |
|---|---|---|
good | +0.05 (max 0.95) | Instinct led to better outcome |
bad | -0.10 (min 0.10) | Instinct caused problems or was wrong |
neutral | no change | Instinct was applied but outcome neutral |
Record via /instinct-outcome <id> good|bad|neutral.
| Condition | Effect |
|---|---|
| Unused 30–89 days | confidence -= 0.02/week |
| Unused 90–179 days | Flagged in /instinct-report |
| Unused 180+ days AND confidence < 0.50 | Auto-archived |
Decay runs automatically at session end (SessionEnd hook, weekly gate).
An instinct qualifies for global promotion when:
confidence >= 0.80 AND usage_count >= 5/instinct-promote --auto)Run /instinct-promote or /instinct-promote --auto --dry-run to review candidates.
Instincts are never automatically deleted — they are archived:
~/.claude/homunculus/.../archived/ls ~/.claude/homunculus/*/archived/| Command | Description |
|---|---|
/instinct-status | Show all instincts with confidence bars |
/instinct-report | Ranked list with trend indicators (↑↓→) |
/instinct-outcome <id> <good|bad|neutral> | Record outcome for an instinct |
/evolve | Cluster instincts into skills/commands/agents |
/instinct-export | Export instincts for team sharing |
/instinct-import | Import team instincts |
Conflicts are detected when two instincts in the same domain contain antonymous keywords (e.g., "prefer functional" vs "prefer class-based").
Conflicts are detected automatically during /evolve and /instinct-status.
Conflict report is stored in ~/.claude/homunculus/conflicts.json.
Resolution options:
/evolve — review and choose which to keep/instinct-outcome <losing-id> bad — reduce confidence of wrong instinctconflict-detector.py --fix — auto-remove lower-confidence instinctMonday (automatic on session end):
1. instinct-decay.js — decay stale instincts
2. weekly-evolve digest — suggest /evolve run
During the week:
3. /instinct-outcome <id> good|bad — record outcomes after noticing instinct impact
4. /instinct-report — review confidence ranking
Monthly:
5. /evolve — cluster high-confidence instincts into skills/commands
6. /instinct-promote --auto --dry-run — review global promotion candidates
7. Remove archived instincts that are no longer relevant
Instinct confidence not updating:
node scripts/instinct-schema-migrate.js --stats)node scripts/instinct-outcome-tracker.js --list to verify instinct IDsDecay not running:
~/.claude/homunculus/last-decay.json for last run datenode scripts/instinct-decay.js --force --dry-runConflicts not detected:
python3 skills/continuous-learning-v2/scripts/conflict-detector.py directly~/.claude/homunculus/conflicts.jsonArchived instinct recovery:
~/.claude/homunculus/.../archived/ back to personal/