Skillpack Creator

Overview

Turn a successful task into a reusable SkillPack. Extract the stable workflow, decide what belongs in a local skill versus pack-level prompts, generate the pack structure, and package it only after the workflow is explicit and repeatable.

Workflow

1. Normalize the source task

Reduce the finished task into a clean execution spec:

• Capture the user goal, concrete deliverable, and the final successful workflow (not the full exploratory transcript).
• List required skills, tools, files, secrets, and environment assumptions.
• Separate deterministic steps from heuristic steps; remove dead ends and debugging noise.
• If the task is still too broad, narrow the scope instead of writing a vague mega-skill. If key success conditions depend on hidden human judgment, mark the pack as a best-effort assistant workflow.

Ask for missing stable facts or infer only the low-risk pieces.

2. Decide what the pack should contain

• Local skill (skills/): reusable procedural knowledge. Keep scripts minimal unless reproducibility depends on exact file generation or repetitive shell steps.
• Scripts (scripts/): repeated shell or file-generation logic where reliability matters.
• References (references/): detailed schemas, API notes, or conventions that should not bloat SKILL.md.
• Prompts (skillpack.json): 1–3 pack-level starter inputs for the UI — not a DAG or state machine. See references/skillpack-format.md for exact pack semantics.

3. Create the pack specification

Before writing files, define the pack spec. Prefer one local orchestrator skill plus a small number of external skills. Example minimal manifest:

json
{
  "name": "company-research",
  "description": "Research a company and produce a summary report",
  "version": "1.0.0",
  "prompts": ["Research {company} and create a report with financials and competitors"],
  "skills": [
    { "name": "research-orchestrator", "source": "./skills/research-orchestrator", "description": "Orchestrate company research across multiple sources" }
  ]
}

4. Create the local orchestrator skill

Create skills/<skill-name>/SKILL.md with frontmatter and imperative workflow instructions:

yaml
---
name: research-orchestrator
description: "Orchestrate multi-source company research. Use when the user wants a structured company report covering financials, competitors, and market position."
---

• Write the stable workflow as imperative steps in the body.
• Add scripts/ only for fragile or repeated operations; add references/ only for detailed information.

5. Materialize the pack

Use scripts/scaffold_skillpack.py when you have the pack spec:

bash
# Basic
python3 skills/skillpack-creator/scripts/scaffold_skillpack.py \
  --manifest /tmp/skillpack.json \
  --output /absolute/path/to/output-pack

# With zip
python3 skills/skillpack-creator/scripts/scaffold_skillpack.py \
  --manifest /tmp/skillpack.json \
  --output /absolute/path/to/output-pack \
  --zip

The script validates the manifest, writes skillpack.json, creates skills/, copies start.sh/start.bat from templates/, and optionally runs npx -y @cremini/skillpack zip.

6. Validate the result

Before handing the pack back, confirm:

• The manifest matches the intended pack scope
• Every declared skill has a valid name, source, and description
• Local skills are present under the target pack's skills/
• Starter prompts are concrete enough to reproduce the workflow
• Zip only after the pack runs as a directory

Output Standard

Produce:

1. A short summary of the stabilized workflow.
2. The target pack structure and skill inventory.
3. The created or updated local skill files.
4. The generated skillpack.json.
5. Whether the pack was zipped and where the zip lives.