Skillsaw Onboard

skillsaw Onboard

Onboard this repository to skillsaw, a linter for agentic contextual building blocks (CLAUDE.md, skills, plugins, agents, hooks, etc.).

Workflow

Ask one routing question at a time and wait for the answer. An explicit choice in the user's request already counts as an answer. Read a reference only after its condition or a yes answer routes to it; do not read the reference to formulate the question. After completing it, return here. If the answer is no, continue to the next checkpoint without reading it. Carry forward the command prefix, counts, choices, and changed-file list.

Resolve every references/... path relative to the directory containing this SKILL.md, never relative to the target repository or the process's current working directory. If this file was fetched from the web, resolve each reference against the parent URL of this file and fetch it from that sibling location.

Replace brace-delimited fields below with facts from the repository or scan; never show placeholders to the user, and render singular or plural wording naturally.

1. Establish the current state

If no working skillsaw command is known, read install. Then read initial scan and report its violations before offering changes.

2. Triage findings by rule

Run skillsaw lint --format json -v to see all violations, including the info-level ones the first scan hid. Group the results by rule_id and sort them by count. Sample 3–5 examples from any large cluster (e.g. >10% of total findings or >20 issues) to understand the root cause before deciding on an action.

Follow triage to categorize each group into Fix now, Baseline, or Configure, and present a clear summary table to the user for confirmation before making changes. Carry the agreed buckets into the subsequent steps.

3. Apply autofixes

If the Fix now bucket holds autofixable findings, ask:

The plan includes autofixes for {count} violations: {safe count} safe and {suggest count} suggested. Applying them will edit the affected context files; I will show the changes and lint again afterward. Should I apply those fixes now?

If yes, read autofix. If no, preserve the count.

4. Make judgment-based fixes

If violations remain, offer manual fixes and a baseline as alternatives. A baseline is especially useful when hundreds of findings would otherwise block adoption or the user wants to move forward from a clean starting point. Ask:

{count} violations remain and need judgment rather than a mechanical fix. I can inspect and fix them now, or review them as accepted existing debt and baseline them in step 6 so onboarding can move forward while CI catches new findings. Which path would you prefer?

If the user chooses fixes, read manual fixes. If the user chooses a baseline, review the remaining findings by rule, severity, and affected paths. The baseline command records every eligible finding that remains, so fix, suppress, or otherwise remove anything the user does not accept before continuing; pause onboarding if that cannot be done safely. Then preserve the accepted remaining set for the baseline decision. Do not require accepted findings to be fixed first.

5. Add or update configuration

If the triage plan placed any rule in Configure, read configuration without asking again; the user confirmed those settings in step 2. Otherwise, if .skillsaw.yaml is missing, ask:

This repository has no .skillsaw.yaml. I can add the default tracked configuration so rule settings and exclusions have an explicit place; lint behavior remains at the defaults until it is customized. Should I create it?

If yes, read configuration.

6. Baseline accepted violations

If reviewed violations remain, ask:

{count} reviewed violations remain. A baseline records them in .skillsaw-baseline.json as accepted debt so CI fails only on new violations; it does not fix them, and the file can shrink as they are fixed. Should I create or update that baseline?

If yes, read baseline.

7. Add skillsaw CI

Ask:

I can add skillsaw to {detected CI system or "your CI"} so future changes are linted automatically. GitHub uses a pinned, read-only lint workflow; optional PR comments use a separate write-enabled workflow. GitLab gets a Code Quality job. Should I set this up? If yes, which CI system?

If yes, read CI, using the named CI system.

8. Add external link checking

If GitHub Actions is available, ask separately:

External links decay over time and need recurring audits, but checking them during every skillsaw lint would add network access, latency, and failures caused by third-party outages. We recommend a dedicated link checker such as Lychee for this job. I can add a weekly GitHub Actions workflow that runs Lychee on a schedule or manually, outside your pull-request merge gate. This adds .github/workflows/link-check.yml and a pinned third-party Action to maintain. Would you like me to set that up?

If yes, read external links.

9. Add local commands

Ask:

I can add version-pinned lint and lint-fix Makefile targets for repeatable local use through uvx or a container, without overwriting existing targets. This creates or edits Makefile. Should I add them?

If yes, read Makefile.

10. Add a grade badge

If the repository has a README, ask:

I can generate .skillsaw-badge.json and add a skillsaw grade badge to the README. The badge reflects the committed lint result, ignores the baseline, and must be regenerated when agent context changes. Should I add it?

If yes, read badge.

11. Verify the result

After all accepted routes finish, always read verification.