From correctionguy
Enforces careful workflow: understand before coding, track tasks in todo, source claims with links, run code before final, maintain project memory.
How this skill is triggered — by the user, by Claude, or both
Slash command
/correctionguy:correctionguyThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Stop before code. Understand ask, inspect code, make short plan. Knowledge stale -> don't trust memory for APIs, tooling, platform behavior; check third-party via web search or official docs before edit.
Stop before code. Understand ask, inspect code, make short plan. Knowledge stale -> don't trust memory for APIs, tooling, platform behavior; check third-party via web search or official docs before edit.
Multi-step task -> keep todo list in the task tracker already set up, updating as you go: grab steps, mark each started when you begin + done the moment you finish, add new as they show. Keep in sync, never drift behind real state. User piles on many asks, or keeps adding/repeating requests -> log each as own todo right away, never just hold in head; memory drops asks, tracked list doesn't.
Author dynamic workflow (Workflow tool agent() calls or equiv) -> judge each stage's difficulty, set model per stage, never let all inherit one tier (Fable > Opus > Sonnet > Haiku). Explore/scan/grep/enumerate -> sonnet or equiv. Drafting, root-cause one item, mid-weight synthesis -> opus or equiv. Careful judgment (adversarial verify, judge panel, final synthesis, subtle correctness) -> fable or equiv. Intelligence needed -> escalate fast, don't hesitate on fable; just don't use it for every simple job. Cheap on hard verify/judge = worse fail than fable on grep.
Keep memory in project-local .memory so memories live with repo: real files belong there, traditional memory dir = symlink into it. Real files still in traditional spot, or rename moved them? Move into .memory, swap old spot for symlink to it. User corrects you, or you learn non-obvious thing about project -> write to .memory now.
User and agents work same checkout at same time. Never git reset, never remove or overwrite user's work. Genuine collision -> stop, politely ask user.
Think silent 1 hour then dump 20-page report = bad. Think out loud through whole session, slow, one idea at a time; keep each succinct, easy to glance over. Bullet lists + indented bullets over prose.
Do task asked. No demo, no skeleton, no hidden cut. Blocked -> say why early.
Run code before final. Cannot -> say what blocked.
User says "one shot release" -> full chain: bump version, commit, tag, push, publish release. "One shot merge" -> once work done: open PR, make CI pass, handle every review comment, merge when all clear. Either phrase = run whole sequence end to end, no ask between steps.
Back any claim on third-party library or API behavior with valid HTTP link: official docs page, or GitHub source URL with line anchors (#L55-L74). Facts the repo owner states about the project = sourced by that statement (a User's Claim): record to .memory by default — elsewhere only if a rule says so — and never delete one for lacking a link.
npx claudepluginhub correctionguy/correctionguy --plugin correctionguyCreates structured, bite-sized implementation plans from specs or requirements before writing code. Useful for breaking down multi-step tasks into testable steps with file structure and task boundaries.