القائمة
Implement a bounded code change from a request, spec, or implementation plan. Use when executing implementation work and choosing between quick, bounded, and tracked modes.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Manage tracked feature work with `.ai/current-work.md`, handoffs, and archive rules.
Turn a bounded functional spec into an implementation plan. Use when architecture, code structure, sequencing, migration, and verification must be decided before implementation.
Turn current context into a concise implementation spec. Use when a user wants to define what to build, capture requirements, or create a bounded technical/product spec before implementation.
Map a large or unclear initiative into facts, decisions, open frontier, and next work without writing a spec or implementation plan. Use when scope, architecture, dependencies, or sequencing are still genuinely uncertain before creating a spec.
Collapses deterministic multi-step repository research, aggregation, indexing, validation, and transformation into one small purpose-built local script and one Bash run. Use when a task would otherwise require 6+ related read/search/tool calls, iterates over many files or records, cross-references results, or requests batch analysis; do not use for small exploration or broad mutations.
Create a concise ADR for a durable, consequential architectural decision.
| name | implementation-workflow |
| description | Implement a bounded code change from a request, spec, or implementation plan. Use when executing implementation work and choosing between quick, bounded, and tracked modes. |
Use for code delivery. Optimize for evidence and signal density, not ceremony.
| Mode | Use when | Default shape |
|---|---|---|
| Quick Task | Small/local, obvious scope | Direct work; no artifacts/subagents |
| Bounded Delivery | Clear change across ~1-3 areas | Direct, or 0-1 scout -> 1 worker -> optional review |
| Tracked Project | Multi-session, expensive restart/handoff, or explicitly requested | Phase contracts + .ai/current-work.md pointer |
Default Quick/Bounded. Do not inherit a stale/completed/unrelated anchor.
Before work, collect only: objective, acceptance criteria, exact paths/symbols, constraints/non-goals, and test commands. Extract this once into a handoff packet. Do not make every child reread full plan/spec/current-work files.
Use a scout only to answer a real uncertainty. Good output: exact paths, symbols, line ranges, tests, and one start point. Bad output: file dumps or generic architecture prose.
Give one worker a coherent vertical slice with owned acceptance tests. A plan bullet, file, type error, or local repair is not automatically another delegation.
The worker should implement, run relevant evals, and fix in-scope implementation/type/test failures before returning. It returns: status, changed paths, eval evidence, one decision/blocker, and one next action.
Parallelize only read-only or isolated work. Never share a mutable checkout between concurrent workers unless the caller explicitly provides safe isolation.
For each phase state: objective, acceptance evidence, child shape, and escalation condition.
0-1 scout -> 1 worker -> optional 1 reviewer.current-work.md is a restart pointer, not a transcript. Keep it under 500 tokens when possible and point to detail.
Update it only at material phase boundary, blocker, decision, handoff, or closeout. Do not update/read it after routine child completions. Keep Todo Tracker at major-phase granularity; detailed work stays in the plan.
At a phase handoff, record only: completed phase and acceptance evidence, active artifacts, one decision/blocker, exact next phase/action, and budget used. On resume, use a fresh parent session when practical; read the anchor plus only the artifact section needed next, never prior tool history.
Report: mode, phase/status, changed paths, eval/test results, one decision/blocker, exact next action, current-work path if used, and review result/recommendation.