Add Seed Skills

Add Seed Skills

Pipeline: seed-skills/<slug>/SKILL.md -> validator -> seed.ts upsert into the PROD database -> live catalog. These are PRODUCT catalog skills (Zod schema with testingTypes/languages), not Claude Code workflow skills; never confuse the two formats.

Format contract (parser reality)

packages/web/src/db/seed.ts parses frontmatter with regex, not a YAML library:

• Every value on ONE line; a wrapped description parses as truncated garbage
• Arrays INLINE ONLY: tags: [a, b, c]. YAML block lists (- item) parse as EMPTY arrays
• Zod limits: name 1-100 chars, description 10-500 chars, version semver, testingTypes >= 1, languages >= 1
• Allowed values come from packages/shared/src/constants/ (testing types, frameworks, languages, domains, agent slugs); check there before inventing one

Template (model the body on seed-skills/playwright-e2e/SKILL.md):

markdownCopy code
---
name: Human Readable Skill Name
description: One line, 10-500 chars, what the skill teaches an agent to do.
version: 1.0.0
author: thetestingacademy
license: MIT
tags: [tag-one, tag-two]
testingTypes: [e2e]
frameworks: [playwright]
languages: [typescript]
domains: [web]
agents: [claude-code, cursor, github-copilot, windsurf, cline]
---

# Skill Title

Real instructions: principles, project structure, code samples, checklists.
This body becomes fullDescription: it renders on the skill page and is what
the CLI downloads. Aim for 100+ lines of substance; an empty or thin body
is a broken product page.

Steps

1. Dedup the slug (directory name = slug)

bashCopy code
ls /Users/promode/qaskills/seed-skills | grep -i "<core-term>"
curl -s -o /dev/null -w '%{http_code}\n' "https://qaskills.sh/api/skills/<slug>"   # want 404

2. Write seed-skills/<slug>/SKILL.md

Frontmatter per the contract above, then the full body.

3. Validate every file

bashCopy code
cd /Users/promode/qaskills
pnpm --filter @qaskills/shared build && pnpm --filter @qaskills/skill-validator build
for d in <slug-one> <slug-two>; do
  node packages/skill-validator/dist/cli.js "seed-skills/$d/SKILL.md" || echo "INVALID: $d"
done

Also eyeball that arrays are inline and the description is one line (the validator uses the real YAML parser and can pass files the seed regex still mangles).

4. Seed PRODUCTION (the dangerous step)

Never use the DATABASE_URL from .env.local. It points at a stale non-prod database; seeding it changes nothing on the live site. This mistake has burned a full session before.

1. Baseline: curl -s 'https://qaskills.sh/api/skills?limit=1' and record the top-level total.
2. Get the prod URL: run vercel env pull .env.vercel-prod --environment=production yourself (worked from the agent as of 2026-07-07; the file is gitignored). If the pull is blocked, ask the user to run that exact command and wait.
3. Seed with the explicit URL (strip any surrounding quotes from the value):

bashCopy code
cd /Users/promode/qaskills
export DATABASE_URL='<prod-url-without-quotes>'
pnpm --filter @qaskills/web db:seed

seed.ts is an upsert (onConflictDoUpdate on skills): safe to re-run, it will not delete the live rows that exist only in prod. Never substitute custom SQL, and never run UPDATE/DELETE against prod without explicit user approval.

5. Verify live (the only proof that counts)

bashCopy code
curl -s 'https://qaskills.sh/api/skills?limit=1'                       # total must equal baseline + N
curl -s "https://qaskills.sh/api/skills/<slug>/content" | head -20     # frontmatter + body, not empty
curl -s -o /dev/null -w '%{http_code}\n' "https://qaskills.sh/skills/thetestingacademy/<slug>"  # 200 (adjust author segment if different)

total unchanged, or slug 404s => you seeded the wrong database. STOP. Do not retry blindly; re-verify which URL was exported and report the mismatch.

6. Commit

bashCopy code
git -C /Users/promode/qaskills add seed-skills/<slug-one> seed-skills/<slug-two>
git -C /Users/promode/qaskills commit -m "feat(skills): add <N> <theme> seed skills"
git -C /Users/promode/qaskills push origin main

No web redeploy is needed for catalog changes (pages read the DB), but commit so the repo stays the source of truth.

Failure modes

Red flags

• Exporting DATABASE_URL from .env.local "because it is right there"
• Skipping the baseline/after total comparison
• Frontmatter-only SKILL.md ("body later")
• Writing DELETE/UPDATE SQL to "fix" prod data