correct-course — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited correct-course (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Run this skill when something has changed — new information, a failed assumption, a reshaped scope — and the current pipeline artifact can no longer be trusted. /correct-course diagnoses the blast radius, names the earliest skill that needs to re-run, and walks you through cleaning up stale artifacts before forward motion resumes.
This is the named, invocable form of the Pipeline Recovery rules in references/SYSTEM-OVERVIEW.md (bundled alongside this skill). Those rules still apply — this skill just gives them a front door.
One question per turn. Throughout this skill, ask one question at a time and wait for the user's answer before asking the next. Backtracking is a conversation, not a form.
>
Prefer single-select. Use single-select multiple choice when the user is choosing one direction, one priority, or one next step.
>
Use multi-select rarely. Reserve it for compatible sets — goals, constraints, non-goals, success criteria — that can all coexist. If prioritization matters, follow up asking which selected item is primary.
>
Use the platform's question tool when available. In Claude Code, useAskUserQuestion; in Codex,request_user_input; in Gemini,ask_user. Otherwise, present numbered options in chat and wait for the user's reply before proceeding.
This is a side-route skill that reconnects to the main pipeline at whichever earlier skill the diagnosis points to.
Use /correct-course when any of the following is true:
/research surfaced version changes or constraints that invalidate the /shape closing summary/write-a-prd reveals the problem was misframed during shaping/prd-to-issues reveals that the total work materially exceeds the stated appetite/execute hits an error whose root cause is that the archived research or the PRD is wrong/pre-merge Dimension 4 surfaces a spec-reality mismatch — the code reflects external reality but the declared spec has rotted (non-existent package, renamed public API, version skew past what the research snapshot pinned)/pre-merge surfaces an architectural concern that warrants rework, not advisory findingsDo not use it for:
/qa)/request-refactor-plan)/shape)Ask one question: "What changed that made the current artifact untrustworthy?"
Let the user describe it in their own words. Do not ask follow-up questions yet. You are trying to understand the class of invalidation, not the full history:
/shape summary marked Likely or Uncertain turned out to be wrongWrite down the class you think this is, in one sentence. You will use it in Step 2.
Based on the trigger, determine the earliest skill whose output is now untrustworthy. The rule of thumb:
| Trigger | Earliest affected skill |
|---|---|
/shape assumption is wrong | /shape |
| Version or API reality diverges from the archived research | /research |
| Problem was misframed — solving symptom, not cause | /shape |
| Appetite or no-gos need to change | /write-a-prd |
| Solution direction is wrong but the problem and appetite still hold | /write-a-prd |
| Slice decomposition reveals work exceeds appetite | /write-a-prd (reshape), not /prd-to-issues (re-decompose) |
| Boundary map contract was wrong, slices are still valid | /prd-to-issues |
| Slice is correct but execution approach hit a wall | /execute (not a backtrack — try a different approach) |
Architectural concern from /pre-merge warrants rework | /execute, possibly /request-refactor-plan first |
Explore to confirm. Read the PRD issue, any slice issues, the research archive entry for this feature, and the relevant parts of docs/solutions/. The goal is to be specific: not "shaping is stale" but "the Ably-vs-Pusher decision in the research is wrong because Ably raised its free-tier limits and Pusher dropped theirs."
State the diagnosis to the user in one or two sentences and ask one question: "Does that match what changed?" If no, refine. If yes, proceed.
Now list every durable artifact that needs to be dealt with before forward progress can resume. Be specific — name the exact issue number, the exact file path. Common stale artifacts:
~/.claude/research/<repo-slug>/<feature-slug>-<date>.md) — if the earliest affected skill is /shape, leave the archive entry in place; it is point-in-time history and cannot mislead downstream skills that no longer reference it. If the earliest affected skill is /research, run /research again to produce a new dated entry superseding the old one; the prior snapshot remains in the archive as context. Archive entries are not deleted during backtracking — they live outside the repo and cost nothing to keep.For each artifact, name it, name what should happen to it (delete, update in place, close with comment), and note the one-line reason. Present the full list to the user as a numbered plan.
Walk the cleanup one artifact at a time. For each:
gh issue close, gh issue comment, or direct file deletion as appropriate.Do not batch the cleanup into a single confirmation. Each artifact is a small, reversible decision, and batching increases the chance that a subtle mistake slips through. The cost of confirming each is small; the cost of closing the wrong issue is real.
If the user hesitates on any artifact, skip it and flag it as "needs review" in the final summary. Do not press.
After the cleanup is complete, state explicitly:
Suggest starting the next skill in a fresh session if the conversation has been long. The closing summary is the compressed handoff.
/qa first to diagnose whether it is a bug or a stale-artifact problem (/qa's per-issue depth check delegates to /triage-issue when root cause is needed)./improve-codebase-architecture or /request-refactor-plan.references/SYSTEM-OVERVIEW.md is still the canonical description of what backtracking means in Skill Kit — this skill makes that reference invocable. For deeper reading, the repo original is at <https://github.com/chrislacey89/skills/blob/main/SYSTEM-OVERVIEW.md>./shape, /research, /write-a-prd, /prd-to-issues, or /execute depending on the diagnosis~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.