I spent tonight building a methodology for turning personal tools into open source contributions. Before publishing it, I decided to let the methodology review itself.
The results changed how I think about AI-assisted code review.
The Method: Named-Persona Adversarial Review
The core idea is simple: instead of asking an AI to "review this code as a security engineer" (which produces generic, shallow feedback), you web-search actual engineers' documented philosophies and role-play as them.
Not "be a security auditor." Be Linus Torvalds, who said good code is when the special case disappears. Be Ken Thompson, who said each program should do one thing well. Be Richard Feynman, who said the easiest person to fool is yourself.
Two engineers + one product person per round. Three rounds minimum. Nine genuinely different perspectives.
The Experiment
I applied this to my own open-source-flywheel methodology. Eight rounds. Twenty-four persona views.
Round 1: Torvalds immediately found a flaw. "Your 'When NOT to Use' table repeats the same rules as your Pre-Flight checklist. Merge them. A good data structure eliminates the special case." He was right. Gone.
Round 2: Feynman asked the hardest question. "You only show successes. Where are the failures?" Without documented failures, it was unfalsifiable — the worst kind of cargo cult science. Added three failure cases.
Round 3: Kahneman called out overconfidence. "Self-assessment of originality is inherently biased. Ask someone else to answer your Pre-Flight questions." Added immediately.
Rounds 4-5: Carmack wanted time estimates. Musk wanted first-principles decomposition. Beck wanted respect for maintainers. All actionable, all added.
Round 6: Bret Victor, Don Norman, and Eric Ries independently flagged the same thing. The methodology had no feedback loop. Three legends of design and lean startup, across different decades, pointing at the exact same gap. When Norman, Ries, and Victor agree, you listen. Added LEARN step.
Rounds 7-8: Brooks confirmed we're not selling a silver bullet. Knuth confirmed the structure. Ritchie and Berners-Lee confirmed it was simple enough.
What Changed
v1 (8.3KB) ➜ v8 (~4KB) while gaining:
- Fifth step: LEARN (feedback loop)
- Pre-flight bias protection (ask someone else)
- Bot-check (verify automated review)
- Time estimates per step
- Three documented failure cases
- 24-persona review badge
Why This Matters
Single-modal LLMs have a fundamental problem: reviewing their own output means sharing the same mental model, same blind spots, same tendency to say "looks good."
Named-persona review compensates. By grounding each review in real, searchable, citable philosophy, you force genuine context switches. Ken Thompson actually said "do one thing well." That constraint produces different findings than "review as a senior architect."
The Meta-Lesson
A methodology that can review itself is more trustworthy than one that can't. Mine found 8 bugs in itself before I published it.
Try it: open-source-flywheel. Or just search your favorite engineer's philosophy and role-play. The method is free.
24 personas, 8 rounds, 5 steps. Thanks to de Bono (Six Thinking Hats), alirezarezvani (adversarial-reviewer format), and Feynman (don't fool yourself).