self-reflection

Use when you need to determine the accuracy of the answer.

You probably got most things right. This skill catches the few things you got wrong.

Quick Flow

text

flag claims → verify in parallel → fix → loop (optional) → done

How-To

1. Flag

Scan your previous output (or user-specified statements) for every factual claim. A factual claim must be small and verifiable — a number, a causal statement, an assertion, a relationship. Skip vague or subjective statements.

2. Verify

For each flagged claim, launch a subagent.

• Each subagent runs in an isolated, independent context — no access to chat history or prior tool results, context should be fair and unbiased.
• Verify in parallel: Fire all subagents simultaneously in a single message !

Every subagent's prompt template as below:

text

You are an independent auditor. Verify the accuracy of the small claim below.

Claim: [exact claim]
Context: [relevant context]

- Answer ONLY from your own tool calls. Do NOT rely on prior knowledge.
**Don't just web search — fetch full sources.** Search's preview are cherry-picked and incomplete.
- Match tool to source: read PDFs, use gh CLI for full issue/PR threads, use yt-dlp Skill for every video content

Output a brief audit result:
- Claim Accuracy: yes / partial / no
- Confidence: 0–10
If not accurate:
- Evidence: quote or summarize what was found
- Reasoning: 1–2 sentences

3. Fix

After all claims are verified, correct every wrong claim:

• Fix obvious errors directly.
• Back corrected statements with clear evidence.
• Strengthen weak statements by adding qualifiers or conditions.
• Clearly flag any unsupportable claims as unreliable.

4. Loop (Optional)

If the user needs a fully correct output (not just a /self-reflection), run the Loop workflow to iteratively verify and fix until the output is indisputably correct.

Start todowrite to track each iteration "Round":

text

R1: flag 5 claims
R1: verify batch 1 (3 subagents, claims 1–3)
R1: verify batch 2 (2 subagents, claims 4–5)
R1: fix statements
R2: flag 3 claims
R2: verify batch (3 subagents, claims 1–3)
R2: fix statements
R2: done — output is correct

Update items in your todo list anytime. Stop when the output is indisputably correct.