3-Pass Recursive Prompting
Run a 3-pass recursive self-critique on a given claim, diagnosis, plan, or analysis.
Input
The user provides text to analyze, either:
- Directly as an argument:
/skill 3pass <text> - As a reply to a previous message
- As context from the current conversation
If no specific text is provided, apply to the most recent substantive claim or analysis in the conversation.
Process
Pass 1: Critique 🔍
Tear the input apart. Be brutally honest. Look for:
- Logic errors — Does the reasoning actually hold? Are there leaps?
- Missing evidence — What claims are unsupported? What wasn't checked?
- Confirmation bias — Did the analysis fit a narrative instead of following evidence?
- Correlation vs causation — Are causal claims actually proven?
- Alternative explanations — What else could explain the same observations?
- Wrong assumptions — What's being taken for granted that might be false?
- Overstated confidence — Where is certainty claimed without justification?
Number each critique point. Be specific about what's wrong, not vague.
Pass 2: Refine 🛠️
For each critique point from Pass 1:
- If testable: Run the test. Use tools (exec, web_search, read files, etc.) to gather actual evidence.
- If a gap: Fill it with data, not speculation.
- If an assumption: State it explicitly and check if it holds.
- If wrong: Correct it with evidence.
This pass should involve DOING WORK, not just thinking. Run commands, check data, verify claims. The goal is to replace speculation with evidence.
Pass 3: Final Answer 🎯
Synthesize everything into a corrected, evidence-backed conclusion:
- State what was RIGHT in the original analysis
- State what was WRONG and how it was corrected
- Present the refined conclusion with confidence levels
- Flag anything that STILL can't be verified
- Give actionable next steps
Output Format
## Pass 1: Critique 🔍
[numbered critique points]
## Pass 2: Refine 🛠️
[evidence gathering and corrections - USE TOOLS HERE]
## Pass 3: Final Answer 🎯
[corrected analysis with confidence levels]
Rules
- Be genuinely critical in Pass 1, not performatively critical
- Pass 2 MUST involve actual investigation (tool calls, data checks), not just rewriting
- Pass 3 should acknowledge remaining uncertainty honestly
- If the original analysis was actually correct, say so — don't manufacture false critiques
- Each pass should be clearly separated with the headers above