← ClaudeAtlas

diff-reviewlisted

Budget-capped adversarial review + simplification pass over a STAGED DIFF only — one agent, two lenses, findings-first. Use for commit-gate reviews instead of dispatching separate whole-repo reviewer and simplifier agents; the caller declares what already ran (tests, lint, gates) and the reviewer verifies claims against the diff's one-hop neighborhood, never the whole tree.
mickeyyaya/evolve-loop · ★ 4 · Code & Development · score 74
Install: claude install-skill mickeyyaya/evolve-loop
# Diff Review — precise, budget-capped, two lenses in one pass > Why this skill exists: two generic review agents (go-reviewer + code-simplifier) > spent 197k + 136k tokens and 97 + 58 tool calls reviewing ONE staged diff > (2026-07-27). Measured waste: re-deriving repo context the caller already had, > re-running test suites the caller had already run, unbounded whole-tree > consumer sweeps, and loading the same diff into two separate contexts. This > skill keeps the review QUALITY disciplines (evidence-first, adversarial, > file:line findings) and deletes the waste structurally. ## Scope contract (RIGID) The unit of review is the STAGED DIFF: `git diff --cached`. Nothing else is in scope unless a hunk drags it in. The caller's prompt MUST declare, and you MUST trust without re-running: - **Already verified**: which test suites/lints/gates ran and their results. Re-running any of them is budget theft — CI re-proves them anyway. - **Intent**: what the change is supposed to do, in 1-3 sentences. - **Known constraints**: pinned contracts, protected surfaces, prior review findings already applied. If the caller declared nothing, spend ONE tool call asking the diff itself (`git diff --cached --stat`) and proceed — never block on missing context. ## Budgets (RIGID) - **≤ 20 tool calls total.** Count them. At 20, stop and report with what you have. - **≤ 15 file reads**, and reads are TARGETED: `Read` with offset/limit around the hunk, never whole large files; the diff