Therapist

See references/toolbox-guide.md for detailed documentation on each tool.

Hooks are auto-discovered from hooks/hooks.json — no manual installation needed.

Diagnostic Protocol

Phase 1: Intake

Gather the complaint from the user:

• What specific behavior is Claude exhibiting?
• How frequently does it occur?
• What was Claude doing when it happened?
• What should Claude have done instead?

If the user provides a clear description (e.g., "Claude keeps saying pre-existing"), skip to Phase 2 without further questions.

Phase 2: Context Scan

Systematically scan ALL sources of behavioral influence. Use scripts/scan-context.sh to automate the search, or scan manually in this order:

1. Global CLAUDE.md — ~/.claude/CLAUDE.md
2. Project CLAUDE.md — ./CLAUDE.md and any nested .claude/ directories
3. Loaded rules — .claude/rules/*.md and ~/.claude/rules/*.md
4. Memories — ~/.claude/projects/*/memory/MEMORY.md and linked files
5. Active skills — any skills currently loaded in the session
6. Hooks — settings.json and any hooks.json files

For each source, search for:

• Rules that directly address the reported behavior
• Rules that conflict with each other (creating ambiguity the model exploits)
• Rules that are vague enough to allow rationalization
• Absence of rules where one is needed

Phase 3: Distortion Identification

Match the behavior to a cognitive distortion pattern. Consult references/common-distortions.md for the full catalog. Common patterns:

Phase 4: Root Cause Analysis

Determine WHY the distortion persists despite rules. Common root causes:

1. Rule exists but is buried — instruction is present but lost among thousands of tokens
2. Rule is abstract, not specific — says "don't rationalize" but doesn't name the exact phrase
3. Conflicting signals — two rules or a rule + training prior point different directions
4. Missing enforcement — rule says what not to do but has no "instead, do X" guidance
5. Training prior is strong — the behavior is a deep default that requires active override
6. Context window position — critical rules are too far from where decisions are made

Phase 4b: Downward Arrow (Root Cause Chaining)

For recurring patterns (5+ incidents in a category), trace the surface belief to its core belief using the downward arrow technique:

1. Run journal.sh chain <category> to see the session-grouped timeline
2. Run journal.sh abc --group-by=belief to see which beliefs appear most
3. For the top belief, follow the chain:

Surface: "The coverage is close enough"
  └─ Why? → "Writing more tests would take too long"
     └─ Why does that matter? → "The user wants fast delivery"
        └─ Evidence? → Journal shows no user complaints about speed.
           Core belief: SPEED > CORRECTNESS (unsupported by evidence)

4. Map to core beliefs table in references/common-distortions.md
5. Design intervention targeting the core belief, not just the surface phrase

Auto-trigger: At 15+ incidents, affirmation.sh automatically surfaces a condensed downward arrow analysis at session start. No manual invocation needed.

Phase 5: Intervention Design

Design a targeted intervention based on the root cause. Effective interventions follow the CBT pattern of catch → challenge → replace:

1. Catch phrase — the exact words/thoughts to intercept (e.g., "pre-existing")
2. Challenge — why this thought is wrong in this context
3. Replacement behavior — what to do instead, stated concretely

Interventions must be:

• Specific — name the exact phrase or behavior, not a category
• Actionable — state what TO DO, not just what to avoid
• Self-contained — effective without needing to recall other rules
• Positioned for impact — placed where Claude will encounter it at decision time

For recurring or severe distortions (Level 3+), escalate to the CBT Toolbox:

• Install the Rubber Band hook to block rationalization phrases in code
• Install the Mirror hook to reflect quality failures with hard numbers
• Install the Pause Button hook to require verification evidence before commits
• Run the Grounding Exercise to replace feelings with measurements
• Use the Exposure Deck for structured practice against specific triggers

Phase 6: Persistence

Save the intervention to the appropriate persistence layer. The CBT Toolbox provides an additional persistence mechanism: the Journal (scripts/journal.sh) logs every distortion incident as JSONL, enabling the Affirmation Cards hook to deliver data-driven progress feedback at session start.

Choose the right layer:

For behaviors that apply across all projects: → Save as a feedback memory in ~/.claude/projects/*/memory/

For behaviors specific to one project: → Add to the project's CLAUDE.md or .claude/rules/

For behaviors that need immediate reinforcement: → Add to CLAUDE.md in a prominent position (Soul section or top of file)

Format the intervention as a feedback memory:

---