Build the active spec to completion, one reviewable pass at a time. The spec is\nthe implementation plan; user constraints and accepted experimental evidence\nremain binding when the architecture changes.
\nA pass (usually one slice) is a commit checkpoint, not a stopping point. The\njob is the whole spec — every slice, every global TODO — not the first green\ncommit. Finishing a pass means starting the next one, not handing back to the\nuser. Only stop when the spec is fully implemented (or a genuine blocker needs a\ndecision only the user can make).
\nWork in parallel wherever the graph allows. Do not walk the ladder one slice\nat a time when slices are independent. Read the spec's dependency graph as a\nwavefront and delegate independent passes to subagents that run concurrently\n(see Rules) — you orchestrate and integrate; only serialize what genuinely\ndepends on prior work.
\nRead the repo README, the spec README, and the next slice before editing.\nLoad any skills named by the spec. Identify the current pickup point, global\nTODOs, required gates, and what must stay green. If a multi-slice spec lacks\na live handoff prompt, add one before the first pass ends.\nFor a validated spike, apply Preserve the winner below before editing.
\nReconcile the plan with the current code. If the slice would preserve a\ndevelopment-only shim, duplicated type, weak wrapper, or obsolete path,\nreplace it with the simpler architecture and update the spec handoff.
\nImplement one coherent pass: usually one slice, one vertical checkpoint, or\none architecture correction. Keep the review surface small enough to audit.
\nVerify the actual contract. For behavior changes, run the focused unit tests\nfirst; for browser-visible work, use the real browser/harness and inspect\nscreenshots so the subject is framed and readable, not merely nonblank.\nA visual CHANGE carries two extra proofs before you claim it: a\nbyte/pixel diff against the pre-change baseline on the production route\n(static checks and re-anchored assertions all pass on a no-op — a\npalette pass once shipped \"verified\" while the production frame was\nbyte-identical), and an unprimed\nscreenshot-critique — at the\nuser's reported framing when the pass answers their visual bug report —\nbefore declaring it fixed; the implementer's eyes are primed by the fix\nand repeatedly pass what fresh eyes catch.\nNever weaken an existing default gate or repin a failing contract without\nproving the old contract is wrong.
\nReview the change list and clean up after every pass, before committing.\nRead git status/git diff --stat line by line and account for every path:\none-off probes, shot scripts, scratch files, nohup.out, ad-hoc screenshot\ndirs, and disposable SPIKE/debug notes never enter a commit — scratch stays\nout of the tree. Preserve frozen references, fixtures, and research evidence\nin the project's reference or spec assets area. A file without a durable\npurpose does not ship.\nDelegated agents leak these; the integrating reviewer re-checks the merged\ntree with the same eye.
Sizing the pass — measuring what it cost and reworking it when the line count\nis out of proportion to the behaviour it delivers — belongs to the shape pass\nin step 6, under refactor-clean.
\nRun review at the end of every pass, before committing.\nIt sequences the three closeout lenses — refactor-clean on the shape,\ncode-review on the settled diff, write-docs on what the change touched — and\nyou apply the fixes from all three. Long specs are where sediment compounds,\nso hold the shape pass to this pass's own output: dev-only shims, duplicated\nconcepts, parallel abstractions, and compatibility wrappers it introduced\ncollapse into the clean contract with one owner, so the code reads as designed\ntoday, not tacked on. Last, run\naudit-choices on the cleaned pass — a pure\naudit that appends every decision made where the spec was silent (your own,\nand each delegated subagent's when integrating) to the spec's choices\nledger (specs/<feature>/choices.md) with verdicts, changing no code\nitself. You act on its findings: redo unsound choices from their corrected\ndecisions, adopt the recorded provisional call on any user-only entry — the\naudit never blocks the run. Rerun the affected checks, then commit only the\nfocused changes from this pass.
Update the spec README's \"Next Agent Prompt\": status, completed work, next\npickup point, blockers, changed gates, and any architecture decision that\nchanged the plan.
\nRun a maintenance checkpoint as part of the loop, not as endgame cleanup.\nTrigger it after a red pass, after every two or three slice commits, after a\nrebase/resume/compaction, before changing feature areas, when evidence\ninvalidates the plan, when the handoff contradicts the TODO/graph, when the\nchoices ledger's entries cluster around one slice, or when the\nactive prompt grows hard to scan. Long specs bloat repeatedly; cleanup is a\nnormal pass, not a cosmetic chore.
\nA checkpoint cleans both plan and code before more feature work:
\nCommit the checkpoint as its own focused pass when cleanup changes the spec,\ncode shape, or handoff enough that future agents would otherwise inherit stale\ncontext. It is done only when a fresh agent can read the README handoff, TODO,\nand slice graph and choose the same next action without conversation history.
\nContinue. If any slice or global TODO is still open, go straight back to\nstep 1 for the next one — same session, no pause for acknowledgement. Keep\nlooping until every TODO is closed.
