Add Claude automated code review

[mattpodwysocki]

Summary

Adds .github/workflows/claude-review.yml using anthropics/claude-code-action@v1.

• Auto-reviews PRs with ≥300 changed lines
• Responds to @claude mentions in PR comments
• Uses ANTHROPIC_API_KEY repo secret

Prompt focus areas

Tailored for a markdown skills repo:

1. Accuracy — technical content correct for referenced Mapbox APIs/SDKs
2. Skill structure — valid frontmatter schema per CONTRIBUTING.md
3. Clarity — actionable guidance, complete runnable code examples
4. Consistency — tone and depth match existing skills
5. Scope — skills stay focused, no unnecessary overlap

Setup required

Add ANTHROPIC_API_KEY to repository secrets:
Settings → Secrets and variables → Actions → New repository secret

🤖 Generated with Claude Code

[github-actions]

Claude finished @mattpodwysocki's task in 1m 54s —— View job

---

Review of PR #56 — Add Claude automated code review + presentation deck

• Gather context (read CLAUDE.md, CONTRIBUTING.md, existing skills)
• Review workflow file (claude-review.yml)
• Review presentation files
• Review ignore file changes
• Post review feedback

---

Summary

This PR adds two things: (1) a GitHub Actions workflow for Claude-powered code review, and (2) a Slidev presentation deck covering the Mapbox Agent Skills project. Overall the PR is well-structured and the presentation content is technically sound. A few issues worth addressing:

Workflow (claude-review.yml)

1. Unnecessary id-token: write permission — Since the workflow authenticates via ANTHROPIC_API_KEY (not GitHub OIDC), this permission should be removed for least-privilege. The commit history suggests this was intentionally removed at one point (fcb84e9) but reappeared. See inline comment →

2. Fragile line-count parsing — The git diff --stat | awk approach can miscount because gsub(/[^0-9]/, " ") doesn't distinguish file counts from insertion/deletion counts. git diff --numstat is more reliable. See inline comment →

3. Long claude_args line — The single-line string is ~430 characters. A YAML block scalar would improve readability. See inline comment →

Presentation (slides.md)

The presentation is well-organized, technically accurate, and provides good coverage of the skills ecosystem. A few minor notes:

4. Workflow prompt references non-existent triggers frontmatter field — The review prompt (line 48 of the workflow) says to check for "triggers, content" in skill frontmatter, but per CONTRIBUTING.md the schema is name + description only. Trigger conditions are embedded in the description text. The slide deck's own example (slide on Skill Structure) is correct, but the workflow prompt should be updated to match. See inline comment →

5. README slide list is out of sync — The README.md lists 15 slides in an order that doesn't match the actual deck (which has ~25 slides). Several slides like "Why Agent Skills?", "Why Mapbox Needs Skills", "Skills as a Strategy", "Skills + MCP: Better Together", and the eval-related slides are missing from the README. See inline comment →

Ignore files

The .prettierignore and .markdownlintignore additions are correct — Slidev's --- slide separators conflict with Prettier's markdown HR formatting and markdownlint rules. Good call excluding the presentations/ directory.

What looks good

• The 300-line threshold for auto-review is a sensible filter to avoid noise on small PRs
• The @claude trigger for on-demand review in comments is well-configured
• The presentation content is technically accurate — the marker decision tree thresholds, iOS annotation manager pattern, Android batch update pattern, and MCP integration examples all match current Mapbox SDK patterns
• Code examples in the slides correctly show anti-patterns (❌) vs solutions (✅), consistent with the repo's skill style

---

[github-actions]

The id-token: write permission is only needed for GitHub OIDC token exchange (GitHub App auth). Since this workflow uses ANTHROPIC_API_KEY directly (and the commit history suggests this was intentionally removed at one point), this permission can be dropped to follow least-privilege:

Suggested change

	actions: read
	contents: read
	actions: read

[github-actions]

This awk extraction is fragile — git diff --stat summary format varies (e.g., " 8 files changed, 9953 insertions(+)" vs " 1 file changed, 3 insertions(+), 1 deletion(-)") and gsub(/[^0-9]/, " ") strips all non-digits then sums the first two numbers, which could pick up the file count instead of insertions.

A more robust approach using --numstat:

Suggested change

	CHANGED_LINES=$(git diff --stat origin/${{ github.event.pull_request.base.ref }}...origin/${{ github.event.pull_request.head.ref }} | tail -1 | awk '{gsub(/[^0-9]/, " "); print $1 + $2}')
	CHANGED_LINES=$(git diff --numstat origin/${{ github.event.pull_request.base.ref }}...origin/${{ github.event.pull_request.head.ref }} | awk '{ added += $1; removed += $2 } END { print added + removed }')

[github-actions]

Nit: This line is 430+ characters. Consider using a YAML block scalar for readability and easier maintenance:

claude_args: >-
  --model claude-opus-4-6
  --allowedTools
  "Bash(gh issue view:*),Bash(gh search:*),..."

[github-actions]

The frontmatter mentions triggers in the prompt focus areas (line 48 of the workflow: "triggers, content"), but per CONTRIBUTING.md, the actual SKILL.md frontmatter schema only requires name and description — there's no triggers field. The trigger condition is embedded in the description text (e.g., "TRIGGER when...").

This slide's example is correct (just name + description), but note the workflow prompt references a field that doesn't exist in the schema.

[github-actions]

The slide order listed here doesn't match the actual order in slides.md. For example, "Why Agent Skills?" and "Why Mapbox Needs Skills" appear before the deep dives in the actual deck but aren't listed here, while "Skills + MCP: Better Together" is in the deck but missing from this list. Consider regenerating this list from the actual slide headings to keep it accurate.