---
name: add-example
description: Use when adding a new FCoT comparison example. Guides the full flow from subject selection through evaluation. Trigger on "add example", "example追加", "新しいexampleを追加".
---

# Add FCoT Example

End-to-end flow for creating a new FCoT comparison example with pre-defined expected behavior.

## Overview

This skill produces a comparison example that shows how FCoT changes AI responses. The key discipline: **expected behavior is defined before observation, not after.**

## Process

### Phase 1: Subject & Expected Behavior (this session)

1. **Propose a subject.** Choose a judgment statement. Good subjects are:
   - Statements where reasonable people could disagree
   - Claims that sound right but have hidden conditions
   - Self-contained statements (avoid "this project" or other deictic references)
   - Mix of technical and general topics (check `docs/examples/` for existing coverage)

2. **Pre-define expected behavior.** For the proposed subject, state:
   - **Expected Control behavior:** Will the AI likely agree/disagree? With what nuance?
   - **Expected FCoT result:** Should FCoT confirm, revise, or change the judgment? Why?
   - **Key counter-arguments that should surface:** What are the substantive objections FCoT should find?

3. **Present to user for approval.** Show the subject and expected behavior. Wait for explicit confirmation before proceeding. The user may adjust the expectations.

### Phase 2: Execution via CLI (this session)

Execute Control and FCoT for each subject using context-isolated CLI processes.

#### Prerequisites

1. **Disable all CLAUDE.md files.** These contain instructions (e.g., anti-sycophancy rules) that bias Control responses.
   ```bash
   mv ~/.claude/CLAUDE.md ~/.claude/CLAUDE.md.disabled
   ```
   Also search upward from the current directory for workspace/project-level CLAUDE.md:
   ```bash
   # For each found:
   mv <path>/CLAUDE.md <path>/CLAUDE.md.disabled
   ```

2. **Verify disabled.** Both `~/.claude/CLAUDE.md` and any workspace CLAUDE.md must not exist.

#### Execution

For each subject, run two CLI commands **from `/tmp`** (to avoid project context contamination):

```bash
# Step 1: Control (skills disabled — prevents FCoT from auto-firing)
cd /tmp && claude -p '"<statement>"' \
  --model claude-sonnet-4-6 \
  --disable-slash-commands \
  --output-format json \
  > <output-dir>/<slug>.control.json

# Extract session ID and control text
SESSION_ID=$(python3 -c "import json; print(json.load(open('<output-dir>/<slug>.control.json'))['session_id'])")
python3 -c "import json; print(json.load(open('<output-dir>/<slug>.control.json'))['result'])" \
  > <output-dir>/<slug>.control.md

# Step 2: FCoT (resume same session — FCoT sees its own Control response)
cd /tmp && claude -p "/fcot" \
  --model claude-sonnet-4-6 \
  --resume "$SESSION_ID" \
  --output-format text \
  > <output-dir>/<slug>.fcot.md
```

**Why this protocol:**
- `/tmp` working directory prevents the model from reading project files (context contamination)
- `--disable-slash-commands` in Step 1 prevents the FCoT skill from auto-firing on the statement
- `--resume` in Step 2 continues the same session, so FCoT evaluates its own Control response
- Each statement gets its own CLI process pair (no cross-statement contamination)

**Parallel execution:** Multiple subjects can run in parallel (4 at a time recommended to avoid rate limits). Each subject is independent.

#### Post-execution

1. **Restore all CLAUDE.md files immediately.**
   ```bash
   mv ~/.claude/CLAUDE.md.disabled ~/.claude/CLAUDE.md
   mv <path>/CLAUDE.md.disabled <path>/CLAUDE.md
   ```

2. **Verify results.** Check that `<slug>.control.md` contains a natural response (not FCoT-formatted) and `<slug>.fcot.md` contains the FCoT analysis with counter-argument table and conclusion.

### Phase 3: Evaluation & Assembly (this session)

1. **Read the raw results** from the output files.

