Using skills at Triumph

Claude Code is the terminal-first Claude — you run claude inside a git repo and it can read/edit files, run shell commands, and call skills. Skills live in the repo itself, so anyone who clones gets the same set. Good fit for engineering workflows: code review, migrations, test writing, refactors, SQL against the warehouse, on-call triage.

Skills are one piece of a broader .claude/ folder you commit to the repo. Everything in here is shared with the team on the next pull. Here's the layout Rock RMS uses — it's a good reference for any Triumph repo:

• skills/ — on-demand workflows that fire when you ask. This is what the marketplace distributes. Each is a folder with a SKILL.md plus optional references/scripts.
• commands/ — short named slash commands like /build, /test, /check. One markdown file per command. Simpler than a skill; good for frequent repo actions.
• rules/ — context Claude loads automatically based on file paths. code-conventions.md and data-model.md are always on; path-specific ones (e.g. block-architecture.md) kick in when you're editing files that match.
• hooks/ — safety guards that intercept commands (e.g. blocking git push --force to protected branches).
• settings.json — shared permissions. Pair it with settings.local.json (gitignored) for personal overrides.

On a skill's detail page, hit Download zip and unzip into .claude/skills/<slug>/. Or copy the install command and paste it into the repo:

$ cd your-project
$ claude skill install triumph/review-prd

✓ Fetched review-prd from Triumph Skills
✓ Added .claude/skills/review-prd/
✓ Indexed 1 new skill — ready to use.

Commit the new folder so the skill ships with your repo.

$ git add .claude/skills/review-prd
$ git commit -m "Add review-prd skill"

Two ways to fire a skill:

• Explicit slash command. /bugfix Finance/TransactionList.ascx.cs resets on page load — always works, always unambiguous. Use this when you need guaranteed behaviour.
• Natural-language match. fix this bug in TransactionList — the grid filter resets on reload — Claude decides whether to invoke the skill based on what you said and what the skill claims to do. Feels magic when it works, confusing when it doesn't.

You'll see skill: bugfix in the output header when a skill fires.

A skill is a markdown file with a YAML frontmatter block — that's it. Here's a minimal one:

---
name: review-prd
version: 1.4.0
description: >
  Structured review of a product PRD.
  Flags missing sections, fuzzy success metrics,
  and undefined dependencies.
runtimes: [code, cowork]
triggers:
  - "review this PRD"
  - "/review-prd"
  - "check .* for gaps"
tags: [product, review]
---

# Review PRD

You are reviewing a PRD for completeness and clarity.

## Checklist
1. Has a one-line problem statement…
2. Success metrics are quantified…
3. Dependencies are named…

## Output format
Return a markdown table with columns:
"Section | Status | Gap | Suggested fix"

Everything below the frontmatter is prose that gets prepended to Claude's context when the skill fires. Keep it tight — every token you add here is a token Claude reads on every invocation.

• Update: claude skill update review-prd pulls the latest version. Skills in your repo pin to a version, so teammates won't get surprise updates.
• Uninstall: delete the folder. rm -rf .claude/skills/review-prd.
• Disable temporarily: add the skill id to .claude/settings.json under "disabled_skills".
• Publish a new version: bump version in SKILL.md, re-zip, click Publish new version on the detail page. No review queue — your version goes live immediately.

• Claude ignores the skill. Run claude skill list to confirm it's loaded. If not, check SKILL.md has valid YAML frontmatter (closing ---).
• Wrong skill fires. Triggers can collide. Be explicit with /skill-id or tighten the triggers regex.
• Skill worked yesterday, broken today. Run claude skill doctor review-prd — it reports version drift, stale assets, and malformed frontmatter.
• Skill has a bug. Use Request modification on the detail page — the author gets a notification. For something unsafe, hit Report.