Self review

Self-Review

A rigorous, two-pass review of all changes on the current branch. Pass 1 catches mechanical errors via automated checks and mechanical checklists. Pass 2 applies the deep design judgment of a senior Staff Engineer to catch the subtler issues that separate correct code from good code.

This workflow is designed to be invoked standalone on any branch, or as an embedded step within a parent workflow (e.g., implement-spec).

---

Pass 1: Mechanical Verification

Purpose: eliminate the rote mistakes that waste human reviewer time. Every check here is binary — pass or fail, no judgment required.

Step 1.1 — Run automated verification

Run the repo's build/test/lint verification for the changeset. Resolve what to run in this order — use the first that applies:

1. Explicit command — the verify_cmd input or a $VERIFY_CMD environment variable.

2. Repo-provided verification entry point — a canonical script or target the repo already defines: e.g. ./scripts/verify.sh, ./bin/verify, make verify, make check, or the verification commands documented in the repo's AGENTS.md / CONTRIBUTING.md. Prefer a changed-files-aware runner if the repo has one.

3. The bundled inference helper — scripts/verify.sh (shipped with this skill) detects the repo's toolchain(s) and runs their standard build/test/lint commands:

   scripts/verify.sh            # exit 0 = pass, 1 = failures, 2 = could not infer
   

4. Manual derivation — if the helper exits 2 (nothing inferable, or a build system like Bazel that needs targeted invocation), derive the exact commands from the repo's docs and build files for the packages you changed, and run those. Never skip verification silently — if truly nothing can be run, state that explicitly in the final summary.

If any check fails, diagnose and fix the issue before proceeding. Re-run until all checks pass. Maximum 3 fix iterations — if still failing after 3 attempts, report the remaining failures and stop.

Step 1.2 — Apply mechanical checklists

Determine which files and languages are affected by the changes:

BASE="${base_branch:-$(git remote show origin | sed -n 's/.*HEAD branch: //p')}"
git diff --name-only "$(git merge-base "origin/${BASE}" HEAD)"...HEAD

Then apply, in order:

1. The general checklist — checklists/general.md (shipped with this skill): language-agnostic mechanical items that apply to any changeset.
2. Repo-specific checklists — from checklists_dir if provided, else auto-discover in the repo (in order): .agents/checklists/, .dev/checklists/, docs/checklists/. Apply every checklist whose language/stack matches the changed files.
3. Derived checklists — for each affected language with no repo checklist, derive a short mechanical checklist before reviewing: read the repo's linter/formatter/compiler configs and agent docs, and extract the rules that are (a) binary and (b) would fail the build or CI (fatal warnings, import rules, naming rules, generated-file policies). Apply that list.

Walk the checklists one category at a time, with that category as your sole focus for the sweep (all files for diff hygiene, then all files for safety, and so on) — controlled experiments show that an explicit "look for X" focus directive raises defect detection far more than checklist possession alone. Fix any violations found. If fixes were made, re-run Step 1.1 to confirm nothing broke.

---

Pass 2: Design Review

Purpose: catch the things that make code good versus merely correct. This is where the frontier model's intelligence earns its keep. No checklist can enumerate these concerns — they require taste, judgment, and deep experience with what makes software maintainable over years.

Independence

The agent that wrote the code is a systematically biased reviewer of it: LLM evaluators favor their own generations, and the effect is amplified when the reviewing context contains the memory of writing — you remember the intent and read the intent into the code. Therefore:

• Where the harness supports subagents, run Pass 2 in a fresh context given only: the base-branch diff, the changed files (to read in full), the standards below, and — if invoked by a parent workflow — the spec/ticket. Not the implementation history or conversation.
• Where it doesn't, apply the fallback discipline: every judgment must be argued from what is on disk, re-read in full — never from memory of writing it. If you catch yourself thinking "this is fine, I know why I did it," re-read the code as evidence instead.

The Reviewer Standards

Review against the standards below, directing your attention to each concern in turn. (The framing is a working stance, not magic: evidence shows role assignment by itself doesn't improve judgment — what does measurably help is explicitly focusing attention on each named concern, which is what the categories below are for.)

You are a Staff Engineer with 15+ years of experience building and maintaining production distributed systems. You have mass-reviewed thousands of PRs across your career. You have seen how innocent-looking code decisions compound into unmaintainable systems over months and years. You have also seen the opposite — code that was a joy to come back to because someone made the right structural choices up front.

You are not a pedant. You don't care about bikeshedding or stylistic trivia — the mechanical checklist in Pass 1 already handled that. You care about the things that determine whether this code will age well or poorly:

Abstraction Quality

• Is each function/class/module doing one thing well, with a clear contract?
• Are the boundaries between components clean and well-motivated?
• Could a competent engineer who has never seen this code understand the intent by reading the types, names, and structure — without needing inline comments as a crutch?
• Is the level of abstraction appropriate? Not so concrete that similar logic is duplicated, but not so abstract that you need a PhD to trace the control flow?

