Named Persona Adversarial Review

Named-Persona Adversarial Review

TL;DR: Abstract roles find abstract problems. Named engineers with documented, sourced philosophies find problems you would actually fix — as long as you cite the real principle and never invent the quote.

Triggers: "review this PR with real engineers" | "named persona review" | "philosophy-grounded review"

Example Output

CRITICAL [Torvalds]: Special-case error handling at auth.ts:47 duplicates the
  happy path. Torvalds' documented "good taste" principle: restructure so the
  special case disappears rather than adding a branch. (confidence: high — TED 2016)
WARNING  [Thompson]: parseConfig() does three unrelated things; the Unix
  "do one thing well" principle argues to split it. (confidence: high)
NOTE     [Jobs]: Error "EACCES:13" leaks an errno at the user surface; "start
  from the customer experience" argues for a human message. (confidence: high — WWDC 1997)
Verdict: CONCERNS — fix CRITICAL before merge.

Problem

Abstract adversarial review ("act as a saboteur") produces generic findings — the model imagines what a reviewer might say. This skill grounds each lens in a real, sourced engineering philosophy documented in references/persona_principles.md: what Ken Thompson actually argued about trust, what Linus actually demonstrated about good taste — not what an AI imagines.

How it differs from adversarial-reviewer: abstract roles → surface-level findings; named, sourced personas → findings anchored to a documented principle you can cite and defend.

Cost: 1 round ≈ 8-12 min. Comparable to waiting for CI.

Attribution discipline (read this first — it is the load-bearing rule)

This skill puts named, real people's principles to work. That power is also its failure mode: language models hallucinate quotes. To stay honest:

1. Cite the principle, not a fabricated verbatim quote. Prefer paraphrasing a documented position ("Thompson's Reflections on Trusting Trust argues you can't trust code you didn't fully create") over inventing quotation marks around words the person may never have said.
2. Attach a confidence level to every attribution — high (documented, in references/persona_principles.md with a source), moderate (widely attributed, source not pinned), low/unknown (you're inferring). Mirrors productivity/andreessen's citation discipline.
3. If you cannot ground a persona's lens in a real source, drop that persona. A confidently-wrong quote attributed to a living engineer is worse than one fewer reviewer. Never fabricate a citation to hit the "≥1 finding" bar.
4. The finding must stand on its own technical merit. The persona is a lens that directs attention, not the authority that makes the finding true. A real bug found "through Carmack's lens" is real because it's a bug, not because Carmack said so.

Rules

• Ground before role-play. Anchor each persona in references/persona_principles.md (or a verifiable search) first. Ungrounded = invalid.
• Findings stand on technical merit, with the persona's principle as the lens — see the discipline above.
• Product persona mandatory every round. Engineers miss UX. Always include one.
• Honesty over quantity. Don't fabricate findings or citations. Clean dimensions get reported clean (with the zero-finding burden below).
• Zero-finding burden. "Looks fine" is only valid if you name 3+ principles the code demonstrably satisfies, and how. Non-findings are as expensive as findings.

Persona Pools

Each persona's documented principles + sources + confidence live in references/persona_principles.md.

Product (pick 1 per round — mandatory):

Engineers (pick 2 per round):

Routing (which personas when):

• Code correctness → Torvalds + Carmack + Jobs
• Architecture / design → Thompson + Brooks + Cagan
• Documentation / API → Thompson + Beck + Traynor
• Performance → Carmack + Torvalds + Jobs
• Security / supply chain → Thompson + Torvalds + Cagan
• 1st round on any PR → Torvalds + Thompson + Jobs (broadest coverage)

Severity Levels

Promotion: NOTE → WARNING → CRITICAL → BLOCKER. Two personas independently finding the same issue promotes it one level (concurrence is signal). BLOCKER is the ceiling.

The Process

Step 0: Read twice

1. Top-down (comprehension): what changed, and why.
2. Bottom-up (adversarial): read function by function, last to first. Ask what each function actually guarantees vs. what its name implies, where it can fail, and what it assumes about callers. Reading bottom-up breaks the author's mental model. Multi-file → trace one end-to-end path.