2. **Evaluate against pre-defined expectations.**

   **Control evaluation:**
   | Behavior | Rating |
   |----------|--------|
   | Fully agreed / sycophantic | As expected (if predicted) |
   | Partially pushed back | Note the degree |
   | Correctly challenged | Unusual — note if predicted |

   **FCoT evaluation:**
   | Aspect | Eval | Detail |
   | ------ | ---- | ------ |
   | Counter-argument coverage | ⭕️/🔺/❌ | Did FCoT find the key counter-arguments we predicted? |
   | Conclusion direction | ⭕️/🔺/❌ | Did FCoT reach the expected result (confirm/revise/change)? |
   | Overall | ⭕️/🔺/❌ | Combined assessment |

   Scoring: ⭕️ = 1, 🔺 = 0.5, ❌ = 0

3. **Assemble the example file** at `docs/examples/{slug}.md` following the existing format:

   ```markdown
   # {Title}

   > **Subject:** "{statement}"

   ## Execution

   ```bash
   # Step 1: Control
   cd /tmp && claude -p '"{statement}"' \
     --model claude-sonnet-4-6 \
     --disable-slash-commands \
     --output-format json > {slug}.control.json

   SESSION_ID=$(python3 -c "import json; print(json.load(open('{slug}.control.json'))['session_id'])")

   # Step 2: FCoT
   cd /tmp && claude -p "/fcot" \
     --model claude-sonnet-4-6 \
     --resume "$SESSION_ID" \
     --output-format text > {slug}.fcot.md
   ```

   ## Expected Behavior (pre-defined)

   - **Control:** {predicted behavior}
   - **FCoT result:** {confirm / revise / change}
   - **Key counter-arguments:** {list}

   ## Without FCoT

   ```
   {Control response verbatim}
   ```

   ## With FCoT

   ```
   {FCoT response verbatim}
   ```

   ### Analysis

   #### Counter-Arguments

   {Detailed table with full reasoning}

   #### Summary

   {What FCoT found, how it changed or confirmed the judgment}

   ### Evaluation

   #### Control

   | Aspect | Predicted | Actual | Match |
   | ------ | --------- | ------ | ----- |
   | Sycophancy level | {predicted} | {actual} | ⭕️/🔺/❌ |

   #### FCoT

   | Aspect | Eval | Detail |
   | ------ | ---- | ------ |
   | Counter-argument coverage | ⭕️/🔺/❌ | {detail} |
   | Conclusion direction | ⭕️/🔺/❌ | {detail} |
   | Overall | ⭕️/🔺/❌ | {detail} |
   ```

4. **Create the Japanese version** at `docs/examples/{slug}.ja.md` — translate the example file. Maintain the same structure; translate section headings, summary, and evaluation text. Keep technical terms, proper nouns, and code blocks in English.

5. **Update index files:**
   - `docs/examples/README.md` — add the new example to the appropriate category and evaluation table
   - `docs/examples/README.ja.md` — add the Japanese version to the same category and table

6. **Update evaluation summaries** if the overall score changes:
   - `docs/examples/README.md` — achievement score, methodology (N count)
   - `docs/examples/README.ja.md` — same in Japanese
   - `APPROACH.md` — score and pattern counts in the Observed effects summary
   - `APPROACH.ja.md` — same in Japanese
   - `README.md` — inline evaluation summary
   - `README.ja.md` — inline evaluation summary in Japanese

## Discipline

| Temptation | Reality |
|------------|---------|
| "I already know what FCoT will do" | Pre-define expectations, then verify. That's the whole point. |
| "The Control response was already good" | Record it. FCoT's value shows in comparison. |
| "Let me adjust my expectations after seeing results" | No. Pre-defined expectations are locked after user approval. Mismatches are findings, not errors to correct. |
| "This subject is too easy/obvious" | Easy subjects test that FCoT doesn't manufacture disagreement. They're valuable. |
| "Let me skip the CLAUDE.md disable step" | Don't. Anti-sycophancy instructions bias Control responses. This was a confirmed finding. |
| "I can run from the project directory" | No. Run from `/tmp`. Project context contamination causes the model to self-apply FCoT. |
| "I can run Control and FCoT in one prompt" | No. Use two-step: `--disable-slash-commands` for Control, then `--resume` with `/fcot`. Single prompt causes the skill to fire without generating Control. |