named-persona-adversarial-review – skill Lập trình cho agent AI

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](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](references/persona_principles.md).

Product (pick 1 per round — mandatory):

| Persona | Documented principle | Best for |

|---------|----------------------|----------|

| Steve Jobs | Start from the customer experience, work back to the tech | UX, onboarding |

| Marty Cagan | Fall in love with the problem, not the solution | PRDs, feature specs, scope creep |

| Des Traynor (Intercom) | The first 30 seconds decide adoption | Docs, READMEs, quick starts |

Engineers (pick 2 per round):

| Persona | Documented principle | Best for | Blind spot |

|---------|----------------------|----------|------------|

| Ken Thompson | Trust boundaries; do one thing well | Architecture, supply chain, API | UX, docs |

| Linus Torvalds | Eliminate the special case ("good taste"); never break userspace | Logic, data structures, compat | User empathy, DX |

| John Carmack | Measure before you optimize; performance as craft | Algorithms, hot paths | Minimalism |

| Kent Beck | Simple design; make it work → right → fast | Process, testability | Performance, security |

| Fred Brooks | Essential vs. accidental complexity | System design, estimation | Low-level perf |

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

| Level | Definition | Action |

|-------|-----------|--------|

| BLOCKER | 2+ personas concur on a CRITICAL, or security / data-loss risk | Fix before any further work |

| CRITICAL | Wrong result, data loss, security hole, or violated core invariant | Fix before merge |

| WARNING | Fragile, misleading, or likely to cause future bugs | Fix, or explain if deferred |

| NOTE | Improvement that doesn't affect correctness | Optional; record for follow-up |

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.

Step 1: Ground the principles first

For each persona, pull their documented principles from references/persona_principles.md (or search "[Name] engineering philosophy principles" and extract only sourced positions) before looking at the code, so you apply the principle rather than retrofitting one to an opinion you already formed.

Step 2: Review (3 independent — 2 engineers + 1 product)

Each persona gets: Mindset (one sentence from their principles), Priorities (3-5 criteria), Findings (each mapped to a documented principle + confidence level), or the zero-finding burden (3+ principles the code satisfies, with how).

Step 3: Synthesize & post

Merge duplicates; count concurrences; promote per the rule; flag single-lens findings (often the most interesting). Post the report as a PR comment (default) or save to .claude/review-[timestamp].md.

Integrity Check (Feynman)

> "The first principle is that you must not fool yourself — and you are the easiest person to fool." — Richard Feynman, *Cargo Cult Science* (Caltech commencement, 1974)

After each round, ask:

1. Would this person's *documented* philosophy actually direct attention here — or am I projecting?

2. Did I cite a real, sourced principle (confidence marked), or dress generic advice in a famous name?

3. Are my findings true on technical merit independent of the name attached?

4. All NOTE-level? Then I'm narrating one perspective in different voices. Switch ≥2 personas and re-review.

Exit Condition

• 1 round minimum for any PR.
• BLOCKER/CRITICAL found → fix, then 1 re-review round.
• CONCERNS (WARNING) → fix or accept risk, then 1 more round.
• CLEAN on 2 consecutive rounds → done.
• CLEAN on round 1 for a low-impact PR → done (1 round is enough).

When to Use

• You want deeper coverage than standard automated checks alone.
• A self-authored PR needs pre-submit hardening.
• adversarial-reviewer findings feel generic and you want sourced specificity.
• Reviewing methodologies or docs (product personas excel here).
• Auth, data, architecture, or public-API changes.

When NOT to Use

• Low-impact PR (cosmetic only, no logic change) → use adversarial-reviewer.
• No web access AND the persona isn't covered in references/persona_principles.md → you can't ground it; don't fabricate.
• Throwaway / prototype code.

Anti-Patterns

Inherits all from adversarial-reviewer. Plus:

| Anti-Pattern | Why wrong |

|-------------|----------|

| Inventing a verbatim quote to sound authoritative | Fabricated attribution to a real person. Cite the sourced principle + confidence, or drop it. |

| "As a senior engineer" without grounding | Not a named, sourced lens. Ground first. |

… (nội dung đầy đủ có trong gói tải về)