\nRun review once more over the whole spec. The\nper-pass reviews each judged one slice against the code as it stood then; this\none judges the finished feature. Scope it to the spec's full diff, not the last\npass — that is the only scope where duplication spread across slices, a shape\nthat only reads wrong once every slice has landed, and docs that describe\nincrements instead of the feature are visible at all. Apply the fixes, rerun\nthe gates, commit.
\nConsolidate the choices ledger, then close. When the last slice lands, the\nchoices.md you've been appending to per pass is build-order sediment: entries\nbanked early carry \"provisional — revisit in slice N\" verdicts that a later pass\nsilently resolved, entries a later pass reverted still sit there, and the same\nchoice may appear twice. The per-pass rule \"banked is settled, never re-listed\"\nis what let that drift accumulate — so the final consolidation is its deliberate\nexception. Before archiving, rewrite choices.md from scratch as the final\nledger: re-audit every banked choice against the final shipped code (not the\npass it landed in), collapse each provisional/needs-later entry to its actual\nend state, drop anything a later pass superseded or reverted, and merge\nduplicates. Keep it choices only — no gate results, e2e evidence, or\nreview-finding narration; those are reported elsewhere and are not decisions the\nuser now owns. Present it per audit-choices: grouped\nby verdict, ranked least-confident-first, every entry ELI5 and standalone.\nThen close the spec with close-spec.
A pass is done when code, spec handoff, verification evidence, review\ncleanup, and a focused commit all agree on the same current truth — then you\nstart the next pass.
\nThe spec is done — and only then is this skill done — when every slice and\nglobal TODO is closed, all gates are green, the whole-spec review has run and its\nfixes have landed, the handoff shows nothing left to pick up, and the spec has\nbeen archived with close-spec.\nAnything short of that is mid-implementation: keep going.
\nThe final handback presents the choices ledger, not the diff, per\naudit-choices — a days-long unsupervised run\nearns its merge through this ledger; it is the user's review surface for\neverything decided without them. Hand over the consolidated ledger from\nstep 11 (final state, verified against shipped code, choices only), never the\nraw per-pass append — a ledger still carrying \"will be done in a later slice\"\nverdicts tells the user you never went back to confirm it was.
\nClose the handback with the size of what you added — always last, after the\nledger. The ledger says what was decided; this says what it cost. A short\ntable over the whole run:
\n| \n | added | \ndeleted | \nnet | \n
|---|---|---|---|
| Production code (excl. comments) | \n\n | \n | \n |
| Comments | \n\n | \n | \n |
| Tests / harness | \n\n | \n | \n |
| Specs & docs | \n\n | \n | \n |
Then one paragraph naming the structural surfaces the run added — a new\ncron, table column, index, endpoint, config flag, dependency — because those\nare what the user now owns and maintains, and a line count alone hides them.\nExclude formatter churn from files the run did not otherwise touch, and say so\nif you excluded any.
\nState the count plainly whatever it is. A large net addition for a small\nbehavioural change is a finding to report, not a number to bury — and if you\nnotice it here rather than at the pass that caused it, say which pass it was.
\nSlices are not the only unit of scope. A spec also records decisions —\nledger rows, decision-table entries, invariants — and a decision can be agreed\nin planning but never turned into a slice, especially when it's orthogonal to\nthe slices' theme. \"All slices closed\" then reads as done while that decision is\nsilently unbuilt. Before declaring the spec done, reconcile every recorded\ndecision against the shipped code, not just the slice list: each one is either\nimplemented, or explicitly marked \"no code needed.\" A decision with no owning\nslice is the classic silent miss — the close-spec audit is the backstop for it,\nnot the first line of defense.
\n","files":[{"path":"SKILL.md","sizeBytes":15845,"isText":true}],"reviewScore":null,"reviewSummary":null,"trust":{"provenance":"trusted-source-unreviewed","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow.","bodySource":null},"bodyLocked":false,"purchaseUrl":null,"sourceUrl":null,"report":{"provenance":"trusted-source-unreviewed","screen":{"ran":true,"outcome":"clean","suspicious":0,"notes":0,"hiddenCharacters":false},"virusScan":{"engine":"clamav","status":"clean","scannedAt":"2026-10-14T21:49:48.271327Z","sha256":"135AA9FFEE8998E637E3D2C61F9E3E368458BED55CF5D21F16E99CD8B034A5B3","sizeBytes":6753},"review":null,"source":{"repositoryUrl":"https://github.com/dzhng/skills","path":"skills/engineering/implement-spec","license":"MIT","commit":"f2fbf92f131fb08bbf3e400a1ef99b7be85bb2bc","subtreeSha":"A08BE132E24619D7DEE8A1C15EE86068FEE6A958D1EA3FE0A18946647FBE6BA4","lastSyncedAt":"2026-10-14T21:49:23.275319Z"},"reviewedAt":"2026-10-14T21:50:21.236994Z","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow."},"install":[{"target":"skills-cli","command":"npx skills add https://github.com/dzhng/skills/tree/main/skills/engineering/implement-spec"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install dzhng-skills@llmmart"},{"target":"git","command":"git clone https://github.com/dzhng/skills.git"}]}