ns-skill-creator

Skill Creator (harness projects)

Create and improve project-local skills in the active harness project. The eval loop, description optimization, and bundled scripts come from the upstream skill-creator (anthropics/skills) — this skill only overrides where files are saved and runs harness sync after each create/update.

Harness discovery

Resolve paths per ../ns-harness/references/harness-discovery.md:

Variable	Value
{product_root}	Repo or monorepo product folder (where AGENTS.md lives)
{skills_canonical}	{product_root}/.agents/skills/

If {harness_root} is missing, tell the user to run npx @nextstage-brasil/harness first — this skill expects a harness-scaffolded project.

Step 0 — Load upstream skill-creator

Harness init installs the anthropics skill-creator bundle (scripts, eval-viewer) at {skills_canonical}/skill-creator/, then installs this wrapper as ns-skill-creator.

For the full upstream workflow, read:

• ~/.agents/skills/skill-creator/SKILL.md when installed globally, or
• anthropics/skills — skill-creator

Use upstream for: intent capture, interview, SKILL.md anatomy, eval/benchmark loop, eval-viewer/generate_review.py, description optimization (run_loop.py), and scripts under {skills_canonical}/skill-creator/scripts/.

Apply the path overrides and post-create sync below on top of upstream instructions.

Path overrides (project)

Artifact	Path	Notes
New/edited skill	{skills_canonical}/<kebab-case-name>/	name frontmatter must match directory
SKILL.md	{skills_canonical}/<name>/SKILL.md	English unless the team defines otherwise in AGENTS.md
references/, scripts/, evals/	Under the skill directory	See references/project-layout.md
Eval workspace	{product_root}/skill-creator-workspace/	iteration-N/eval-<id>/ per upstream
Upstream tooling	{skills_canonical}/ns-skill-creator/	scripts/, eval-viewer/ from anthropics bundle

Never save project skills to skills/ at repo root (that layout is for the nextstage-brasil/skills catalog repo only).

Project conventions

Read references/project-layout.md when drafting. If {product_root}/AGENTS.md defines language or skill conventions, follow them.

Summary:

• Pushy description in frontmatter for triggering (per upstream guidance).
• Keep SKILL.md under ~500 lines; move detail to references/.
• evals/evals.json — 2–3 realistic prompts when evals add value.
• Declare depends: in frontmatter only when the skill references other installed skills (e.g. ns-harness).

Workflow

Follow upstream ns-skill-creator stages (capture intent → draft → eval → improve → repeat). At each write step, use the path overrides above.

Post-create — harness sync

After skill files are written or materially updated, run from {product_root}:

npx @nextstage-brasil/harness sync

This regenerates Claude skill symlinks (.claude/skills/) so Claude Code discovers the new skill. Cursor reads .agents/skills/ directly.

Use npx @nextstage-brasil/harness sync --check locally after editing canonical rules to verify adapters on disk match before committing .nextstage-harness/ changes.

Report sync results to the user. Do not claim the skill is discoverable until sync completes successfully.

Description optimization

After the skill is stable, offer upstream description optimization per upstream ns-skill-creator (run_loop.py). Use the model ID from the active session. Update {skills_canonical}/<name>/SKILL.md frontmatter, then run harness sync if needed.