counselors-log

name	counselors-log
description	Analyze observation and learning queues, cluster patterns, propose instincts with confidence scores, and suggest capability evolution when patterns cluster.
tags	["learning","self-improvement","patterns","evolution"]

/counselors-log — Pattern Analysis + Instinct Management

Pipeline role: Observation analyzer. Reads observations-queue.json and proposes instincts. For full cross-source analysis (all 5 learning sources including K-LEAN and knowledge graph), use /borg-assimilate instead.

Analyze captured observations and learnings. Cluster patterns by domain. Propose instincts (reusable behavioral rules) with confidence scoring. Suggest new skills/agents when 5+ instincts cluster.

Arguments

• /counselors-log — Full analysis of both queues
• /counselors-log --instincts — List current instincts and their confidence levels
• /counselors-log --evolve — Check for evolution threshold (5+ instincts in a domain)

Step 1: Read Queues

Read both queue files:

• ~/.claude/cache/learnings-queue.json — corrections, preferences, positive feedback
• ~/.claude/cache/observations-queue.json — tool patterns, error-fix patterns, style preferences

If both are empty or missing, output:

No observations or learnings to analyze. Work normally and the hooks will capture patterns.

Then stop.

Step 2: Read Existing Instincts

Read all files in .claude/instincts/ (project root) if the directory exists. These are previously approved instincts that should not be duplicated.

Step 3: Group by Domain

Categorize each queue entry into one of these domains:

• code-style — formatting, naming, syntax preferences
• testing — test patterns, TDD practices, coverage
• debugging — investigation patterns, common fixes
• tooling — tool usage patterns, workflow preferences
• architecture — design patterns, structure decisions
• deployment — build, deploy, release patterns
• communication — response style, verbosity, explanation depth
• security — security-related patterns and practices

Step 4: Cluster Analysis

For each domain with 3+ entries:

1. Identify the common theme connecting the entries
2. Propose an instinct — a reusable behavioral rule:

### Proposed Instinct: {name}

**Domain**: {domain}
**Trigger**: "{when this situation occurs}"
**Action**: "{do this}"
**Confidence**: {0.3 = tentative, 0.5 = emerging, 0.7 = established, 0.9 = near certain}

**Evidence** ({N} observations):
- {observation 1 summary}
- {observation 2 summary}
- {observation 3 summary}

**Contradicts**: {existing instinct name, if any — flag for resolution}

3. Check against existing instincts for:
   • Duplicates: Same trigger + action → skip, note "already captured"
   • Contradictions: Same trigger, different action → flag for user resolution
   • Refinements: Same domain, more specific → propose as refinement of existing

Step 5: Present for Approval

Show all proposed instincts grouped by domain. For each:

• Accept → write to .claude/instincts/{domain}-{name}.md
• Reject → discard
• Edit → user provides revised wording

NEVER auto-apply. Always wait for explicit approval.

Step 6: Evolution Check

After processing, count instincts per domain (existing + newly approved):

If any domain has 5+ instincts:

COUNSELOR'S LOG: {N} instincts detected in "{domain}" domain.

Pattern: {summarize the common thread across instincts}

Suggested evolution:
- [ ] Create a "{domain}-protocol" skill encoding these patterns
- [ ] Create a "{domain}-reviewer" agent with domain expertise
- [ ] Add a "{domain}" section to CLAUDE.md (costs {N} instruction budget lines)

Which evolution, if any? [1/2/3/none]

If user selects an evolution, describe what would be created but do NOT auto-create it. Instead, output the recommended file contents for user review.

Step 7: Clean Up Queues

Remove processed entries from both queue files:

• Entries that led to approved instincts → remove
• Entries that were rejected → remove
• Entries that didn't cluster (fewer than 3 in domain) → leave for future analysis

Step 8: Summary

## Counselor's Log — {date}

**Analyzed**: {N} observations + {N} learnings
**Domains with patterns**: {list domains with 3+ entries}
**Instincts proposed**: {N}
**Instincts approved**: {N}
**Evolution threshold**: {domain at 5+ or "none"}

### Active Instincts
| Domain | Count | Newest |
|--------|-------|--------|
| {domain} | {N} | {instinct name} |

Instinct File Format

When writing instinct files to .claude/instincts/:

---
domain: {domain}
confidence: {0.3-0.9}
created: {ISO date}
evidence_count: {N}
---

# {Instinct Name}

**Trigger**: {when condition}
**Action**: {what to do}

## Evidence
- {observation summaries}

Notes

• Instincts are NOT injected into CLAUDE.md (budget protection)
• They live in .claude/instincts/ and are consulted on demand
• /reflect also checks instincts for consistency when processing learnings
• Token cost: ~2,000-3,000 per invocation
• Always-on overhead: 0