skills/kickoff/SKILL.md
Re-read the approved SPEC or PLAN, then build it while silently maintaining a running implementation-notes file that records every design decision, deviation, tradeoff, and open question where the build diverges from or interprets the spec. Use when the user says "start work", "begin implementation", "let's build this", or "kick it off" to start implementing a plan or spec they have just reviewed, even when they do not name the spec file. This is the handoff from plan review to coding.
npx skillsauth add shihyuho/skills kickoffInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
The user just reviewed a SPEC or PLAN and handed it to you to build. Build from the approved document — not a half-remembered one — and leave its author a record of where the build diverged from it.
1. Re-read the spec. Read the SPEC/PLAN in full from disk before touching code — the user likely edited it during review, so your context is stale. If you can't tell which document, or there are several, ask. If none exists, say so — kickoff assumes a reviewed plan.
2. Pick the notes format. Unless the user already picked one, ask: Markdown or HTML. Don't create the file yet — nothing to write until the build diverges.
3. Build, logging divergences as they happen. Implement with whatever workflow fits. The moment the build departs from or interprets the spec, log an entry — then and there, not batched for the end, when the alternatives you weighed are forgotten.
That first entry creates the file — implementation-notes.md (or .html) beside the spec, repo root if the spec lives there. Structure it however reads best; no fixed template — just keep the four kinds of note below distinct. If a notes file already exists for this work, append rather than overwrite. A build that never departs from the spec produces no file — don't create an empty one.
Four kinds of note, each answering a question the spec's author would ask:
The test: would the spec's author be surprised, or want a say? If yes, log it. If the spec determined the choice, or it's a routine detail, skip it — a notes file padded with the obvious goes unread.
The notes file is an extra record alongside your work, never a driver of it. Build exactly as you would if it did not exist.
Don't pause to walk the user through it, and don't announce entries — they open it when they choose. Don't let it suppress anything either: whatever you would normally stop and raise — a blocking question, a decision you would surface anyway — you still raise, log or no log. Logging an open question is no substitute for asking one that genuinely blocks you.
development
Write a short author's briefing to hand to a code reviewer whose agent already has its own review skill, so it supplies the context that skill can't see instead of repeating how to review. Right after you finish a piece of work, it mines this session (and any kickoff implementation-notes) for what the reviewer most needs flagged — the easy-to-miss changes, the parts you're least sure about, the looks-wrong-but-intentional bits, and the blast radius — plus the exact commit range to review. Use when you've just finished work and want to hand the review off to another agent, chat, or teammate, when you want a "heads-up for the reviewer", or when packaging a change for review elsewhere. It does not perform the review and does not re-specify severity tiers or output format — that's the reviewer's own skill's job.
testing
Use when creating, rewriting, pruning, or reviewing `AGENTS.md` or `CLAUDE.md`, especially to remove repo summaries, stale rules, and other low-signal global instructions. Trigger when deciding what belongs in always-on agent files versus a task-specific skill.
development
Drive a structured tutoring workflow that turns Claude into a learning onramp accelerator — consultative diagnosis → custom syllabus → unit-by-unit guided lessons with notes/whiteboard → dynamic adjustment from an accumulating learner profile. Use when the user states a learning goal ("I want to systematically learn X", "teach me Y", "help me prep for Z exam"), uploads study materials and asks for a course plan, or signals sustained guided study (mentions tutor, syllabus, course, lessons, study plan, curriculum, 家教, 學習路徑). Skip for one-shot factual Q&A or quick code-context explanations.
tools
Produce a TL;DR of a given target. Use when the user asks for a tldr, tl;dr, or quick summary of anything.