Skill Authoring Goal Prompts

# Skill-authoring goal prompts (deliverable is itself an agentskills.io SKILL.md package)

Use this reference when the user's prompt asks the generator to produce a goal whose **deliverable is a coding-agent-agnostic agent skill** (a `SKILL.md` + optional `references/`, `scripts/`, `templates/`, `assets/`, conforming to https://agentskills.io/specification). Real example: "create a coding-agent-agnostic agent skill named `design-with-pencil`". This is a recurring class — encode it once.

## Distinguishing signals

- The deliverable is a directory shaped like `<skill-name>/SKILL.md` + supporting dirs.
- The user names a methodology repo (e.g. `srinitude/design-craft`), a CLI/SDK to wrap (e.g. Pencil CLI, OpenHue), and a spec URL (agentskills.io, anthropic guide PDF, llms.txt).
- The skill must be agent-agnostic: no Claude-/Cursor-/Windsurf-/Replit-/IDE-specific behavior, no model-provider-specific syntax.
- The user demands exactly one deterministic workflow path through the skill.

## Required research (front-load before the helper runs)

1. **`firecrawl map --limit 5000 --json --pretty -o research/url-maps/<tech>.json`** for every doc root the user cites:
   - The wrapped tool's official docs root (e.g. `https://docs.pencil.dev/`).
   - The agentskills.io site (`https://agentskills.io`).
   - Each map file's `data.links` is a list of strings or `{url, title, description}` objects depending on doc shape — defensively check `isinstance(link, str)` before using it as a URL.
2. **`firecrawl scrape <page> --format markdown --json -o research/scrapes/<page>.json`** for the most important pages (per `firecrawl-cli-1-16-quirks.md` §3a-pre — always `-o`, never piped to stdin):
   - The wrapped tool's CLI reference page (full command surface, every flag, every subcommand, every env var, every supported MCP tool, every example invocation).
   - `https://agentskills.io/specification` — the authoritative frontmatter contract (name ≤64 chars, description ≤1024 chars, optional `license`/`compatibility`/`metadata`/`allowed-tools`, optional `scripts/` `references/` `assets/` dirs, progressive disclosure).
   - `https://agentskills.io/skill-creation/best-practices` — start from real expertise, design coherent units, moderate detail, progressive disclosure, plan-validate-execute, bundle reusable scripts.
   - `https://agentskills.io/skill-creation/optimizing-descriptions` — how skill triggering works, trigger eval queries, should/should-not split.
   - `https://agentskills.io/skill-creation/quickstart` — minimal directory shape (`.agents/skills/<name>/SKILL.md`).
   - `https://agentskills.io/skill-creation/using-scripts` — design scripts for agentic use (no interactive prompts, `--help` docs, helpful error messages, structured output, version-pinned `uvx`/`npx`/`bunx`).
3. **`opensrc fetch <org>/<repo>`** for any methodology repo the user names. Read its `SKILL.md` (if it ships one), its `README.md`, its `references/` directory, and its `scripts/` directory deeply enough to enumerate:
   - The methodology's pipeline / phases / structure (e.g. Design Craft's THINK→RESEARCH→SHAPE→BUILD→VERIFY→ITERATE→SHIP).
   - The reference catalog and what each reference governs.
   - Naming conventions, design primitives, component hierarchy guidance, validation practices, anti-patterns.
   - Any bundled scripts, schemas, templates that the authored skill should leverage or analogously bundle.
4. **Anthropic skill-building PDF** (`https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf`) — `firecrawl scrape ... --format markdown --json -o ...` works on the PDF URL; otherwise extract via `firecrawl agent --schema-file ... --urls ...` per SKILL.md step 11.

Persist every research artifact under `~/.hermes/goal-prompts/research/{url-maps,scrapes}/` so the YAML's `validation_evidence.firecrawl.required_maps_present[*]` and `.evidence[]` can cite real on-disk paths with real link counts.

## Pre-helper prompt construction (file-fed, not heredoc)

Per SKILL.md pitfall #16, write the full source prompt to `/tmp/<project>_prompt.txt` first, then call: