Pattern Capture

Detect repetitive feedback across sessions and convert it into the right enforcement artifact — a memory entry, a validation hook, an enforcement pattern, or a standalone skill.

The Problem This Solves

Users give the same corrections repeatedly:

• "Don't mock the database in tests" (session 1, 3, 7, 12)
• "Use jq explicit syntax, not shorthand" (session 2, 5, 8)
• "Check the build before claiming it works" (session 1, 4, 6, 9, 11)

Each correction costs the user time and erodes trust. The DRY principle applies to prompting: if you've said it twice, it should be automated.

When to Use

• User says "I keep telling you..." or "Again, don't..."
• You notice you're receiving the same type of correction
• At end of session, to audit what feedback was given
• Proactively, when continuous-learning detects user_corrections patterns
• User explicitly asks to capture a pattern or DRY their prompting

Process

Step 1: Gather Evidence

Collect instances of the repeated pattern. Sources (check in order):

1. Memory files (fastest)
   → Grep pattern="<keyword>" path="<memory_dir>" glob="*.md"
   → Look for feedback-type memories

2. Session transcripts (if CLAUDE_TRANSCRIPT_PATH is set)
   → Grep for user corrections: "no", "don't", "stop", "again", "I said"
   → Count occurrences of similar corrections

3. Spotless archives (if available, cross-session)
   → Search conversation history for repeated correction patterns

4. User report (always valid)
   → User says "I keep having to tell you X" = sufficient evidence

Minimum evidence threshold: 2 independent instances (same correction, different contexts). A single user report of "I keep telling you" counts as meeting threshold — trust the user's observation.

Step 2: Classify the Pattern

Every repeated pattern maps to exactly ONE artifact type. Use this decision tree. Evaluate branches top-to-bottom; stop at the FIRST match.

Is the pattern about WHEN to do something?
  YES → Is it about tool/command selection?
    YES → MEMORY (feedback type)
    NO  → Is it about workflow sequencing?
      YES → ENFORCEMENT PATTERN (add to existing workflow skill)
      NO  → MEMORY (feedback type)
  NO  →
Is the pattern about HOW to do something?
  YES → Is it a single rule (< 3 sentences)?
    YES → Is it project-specific?
      YES → MEMORY (project type)
      NO  → MEMORY (feedback type)
    NO  → Does it require multi-step verification?
      YES → VALIDATION HOOK
      NO  → Is it reusable across projects?
        YES → SKILL (learned skill)
        NO  → MEMORY (feedback type)
  NO  →
Is the pattern about WHAT NOT to do?
  YES → Can the violation be detected programmatically?
    YES → VALIDATION HOOK
    NO  → RED FLAG (add to existing skill's Red Flags table)
  NO  →
Default → MEMORY (feedback type)

Artifact Type Reference

Step 3: Generate the Artifact

Based on classification, generate the appropriate artifact:

For MEMORY entries

---
name: feedback_<descriptive-slug>
description: <one-line description specific enough to match in future>
type: feedback
---

<The rule, stated clearly>

**Context:** <Why this matters — what went wrong when it was violated>
**Source:** <How this was discovered — "corrected N times" or "user reported">

Write to memory directory and update MEMORY.md index.

For ENFORCEMENT PATTERNS (added to existing skills)

Identify which skill the pattern belongs to, then add the appropriate enforcement element:

Iron Law (for high-drift actions the agent rationalizes skipping):

<EXTREMELY-IMPORTANT>
**[RULE IN ALL CAPS]. This is not negotiable.**

[One sentence explaining concrete user harm if violated.]
</EXTREMELY-IMPORTANT>

Rationalization Table entry (for patterns where the agent makes excuses):

| Excuse | Reality | Do Instead |
|--------|---------|------------|
| "<exact excuse the agent generates>" | "<why this is wrong>" | "<correct action>" |

Red Flag entry (for observable wrong actions):

| Action | Why It's Wrong | Do Instead |
|--------|----------------|------------|
| "<observable behavior>" | "<concrete harm>" | "<correct alternative>" |

For VALIDATION HOOKS

Generate a PreToolUse or PostToolUse hook:

// hooks/<hook-name>.ts
// Pattern: <description of what this catches>
// Source: User corrected this N times across sessions

export default {
  event: "PreToolUse",  // or PostToolUse
  name: "<tool-name>",  // e.g., "Bash", "Write", "Edit"
  async handler({ input }) {
    // Detection logic
    const violation = /* check for the anti-pattern */;
    if (violation) {
      return {
        decision: "block",  // or "ask"
        reason: "<explanation of why this is blocked>"
      };
    }
    return { decision: "approve" };
  }
};

For LEARNED SKILLS

Delegate to skill-creator:

Skill(skill="skill-creator", args="Create skill from captured pattern: <description>")

Provide the skill-creator with:

• Pattern description and evidence
• Example correct/incorrect behaviors
• Suggested enforcement level (from classification)

Step 4: Verify Integration

After generating the artifact, verify it's properly integrated:

Step 5: Report

Output a summary:

## Pattern Captured

**Pattern:** <one-line description>
**Evidence:** <N instances across M sessions>
**Classification:** <artifact type>
**Artifact:** <file path or location>
**Prevention:** <how this prevents future repetition>

Proactive Detection

When invoked without a specific pattern (e.g., "find repeated feedback"), scan all available sources:

1. Read all feedback-type memory files
2. Search session transcripts for correction language:
   • "no,? (don't|stop|not|never|instead|again)"
   • "I (already|just) (told|said|asked|mentioned)"
   • "(wrong|incorrect|that's not|not what I)"
3. Group similar corrections by semantic similarity
4. For each group with 2+ instances, run the classification tree
5. Present findings to user for confirmation before generating artifacts

Iron Laws

Fabricating patterns the user hasn't actually repeated leads to over-engineered enforcement that constrains legitimate behavior. Every artifact must trace to specific observed instances. </EXTREMELY-IMPORTANT>

Adding a Red Flag or Iron Law without understanding the skill's existing enforcement creates conflicts, duplicates, and confusion. Read the entire SKILL.md before modifying it. </EXTREMELY-IMPORTANT>

Red Flags - STOP If You Catch Yourself:

Rationalization Table

Integration Points

Examples

Example 1: Simple Behavioral Rule → Memory

Evidence: User said "stop summarizing at the end" in 3 sessions Classification: WHEN to do something → tool selection? No → workflow? No → MEMORY (feedback) Artifact:

---
name: feedback_no_trailing_summaries
description: Do not add summary paragraphs after completing a task — user reads diffs directly
type: feedback
---
Do not summarize what you just did at the end of responses. The user reads diffs and tool output directly.

**Context:** Trailing summaries waste time and feel patronizing to experienced users.
**Source:** Corrected 3 times across sessions.

Example 2: Build Verification → Enforcement Pattern

Evidence: Agent claimed "build passes" without running build in 4 sessions Classification: HOW → multi-step verification? Yes → but dev-verify already handles this → ENFORCEMENT PATTERN Artifact: Add to dev-verify's Red Flags table:

| "Build should still pass from earlier" | Earlier results are stale — any code change invalidates them | Run `npm run build` fresh RIGHT NOW |

Example 3: No Mocks in Integration Tests → Validation Hook

Evidence: Agent used jest.mock() in integration test files 3 times Classification: WHAT NOT TO DO → programmatically detectable? Yes (grep for jest.mock in tests/integration/) → VALIDATION HOOK Artifact: PostToolUse hook on Write/Edit that warns when jest.mock appears in integration test files.

Example 4: Project-Specific Convention → Project Memory

Evidence: User corrected jq syntax 3 times — container uses jq 1.6, not 1.7 Classification: HOW → single rule → project-specific → MEMORY (project) Artifact:

---
name: project_jq_16_explicit_syntax
description: NanoClaw container runs jq 1.6 which requires explicit field syntax, not shorthand
type: project
---
Container runs jq 1.6. Always use explicit syntax: `{title: .title}` not `{title, location: expr}`.

**Applies to:** NanoClaw container agent, any jq commands in container scripts.
**Context:** jq 1.6 does not support mixing shorthand + explicit fields. Causes silent failures.
**Source:** Corrected 3 times across sessions.

Example 5: Complex Debugging Procedure → Learned Skill

Evidence: User walked through the same pixi debugging steps in 3 sessions Classification: HOW → single rule? No (5+ steps) → multi-step verification? No → reusable? Yes → SKILL Artifact: Delegate to skill-creator with the debugging steps as input.

References

• Classification quick reference: references/classification-guide.md — fast-path matrix and enforcement strength ladder
• Artifact templates: references/artifact-templates.md — copy-paste templates for all artifact types with formatting guidance
• Enforcement checklist: references/enforcement-checklist.md — full 12-pattern reference (when adding enforcement to existing skills)
• Continuous-learning: ../continuous-learning/SKILL.md — upstream pattern detection (feeds into this skill)
• Skill-creator: ../skill-creator/SKILL.md — downstream skill generation (this skill delegates to it for learned skills)