AI skills for capturing and pressure-testing how work actually happens. Business processes live in people's heads; these skills turn them into accurate, inspectable artifacts — and audit the artifacts that already exist. Some skills generate (transcripts → service blueprint), others assess (score, decompose, and trace an existing diagram). The collection grows over time.
| Skill | Type | What it does |
|---|---|---|
| service-blueprint | Generate | Turns discovery-call transcripts into an interactive HTML service blueprint: swimlanes, systems of record, the core happy path vs. deviations, decision points, and pain points — with print/PDF output. |
| assess-diagram-quality | Assess | Report card. Scores a diagram across seven quality dimensions, gives an overall band, and separates what it can judge from the artifact from what a human must verify. Start here for existing diagrams. |
| pressure-test-diagram | Assess | Decomposition scalpel. Flags steps that hide complexity or add noise, then emits a before/after inventory and a fix checklist. |
| interrogate-data-sources | Assess | Provenance scalpel. For each box, identifies which system every datum comes from, tags the box, and maps identifiers across system boundaries. |
They compose: generate an as-is blueprint from transcripts with service-blueprint, then run the assessment skills against it — or against any process diagram you already have.
- Audit, not fabricate. Where a data source, owner, or step can't be determined from the source material, the skills flag it as an open question (or lower its confidence) rather than inventing an answer.
- Evidence-grounded. service-blueprint cites which transcript supports every step and tracks second-hand hearsay; the assessment trio builds one shared numbered
Box #inventory so their reports line up box-for-box. - Internal quality ≠ external truth. A clean artifact means the artifact is sound, not that it matches reality. Every output names what a human who runs the process still has to confirm.
These follow the Agent Skills open standard — one <name>/SKILL.md folder per skill — so they work across several agents. Skills do not sync between surfaces; install them wherever you want to use them.
Claude Code — copy the folders in; no upload step. Invoke with /service-blueprint (etc.), or just describe the task and let them auto-activate.
cp -r skills/* ~/.claude/skills/ # personal — all your projects
cp -r skills/* .claude/skills/ # project — checked in for the team
They can also ship inside a Claude Code plugin for one-command install from a marketplace.
Claude.ai / Claude Desktop — zip a skill folder and upload it under Settings → Features → Skills. Requires a Pro, Max, Team, or Enterprise plan with code execution ("Create and edit files") enabled; uploaded skills are per-user, not org-wide.
cd skills && zip -r service-blueprint.zip service-blueprint
Claude API (and Claude on AWS / Microsoft Foundry) — upload via the /v1/skills endpoint, then reference the returned skill_id in the container parameter of a request that uses the code execution tool (beta header skills-2025-10-02). Shared workspace-wide.
GitHub Copilot — Copilot reads the same standard; drop the folders in .github/skills/ (repo) or ~/.copilot/skills/ (personal). Works in VS Code, JetBrains, the Copilot CLI, and the cloud agent.
Any other agent — the format is just Markdown. Point the agent at a SKILL.md and it has everything it needs.
Sandbox-friendly. Every skill here is Markdown plus — for service-blueprint — one stdlib-only Python script and a fully self-contained HTML template. No network calls, no package installs, so they run as-is inside the locked-down API and claude.ai containers.
skills/
assess-diagram-quality/SKILL.md
interrogate-data-sources/SKILL.md
pressure-test-diagram/SKILL.md
service-blueprint/
SKILL.md # entry point + pipeline overview
README.md # human-facing overview
references/ # extraction, synthesis, schema, rendering docs
assets/blueprint-template.html # self-contained render target (single-page PDF export)
examples/blueprint-model.example.json
scripts/render.py # stdlib-only validate + render
theme-source/ # provenance: the "Warm Gray Ember" theme snapshot
Each SKILL.md carries YAML frontmatter (name, description, and optionally argument-hint, allowed-tools, model) followed by the skill instructions. Edit these files directly — they are the source of truth.
MIT — see LICENSE.