Flow review
Run an independent multi-lens review of a change — correctness, integration and blast radius, production robustness, tests, security — and record verified, ranked findings in the task ledger. Use this skill when the user says "проревьюй", "посмотри что не так", "ищи баги", "review this", "check my changes before the PR", when an implementation has been tested and cleaned and is about to be committed or sent for review, or after fixing a previous round of findings. Prefer it over reading the diff yourself: you wrote the code, and the blind spots that produced the bug are the same ones that will hide it from you.From its SKILL.md
npx -y skills add maklybae/maklyskills --skill flow-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.8 KB, as published. Nobody here has run it
Flow review
The value of this stage comes from independence, not from thoroughness. You already know what the code was meant to do, so reading it again mostly confirms your intent. A subagent that sees only the diff, the task statement and one lens has no such loyalty — and five of them, unable to see each other's reasoning, will not converge on the same comfortable story.
1. Scope the diff
Take the base ref and the diff command from the profile's ## VCS section and
the ledger's base: line. Run it. Confirm it is non-empty and that its size is
what you expect before spending five subagents on it — a bad ref discovered
inside the fan-out wastes the whole round.
2. Build the context packet
Everything the reviewers share, assembled once:
- The diff command and the file list — not the diff pasted inline. They need to read the surrounding code anyway, so give them the means, not a snapshot.
- The task — the ledger's
## Task, so a finding can be judged against intent instead of guessed at. - The profile's stack, commands and constraints.
- Previously settled findings — every ledger finding with a verdict of
acceptedorrejected, verbatim with its reason. This is what makes the loop converge: without it a fresh reviewer re-derives the same suggestion every round, forever. - Out of scope, stated plainly: comments, formatting and naming style
(
flow-cleanupowns them), and conformance to the written spec unless the user asked for that lens.
3. Dispatch the lenses in parallel
Pick the lens set from references/lenses.md — the five
defaults unless the change or the user calls for more. Fill
references/reviewer-prompt.md once per lens and
send all of the Agent calls in a single message so they run concurrently.
Dispatch each lens as a flow-reviewer agent — the bundle ships it, and its
tool set carries no edit tools, so read-only is a property rather than a
promise. Where the bundle's agents are not installed, fall back to
general-purpose; either way reviewers read and run commands, they never edit.
The prompt template carries the parts that decide review quality: findings must be proven before they are reported, findings are capped and ranked, and finding nothing is an acceptable answer. Do not paraphrase it — pressure to produce findings is what manufactures false positives, and a reviewer that has to fill a quota will fill it with noise.
4. Merge
- Deduplicate. The same defect will arrive from two lenses in two vocabularies. Keep the version with the concrete failure scenario.
- Drop what is already settled. If a finding restates one the ledger marks
acceptedorrejected, drop it — unless it brings genuinely new information, in which case say what changed. - Rank by severity: blocker, serious, minor. The ladder is defined in the prompt template; use it unchanged so severities mean the same thing across rounds.
- Sanity-check the blockers yourself. A claimed blocker justifies thirty seconds of your own reading. If the trigger path does not hold up, demote it and say why — passing an unverified blocker to the user costs their trust in the whole stage.
5. Record and report
Append each surviving finding to the ledger's ## Findings with a fresh id —
the ledger is at
~/.claude/projects/<project-root-as-dashes>/flow/<task-slug>/ledger.md, schema
in ../flow/references/ledger.md relative to this skill's directory. Ids are never reused, so that "R7 is back" stays a meaningful
sentence.
Report to the user compactly — id, severity, file:line, and the one-line
failure scenario. No praise section, no summary of what the code does: they
wrote it. If a lens found nothing, say so in one line; that is information too.
Then hand over to flow-resolve, which gives every finding a verdict.
What this stage does not do
It does not fix anything. Mixing review and repair loses the record of what was
found, and a reviewer that starts editing stops reviewing. It also does not
re-run the tests — flow-test did that, and its results are in the ledger.