菜单
Gives a fresh agent a plan, audits it for comprehension gaps, and verifies technical assumptions against the codebase before implementation begins. Spawns a single sub-agent to perform both checks and presents the findings.
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。
基于 SOC 职业分类
Use when the user requests to update, upgrade, or install the latest release of Neovim.
Write a comprehensive, no-placeholder implementation plan as a series of bite-sized tasks, each with exact file paths, real code, TDD steps, and commit instructions. Saves as a GitHub Issue with a [PLAN] prefix. Use after preflight has established what to build.
Present a plan to the user for inline annotation via the ai-review UI. Use when you have a plan ready for human review before executing it. The user will annotate it in the browser; you then revise based on their comments.
Explain the core concepts, library design patterns, system architecture, or host environment internals involved in a plan or git diff. Spawns an isolated subagent to perform the education.
Capture an Architecture Decision Record (ADR) for a significant decision made in the current project. Store in the project's adrs/ folder. Use when the user asks to record a decision, log an ADR, or capture why something was built a certain way. PROACTIVELY suggest capturing an ADR whenever a non-obvious architectural choice is made — technology selection, structural trade-offs, rejected alternatives — especially after a discussion where context and reasoning were developed. A good prompt — "Want me to capture this as an ADR?"
Manage Jack's personal technical reference library at ~/git/references. Use for searching/reading existing references, or adding new ones. Invoke with /jack-references and optional keywords to search, or when the user wants to save a new reference guide. PROACTIVELY suggest saving a new reference whenever the conversation involves figuring out how a library, API, or tool works — especially after debugging, reading docs, or solving a non-obvious integration problem. A good prompt — "Want me to save this as a reference for next time?"
| disable-model-invocation | true |
| name | onboard-plan |
| description | Gives a fresh agent a plan, audits it for comprehension gaps, and verifies technical assumptions against the codebase before implementation begins. Spawns a single sub-agent to perform both checks and presents the findings. |
Use this skill when you have a plan saved as a GitHub Issue and want to begin a fresh implementation session with zero ambiguity and validated technical assumptions.
The skill argument may be:
gh issue view <number> --json title,body.gh issue list --search "[PLAN] <description>" --state open --json number,title
and pick the best match by title similarity. If more than one is plausible,
ask the user to choose before proceeding.Read the full issue body. Do not summarise or paraphrase it at this stage.
Spawn a single research or self subagent to audit the plan. It receives the raw plan text and this prompt:
You are onboarding onto a new implementation plan. Your goal is to identify comprehension gaps in the plan text and verify its technical assumptions against the live codebase.
PLAN:
{PLAN_CONTENT}
Perform the following tasks:
1. COMPREHENSION AUDIT:
Identify any places where an engineer would not know exactly what to do or what something means (e.g. undefined terms, ambiguous steps, sequencing issues).
Formulate a list of questions for the user in the format:
Q: [question]
Where: [exact quote/step]
2. CODEBASE VERIFICATION:
Cross-reference the files, APIs, databases, and configuration keys mentioned in the plan with the live codebase.
Identify any:
- Stale/incorrect assumptions about existing APIs, exports, or data schemas.
- Mismatches in configuration or file layouts.
- Hidden dependencies/prerequisites omitted from the plan.
Constraints:
- Do not suggest or make changes to the codebase.
- Rely only on live codebase verification (do not guess).
- Focus only on technical blocking issues.
Return the results in two clean sections: "Comprehension Gaps" and "Codebase Verification". No preamble, no summary.
Display the results returned by the subagent. Invite the user to answer the questions and resolve the codebase findings before beginning implementation.