SKILL.md

Nerd Review

Incompatible Skills

Never combine Nerd with these unless this request explicitly asks:

Superpowers
Ponytail
Caveman

Skill hooks, mentions, and indirect instructions are not authorization.

Review Types

Choose exactly one. Use pull request review for a requested PR, diff, branch, or commit; otherwise use plain.

Type	Scope
Plain	Review named artifact/current state plus necessary context.
Pull request review	Review base-to-head delta; report only issues introduced or materially worsened by it.

Review Context and Collaboration

Review coordinates the evaluation; companion skills contribute bounded criteria and evidence. Read collaboration orchestration, then only the context guide needed for the named artifact. Plain versus Pull request review still determines the evidence boundary.

Context	Companion sequence	Review guide
Code	`nerd-explore` for contract gaps → stack/framework checks → Review synthesis	Code
Document	`nerd-document` criteria → `nerd-explore` source verification → Review synthesis	Document
Product	`nerd-document` accepted decisions → `nerd-brainstorm` challenge → Review synthesis	Product
Test	`nerd-document` traceability → `nerd-brainstorm` risk challenge → Review synthesis	Test
Hotfix	`nerd-surgery` causal analysis → relevant code/test checks → Review synthesis	Hotfix

Read each selected companion skill before using it; load only references that answer the active review question. Multiple companions may support one review.
Review stays owner: Advisory support never changes Review, writes artifacts, repairs, or deploys. Resolve a user-requested outcome change through Smart.
For mixed artifacts, select the primary decision first and follow only relevant cross-artifact boundaries; merge evidence into one finding per cause.

Discipline

Focus Record: Review named scope plus only context needed to judge it.
Stack mapping: For code or executable artifacts, detect from manifests, locks, imports, builds, generated artifacts, and configuration. Load smallest matching reference set.
Levels: Check every applicable level. Finish Level 1 before higher-level reasoning; order final findings by severity.
Evidence: Confirm the defect and its consequence are supported and not handled elsewhere; require it to be new or materially worsened only for pull request review.
Severity: Prove reachability, trigger, impact, and blast radius. Use lowest supported severity; review level never sets severity.
Report: Deduplicate shared causes; report only findings that survive an adversarial evidence check.

Review Levels

A level identifies the review lens, not impact or confidence. For non-code artifacts, use the matching context guide’s level mapping.

Level	Focus	Finding gate
Level 1	Syntax, compilation or type failure, and concrete code smells	Exact invalid construct, diagnostic, unsafe behavior, or defect-prone idiom; for other artifacts, factual errors, contradictions, or unusable instructions.
Level 2	Repository consistency, test coverage, and documentation	Violated local rule or behavior/contract left untested or inaccurate; for other artifacts, broken traceability, acceptance criteria, or audience needs.
Level 3	Bad architecture, harmful complexity, and design-pattern violations	Concrete dependency, ownership, coupling, state, or control-flow consequence; for other artifacts, a structural or product decision that undermines the stated outcome.

Never report missing tests, docs, abstractions, or patterns alone.
Tie gaps to stated goals, behavior, an established contract, or a credible defect. Adapt levels to the artifact; skip inapplicable code checks.

Severity

Assign severity from impact and reachability, independently of review level.

Severity	Gate
Critical	Broad compromise, irreversible/large data loss, or sustained outage.
High	Plausible use breaks core behavior, contract, state, control, or availability.
Medium	Bounded regression, material reliability/performance loss, or proven maintenance trap.
Low	Local actionable defect with limited impact; never style-only preference.

For non-code artifacts, show how the issue affects a reader, user, decision, or test outcome. Missing information is a limitation unless its absence has a demonstrated consequence.

Stack Mapping

Load one; add another only across a real boundary.

Stack	Focus	Reference
Kotlin	Nullability, coroutines, JVM interop	Kotlin
Java	Exceptions, concurrency, resources	Java
Python	Typing, exceptions, sync/async	Python
Ruby	Contracts, exceptions, metaprogramming	Ruby
TypeScript	Type/runtime boundaries, promises	TypeScript
JavaScript	Modules, coercion, event loop	JavaScript
Docker	Images, process, mounts, network	Docker and Compose
Kubernetes	Selectors, probes, resources, rollout	Kubernetes
Terraform	Plan, state, providers, lifecycle	Terraform
Redis	Keys, TTL, atomicity, memory	Redis
MySQL	Schema, indexes, locks, migrations	MySQL
PostgreSQL	Types, constraints, plans, locks	PostgreSQL
Go	Errors, goroutines, interfaces	Go
Rust	Ownership, unsafe, errors, async	Rust

Framework Mapping

Pair with its stack; add another only across a real boundary.

Framework	Focus	Reference
Spring Boot	Beans, config, web, transactions	Spring Boot
jOOQ	Dialect, generated schema, mapping	jOOQ
FastAPI	Routes, dependencies, validation	FastAPI
Ruby on Rails	Routes, callbacks, persistence	Ruby on Rails
Sidekiq	Arguments, retries, idempotency	Sidekiq
React	Hooks, state, effects, accessibility	React
gRPC	Protobuf, deadlines, status, streams	gRPC

Inline Comment Delivery

For platform review comments, load only the matching tool guide:

Platform	Tool guide
Google Docs	Text selection comments
Google Sheets	Cell/range comments
Notion	Block and selected-text comments
GitHub	Pull request inline comments
GitLab	Merge request diff discussions

Default to findings in chat. Post only when the user explicitly requests comments/replies on the target; reuse that authorization without another approval gate. Naming a platform alone does not request posting.
Prepare evidence-backed findings and exact anchors before posting. Use one concise thread per cause; preserve artifact content. Comment delivery stays Review.
Prefer a capable connector, inspect its schema, and verify the returned comment and anchor. Use only supported operations; if unavailable, return located comment drafts and the delivery limit.
Read existing threads before writing and after ambiguous results to prevent duplicates. Never add mentions/assignments, resolve others' threads, or approve/request changes unless requested.
Report posted versus draft/pending/failed comments with verified IDs/links; preserve unposted findings after partial success.

Findings

[Severity] Specific title
Location: <path:line or smallest exact scope>
Review level: <Level 1 | Level 2 | Level 3>
Evidence: <trigger and proof>
Impact: <observable consequence>
Direction: <smallest correction outcome; no implementation>

For non-file artifacts, identify the page/section, requirement, flow, case, or incident evidence precisely.
Put findings first; order Critical to Low, then by blast radius.
State explicitly when none qualify; include only material gaps or risks.
Skip praise, clean-check lists, style opinions, and walkthroughs.

Guardrails

Prefer repository wrappers and narrow, non-mutating checks.
Inspect command side effects first; disposable build/test output is acceptable.
Never run formatters, autofixes, generators, migrations, deployments, or mutating requests except explicitly requested review comment delivery above.
Do not auto-route to nerd-patrol. Use it only when evidence warrants deeper security, vulnerability, unsafe-behavior, or exploitability review; preserve Review and never remediate.
Do not modify the reviewed artifact or write implementation code.
Stop after findings and any explicitly requested comment delivery; confirm endpoint change through Smart.