用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/marcellocurto/skills --skill implementation-planner命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | implementation-planner |
| description | Turn a defined feature or refactor into a bounded, implementation-ready plan without changing code. |
Plan only. Do not edit files, apply patches, commit, or implement unless the user explicitly changes the request.
Produce one repository-grounded plan that another engineer can implement without reconstructing the conversation or rediscovering the relevant code path.
Identify:
Treat examples, suggestions, rejected approaches, and future ideas as context unless the source explicitly adopts them. Do not weaken requirements to make the implementation easier.
Start from the request, issue or specification, repository instructions, glossary, ADRs, and prior decisions. Search by domain concept, then inspect only the code needed to understand the current path, ownership, stable seams, consumers, relevant tests, and repository validation commands.
Trace at least one representative path from its caller or entry point to the behavior, state change, or output being modified. Distinguish confirmed repository facts from inferences and unknowns. Stop inspecting once the plan's change boundary and validation path are supported by evidence.
Choose one primary implementation path. Prefer existing modules, interfaces, helpers, types, and data shapes when they can satisfy the requirement cleanly.
Every proposed file, abstraction, dependency, schema change, state mechanism, configuration option, or public interface must have a direct requirement or repository-based reason to exist. Preserve essential complexity that represents real domain rules, compatibility, durability, recovery, security, or operational behavior.
Name the expected files and symbols to change and why. Treat that list as the evidence-backed expected surface, not certainty that later implementation may never challenge. Identify protected contracts or consumers when the change could affect them.
Define validation through observable behavior or a durable module seam. Each proposed test must name the realistic regression it would catch. Include exact existing commands when they can be established from the repository; do not invent commands or require tests that merely mirror implementation details.
Taste-check the draft and revise it in place:
Keep intentional complexity only when the requirement or repository evidence justifies it. Preserve useful problem-specific reasoning, but omit the draft and refinement history from the final answer unless a discarded approach explains an important decision.
Call the plan ready only when its requirements, current-code evidence, proposed behavior, expected change surface, ordered steps, and validation are concrete enough to implement.
A local, reversible assumption may remain when repository evidence supports it; record it. If a missing decision could materially change user-visible behavior, public contracts, data semantics, security, identity, routing, scope, acceptance criteria, or external authority, do not choose silently. Mark the plan not ready and state the exact decision or evidence needed.
Keep the plan concise and use only headings that add decision value. A ready plan normally includes:
Stop when the refined plan is implementation-ready or when the remaining material blocker is stated precisely. Do not continue reading merely to make the plan look exhaustive.