Naming as Design

• Do names reveal intent and domain meaning, not implementation details?
• Would someone reading a call site understand what's happening without jumping to the definition?
• Are boolean parameters and return values self-documenting? (e.g., forceRefresh = true vs a bare true)
• Do collection variable names indicate what they contain, not just that they're collections?

DRY Without Over-Abstraction

• Is there duplicated logic that should be a shared utility or method?
• But equally important: is anything abstracted prematurely? Is a "reusable" component actually used in only one place, adding indirection without benefit?
• Does factoring out shared code actually reduce total complexity, or does it just move it somewhere harder to find?

Error Handling as a Design Choice

• Are errors handled at the right level of the call stack — not too deep (swallowing context), not too shallow (leaking implementation details)?
• Is enough context preserved for debugging in production? If this fails at 3am, will the error message tell the on-call engineer what happened?
• Are failure modes explicit and visible, not hidden behind silent defaults, empty fallbacks, or swallowed exceptions?
• Is the error handling strategy consistent with adjacent code in the same module?

Extensibility and Change Resilience

• What is the next feature someone will likely want to add in this area? Does this design make that easy, or will it require reworking the current change?
• If the underlying data model changes (a new field, a new enum value, a new source type), how many files need to be touched? Is the blast radius proportional to the change?
• Are there implicit assumptions or magic constants that will silently break when requirements evolve?

Codebase Coherence

• Does this code look like it belongs next to the adjacent code?
• Does it follow the idioms and patterns established by the rest of the codebase — or does it introduce a new way of doing something that already has an established pattern?
• If it introduces a new pattern, is there a compelling reason? Or is it just the agent's default style leaking through?

Simplicity and Proportionality

• Is this the simplest solution that handles all the actual requirements?
• Could any part of the change be deleted without losing correctness or meaningful capability?
• Is the complexity of the implementation proportional to the complexity of the problem it solves? Over-engineering is a design defect, not a virtue.

Test Quality (where tests exist or should exist)

• Do tests verify behavior and contracts, or do they just exercise code paths for coverage?
• Are meaningful edge cases covered — empty inputs, error paths, boundary conditions?
• Would a test failure give you enough information to diagnose the bug without re-running with debug logging?
• Are test names descriptive enough to serve as documentation of expected behavior?

The Gestalt

• Step back and look at the full set of changes as a whole. Does this changeset tell a coherent story? Is the scope focused, or has it drifted into unrelated cleanup?
• If you were the reviewer, would you feel confident approving this after reading it once? Or would you need to ask clarifying questions?
• Is this the kind of code that builds trust with reviewers, or the kind that erodes it?

Procedure for Pass 2

1. Get the diff against the base branch:

   git diff "origin/${BASE}"...HEAD
   

2. Read every changed file in full — not just the diff. The diff shows what changed, but correctness and design quality depend on the surrounding code. A one-line filter change is only correct if downstream code handles the new possible values.

3. Read adjacent files to understand context, existing patterns, and how the changed code fits into the broader module. The goal is to see the change the way a reviewer who knows the codebase would see it.

4. For each file, evaluate against the design criteria above — one category at a time; a focused sweep per concern outperforms one diffuse read. Be honest and critical. The point is not to validate your own work — it's to find the things a rigorous human reviewer would find.

5. For genuine issues: fix them. Don't just note problems — resolve them. A self-review that produces a list of "consider doing X" is not a review, it's procrastination. Either it's worth fixing or it's not worth mentioning.

6. After all fixes, re-run Pass 1 (Step 1.1) to ensure nothing broke.

Calibration

• Only flag genuine issues. A Staff Engineer doesn't leave nitpick comments on code that's already style-consistent and functionally correct. If the code follows established patterns and handles its cases, let it stand.

• Every flag must be demonstrable. For each issue you raise, you must be able to state at least one of: the input/sequence that makes it fail, the contract or invariant it violates, or the concrete maintenance cost it incurs. LLM critics are known to hallucinate plausible-sounding bugs; if you cannot articulate the demonstration, the issue isn't real — drop it.

• Pragmatism over perfection. The goal is production-quality code, not platonic-ideal code. If a minor abstraction improvement would require touching 10 additional files for marginal benefit, that's not worth doing in this changeset.

• "Maybe consider..." is not an action. If you find yourself hedging, that's a signal it's not a real issue. Either commit to fixing it or move on.

• Respect existing patterns. If the rest of the codebase handles a concern in a particular way, follow that way — even if you'd prefer a different approach in a greenfield project. Consistency is more valuable than local optimality.

• Shared-abstraction extraction: factor repeated patterns into a shared utility/component only where it genuinely reduces complexity and the extraction would be used in 2+ places. Don't extract something used once — that's just indirection.

---

Completion

After both passes are done and all fixes have been verified:

• Report a brief summary: how many mechanical issues were found and fixed, how many design issues were found and fixed, and the final verification status (including anything that could not be verified and why).
• If invoked as part of a parent workflow, return control to it.
• If invoked standalone, optionally commit and push the fixes.