From superpowers-plus
Generates context-free prompts to reframe analysis when stuck, reducing anchoring bias. Experimental — ~20% false positive rate, verify outputs manually.
How this skill is triggered — by the user, by Claude, or both
Slash command
/superpowers-plus:experimental-self-promptingThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
> **Wrong skill?** Getting unstuck → `think-twice`. Research → `perplexity-research`. Brainstorming → `brainstorming`.
Wrong skill? Getting unstuck →
think-twice. Research →perplexity-research. Brainstorming →brainstorming.WARNING: This skill is EXPERIMENTAL. It has been validated in a controlled experiment but is NOT production-ready. Expect ~20% false positive rate. ALWAYS verify outputs manually before acting on findings.
Winner: Condition B (Reframe-Self) - Write prompt, answer yourself (no external model)
| Condition | VH | HR | Avg VH/Round | HR Rate |
|---|---|---|---|---|
| A: Direct | 19 | 1 | 3.8 | 20% |
| B: Reframe-Self | 21 | 1 | 4.2 | 20% ← WINNER |
| C: Direct-External | 23 | 4 | 4.6 | 80% |
| D: Reframe-External | 18 | 6 | 3.6 | 100% ← WORST |
Key Insight: Reframing helps Claude (+10% VH), but HURTS external models (+400% HR).
Condition D (reframe + Gemini) had 100% hallucination rate.
Every single round had at least one false positive. The detailed prompts give external models more rope to hallucinate confidently.
Even with Claude answering its own prompts, expect 1 in 5 findings to be wrong. Always verify findings before acting on them.
This skill was validated on 5 genesis-tools projects:
It may not generalize to other codebases.
| Trigger | Description |
|---|---|
| Complex system review | Multi-component systems with alignment concerns |
| Adversarial analysis | Looking for gaming vulnerabilities or edge cases |
| Independent verification | Verify claims from external sources (Gemini, GPT) |
| Pre-commit review | Final check before major commits |
Explicit invocation required:
Use the experimental-self-prompting skill to analyze [system]
Create a context-free prompt that any engineer could pick up cold:
You are an expert [ROLE] performing [TASK TYPE] on [SYSTEM].
## CONTEXT
[Explain the system, its components, and their relationships]
## THE PROBLEM
[What misalignment/issue pattern you're looking for]
## YOUR TASK
[Specific things to check]
## VERIFICATION REQUIREMENTS
For EACH finding:
1. State the claim
2. Cite exact file and line number
3. Show evidence (grep/code)
4. Categorize: VERIFIED | FALSE POSITIVE | NEEDS INVESTIGATION
## FILES TO EXAMINE
[List specific files with full paths]
Focus on ACTIONABLE findings with EVIDENCE. No speculation.
Treat the prompt as if you've never seen the code before. Answer systematically.
CRITICAL: Never trust findings without verification.
For each finding:
Create summary with:
This skill will be promoted to production when:
superpowers-plus/docs/plans/experiment-results-v2/superpowers-plus/docs/plans/experiment-results-v2/STATISTICAL_ANALYSIS.mdsuperpowers-plus/docs/SKILL_COMPARISON_self-prompting_vs_think-twice.md# Generate a context-free prompt for fresh analysis
echo "Analyze this codebase for [specific concern].
Constraints: [list constraints].
Output format: [specify format].
Do NOT reference prior analysis." > /tmp/self-prompt.md
npx claudepluginhub bordenet/superpowers-plus --plugin superpowers-plusGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.