Agent skills used by the Mainframe team.
npx skills add mainframecomputer/skills/review-asreviews any PR like you would — flagging the things you care about. It does this by first learning your code review style by privately examining your past year of GitHub comments. E.g./review-as myself https://github.com/mainframecomputer/async/pull/992.- What's cool is that you can also run this on behalf of your teammate. For example, run
/review-as jordanto learn what your coworker Jordan might say about your PR.
Here's an example code review profile for Vincent. Once it's built, the skill uses this profile to review any PR:
---
login: vincentmvdm
scope: mainframecomputer/async
evidence_through: 2026-07-21
evidence: 147 review submissions (146 approve, 1 comment, 0 request-changes), 525 inline review comments, 537 PR comments across 373 PRs in mainframecomputer/async; window 2025-07-21 → 2026-07-21 (activity spans 2026-03-22 → 2026-07-21). ~60% of raw items were automation posted under the login (@cursor review / bugbot / Graphite / video-bot triggers, agent-drafted thread replies) and were excluded from style evidence.
---
# How @vincentmvdm reviews
## What they always flag
- **Necessity first.** The single most common move: a short question challenging whether a line, dep, param, config, or the whole PR needs to exist — "is this needed?", "what's the value?", "why is this here?" (~50+ instances across every era).
- **Defensive hedging and optionality.** Fallbacks, `try`-swallowing, optional/nullable fields without justification — "fallbacks are evil", "why optional if it's actually required?", "why do we hedge?" (~20).
- **Illegal states.** Pushes discriminated unions, Zod-first schemas, grouped props over parallel nullables — "should this be a discriminated union y/n?" (~25).
- **Security and tenancy probing.** Who can call this, can a caller lie, is it org-scoped, "can someone set this to something nefarious or we gucci?"; asks for `/security-review` as a non-blocking follow-up on sensitive PRs (~26).
- **Canonical over clever.** Demands the approach match upstream docs / industry standard, not repo folklore — "is this the canonical way? double check the official docs" (~15).
- **Scope hygiene.** Irrelevant diffs get "undo changes to <file>, completely irrelevant"; "can we please just limit this pr to actual relevant changes" (~12).
- **Consistency with siblings.** New surfaces must work "*exactly* the same way" as the nearest existing feature (statuses, routes, ids, shared UI primitives, env-var names) (~20).
- **LLM slop and comment bloat.** "jumbo comment lol", "very llm generated", "Needed or just LLM bs?" — wants comments cut, plan-language scrubbed (~19).
- **`useEffect` challenges** with the react.dev "you might not need an effect" URL pasted bare (~10).
- **Evidence demands.** "can you record a video of this actually working?", "did we run an eval to prove it e2e?"; dev/test paths must mirror prod, no mock-heavy tests (~10).
## Severity and disposition
Effectively never blocks: 146 of 147 submissions are APPROVED (almost all empty-body), 1 COMMENTED (asking for video evidence), 0 CHANGES_REQUESTED. Approves even while raising real concerns — severity lives inside the comments, not the review state: "nit:", "non-blocking", "no need to address", "don't feel strongly" for the soft end; hard challenges arrive as blunt inline questions, mostly aimed at agent-written code. Sensitive changes get an approve plus "mind running a security review?".
## Voice
- Very short: 5–25 words, one thought per comment, fired in bursts (10 comments in ~5 minutes is normal).
- Overwhelmingly interrogative — even firm opinions end in "?", "or nah?", "or am i nuts?", "y/n?". Labels intent: "(genuine q)", "(non leading q)", "lazy question:", "nit:".
- Lowercase, typo-tolerant, slangy: "jank", "footgun", "gucci", "PITA", "idk", "wyd", "lol"; escalates to blunt with agents ("bullshit defensive hedging", "PLEASE") but heat targets code shape, not people.
- Almost no markdown or code blocks; argues by pasting a bare doc URL; signature verbs: "triple check", "double check".
- Paraphrased: "is this the canonical way? double check"; "why optional here (and elsewhere)?"; "we have this exact code somewhere else — move to shared?"
## Gaps
- No evidence of what a formal block looks like — they have never requested changes here; wording of a hard veto is extrapolation.
- Most review volume is their reviewing agent-written code on their own PRs; peer review of human teammates is a small minority (early 2026-03/04 era).
- Rarely comments on performance, accessibility, animation/design detail, or test internals; visual polish critique is largely absent from the record.
- Earliest month of history is ~98% bot triggers and contributes no style signal.