mine-patterns

Mine Patterns

Analyze conversation history from the ai-chat-browser MCP to discover automatable patterns.

Two Modes

Incremental (from SessionStart queue)

Process specific sessions queued by the SessionEnd hook.

Bulk Sweep (from /mine-patterns command)

Scan all indexed sessions or a filtered subset.

Process

Step 1: FTS Pre-Filter

Query the ai-chat-browser MCP search_messages tool for high-signal messages.

Correction signals (user correcting the assistant):

• Query: "no don't" OR "wrong" OR "stop" OR "that's not" OR "I said"

Confirmation signals (user validating an approach):

• Query: "perfect" OR "exactly" OR "yes that's right" OR "keep doing"

Instruction signals (user giving persistent directives):

• Query: "always" OR "never" OR "make sure" OR "from now on" OR "remember to"

For each query, use limit: 50 and context_messages: 3.

Step 2: Context Extraction

For each hit from Step 1, use get_conversation with the search parameter to retrieve the matching message with ±3 messages of surrounding context.

Step 3: Pattern Classification

For each high-signal hit with context, classify the pattern:

Pattern Type	Signal	Proposed Output
correction	User says "no/don't/wrong" then provides right way	feedback memory or hook rule
repeated_workflow	Same tool call sequence across 3+ sessions	Skill candidate
repeated_delegation	Same autonomous task requested 3+ times	Agent candidate
validated_approach	User confirms with positive signals	feedback memory
recurring_instruction	Same instruction in 3+ sessions	Rule or CLAUDE.md update

Step 4: Confidence Scoring

Assign 0-100 confidence based on:

• Frequency (40%): How many sessions show this pattern? (1 session = low, 3+ = high)
• Clarity (30%): How explicit is the user's correction/instruction? (ambiguous = low, direct quote = high)
• Actionability (30%): Can this be converted to a concrete automation? (vague preference = low, specific rule = high)

Step 5: LLM-Based Dedup

Before creating a nomination:

1. Keyword pre-filter: Use Grep to search existing files:

   • ~/.claude/projects/<project>/memory/knowledge/ (existing memories)
   • ~/.agents/rules/ (existing rules)
   • ~/.agents/skills/ (existing skills)
   • ~/Projects/patryks-treadmill-claude-plugins/plugins/*/skills/ (treadmill skills)

2. LLM similarity check: If keyword matches found, evaluate:

   • "Given existing automation X, is candidate Y a duplicate, an extension, or genuinely new?"
   • Duplicate → discard
   • Extension → nominate as update to existing (include diff)
   • New → nominate as creation

Step 6: Write Nominations

Auto-merge (confidence 80+, memories and rules only):

• Write directly to knowledge/ or ~/.agents/rules/
• Update MEMORY.md
• Log as auto-merged for awareness during /review-nominations

Nominate (confidence 50+, or any skill/agent/hook):

• Write to ~/.claude/projects/<project>/memory/nominations/
• File format:

---
id: candidate_<8-char-hash>
confidence: <score>
type: feedback | skill | agent | hook | rule | claude_md
target: new | update:<existing-file-path>
source_sessions:
  - <session-id-1>
  - <session-id-2>
discovered_at: <ISO-8601>
pattern_type: correction | repeated_workflow | repeated_delegation | validated_approach | recurring_instruction
auto_merged: false
---

# <Title>

## Evidence

<Relevant message excerpts from source sessions>

## Proposed Content

<The actual memory/skill/hook content to create or diff to apply>

Discard (confidence < 50): Skip silently.

Step 7: Report

After processing, summarize:

• Total sessions analyzed
• Patterns found by category
• Auto-merged items (with file paths)
• Nominations created (pending review)
• Discarded (count only)

Suggest running /review-nominations if there are pending items.