business-blueprint-skill（商业蓝图技能）

Business Blueprint Skill

Use the Python scripts in this repository as the execution surface.

Output Directory

All generated files (blueprint JSON, viewers, exports) go into projects/workspace/ — not the repository root.

python -m business_blueprint.cli --plan projects/workspace/solution.blueprint.json --from "..."
python -m business_blueprint.cli --project projects/workspace/solution.blueprint.json
python -m business_blueprint.cli --export projects/workspace/solution.blueprint.json

Industry Selection

Choose --industry from exactly one of: "common", "finance", "manufacturing", "retail". Select the closest match based on the user's domain and materials; do not invent other values.

How to Generate a Blueprint

The AI agent is responsible for entity extraction. The Python tool handles JSON writing, visualization, and export.

Step 1: Read industry hints

Read the seed template at business_blueprint/templates/{industry}/seed.json and get the industryHints.checklist.

Step 2: Extract entities from source text

Using the user's source material AND the industry hints checklist, extract:

• capabilities: business capability areas (name, description)
• actors: roles/people involved (name)
• flowSteps: business process steps (name, actorId, capabilityIds, stepType)
• systems: IT systems that support capabilities (name, description, capabilityIds)

Step 3: Write the blueprint JSON

Write the JSON file directly to the output path. Use this schema:

{
  "version": "1.0",
  "meta": {
    "title": "...",
    "industry": "retail",
    "revisionId": "rev-YYYYMMDD-NN",
    "parentRevisionId": null,
    "lastModifiedAt": "ISO8601",
    "lastModifiedBy": "ai"
  },
  "context": {
    "goals": [],
    "scope": [],
    "assumptions": [],
    "constraints": [],
    "sourceRefs": [{"type": "inline-text", "excerpt": "..."}],
    "clarifyRequests": [],
    "clarifications": []
  },
  "library": {
    "capabilities": [
      {"id": "cap-xxx", "name": "...", "level": 1, "description": "...", "ownerActorIds": [], "supportingSystemIds": []}
    ],
    "actors": [
      {"id": "actor-xxx", "name": "..."}
    ],
    "flowSteps": [
      {"id": "flow-xxx", "name": "...", "actorId": "actor-xxx", "capabilityIds": ["cap-xxx"], "systemIds": [], "stepType": "task", "inputRefs": [], "outputRefs": []}
    ],
    "systems": [
      {"id": "sys-xxx", "kind": "system", "name": "...", "aliases": [], "description": "...", "resolution": {"status": "canonical", "canonicalName": "..."}, "capabilityIds": ["cap-xxx"]}
    ]
  },
  "relations": [
    {"id": "rel-xxx", "type": "supports", "from": "sys-xxx", "to": "cap-xxx", "label": "支撑"}
  ],
  "views": [],
  "editor": {"fieldLocks": {}, "theme": "enterprise-default"},
  "artifacts": {}
}

Step 4: Generate visualizations

python -m business_blueprint.cli --export <blueprint.json>

This generates SVG + HTML viewer by default. Use --format drawio|excalidraw|mermaid for other formats.

Export View Selection Policy

Treat export view choice as a routing decision, not a styling preference.

• If a request matches a supported, standard export template, use that template.
• If there is no standard export template for the requested diagram, fall back to freeflow.
• Do not substitute swimlane, matrix, product tree, or other generic views just because they are available.
• When embedding a blueprint diagram into a report or ad hoc analysis, freeflow is the safe default unless the user explicitly asks for a supported standard template.

Step 5: Generate downstream projection

python -m business_blueprint.cli --project <blueprint.json>

This generates solution.projection.json, the canonical machine projection consumed by downstream report/slide workflows.

Workflow Decision Tree

User provides raw requirements / meeting notes?
  → AI agent reads hints, extracts entities, writes blueprint JSON
  → Optionally run --project for downstream machine handoff
  → Then run --export for visualization

User needs diagram files (SVG, draw.io, etc.)?
  → --export (default: SVG + HTML viewer)

User unsure about blueprint quality?
  → --validate

User wants downstream report / slide generation?
  → --project

Commands

Export Formats

Collaboration Boundary

This skill produces semantic intermediate artifacts. Downstream skills consume them:

• report-creator consumes solution.projection.json → assembles reports
• slide-creator consumes solution.projection.json → assembles presentations
• Other skills may consume relations → generate PlantUML or other diagram syntax
• Downstream skills should never directly edit solution.blueprint.json
• solution.handoff.json is viewer-only metadata, not a downstream narrative input

Sandbox Execution

When running in an isolated Python sandbox (Jupyter, notebook, cloud REPL) that auto-installs dependencies:

1. The sandbox auto-installs kai-business-blueprint from PyPI. Do NOT try to pip install -e or use sys.path hacks with __file__ — __file__ is undefined in Jupyter.

1. For --export (after blueprint JSON exists):

   from pathlib import Path
   from business_blueprint.model import load_json
   from business_blueprint.export_svg import export_svg_auto
   from business_blueprint.export_html import export_html_viewer

   blueprint_path = Path("solution.blueprint.json")
   blueprint = load_json(blueprint_path)
   export_dir = blueprint_path.with_name("solution.exports")
   export_dir.mkdir(parents=True, exist_ok=True)
   # Default fallback is freeflow unless a standard export template applies.
   export_svg_auto(blueprint, export_dir / "solution.svg")
   export_html_viewer(blueprint, blueprint_path.with_name("solution.blueprint.html"))

1. Prohibited patterns in sandbox:

- __file__ — undefined in Jupyter - sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) — will raise NameError - subprocess.run(["business-blueprint", ...]) — sandbox runs Python cells, not shell - os.system() — same reason

Architecture Diagram Generation

When user requests an architecture diagram (keywords: "架构图", "architecture diagram", "--export", "diagram"):

1. Read references/architecture-design-system.md for the complete design system.
2. Read the appropriate template from references/architecture-templates/ based on the user's domain:

- AWS/Serverless/Lambda → serverless.md - Microservices/Kubernetes/微服务 → microservices.md - Other → use serverless.md as a structural reference

1. Read the blueprint JSON to extract entities and flow steps.
2. Generate a self-contained HTML file with inline SVG following the design system rules.
3. Write the output file to the same directory as the blueprint JSON.

If the request does not match one of the supported standard templates above, stay on the default freeflow export path. Do not switch to another generic view type as a fallback. If a standard template would create a squeezed, clipped, or overcrowded diagram, stop using the fixed template geometry and fall back to freeflow or a wrapped multi-row layout.

Route eligibility matrix

Use an explicit route contract before rendering:

Do not invent route heuristics ad hoc inside a renderer. Route eligibility must stay explicit and reviewable.

Generation Rules

• Use dark mode by default (#020617 bg + 40px grid). Only use light mode when the user explicitly asks for it.
• L→R data flow: Clients(左) → Frontend → Backend → Database(右)
• Map systems[].category to semantic colors from the design system
• Map systems[].properties.type == "aws" → AWS Region boundary box
• Map systems[].properties.type == "k8s" → Kubernetes Cluster boundary box
• Use flowSteps[].seqIndex for L→R ordering
• Component sizing: 0-1 cap = small(44px h), 2-4 = medium(80px h), 5+ = large(80px h)
• Layout must be content-driven. Never force every node in a layer into one fixed row if that creates toothpaste-style squeezing.
• When a layer has more than 3 peer nodes, or labels/features become tight, wrap into multiple rows or widen the canvas before shrinking the content.
• Render users/actors as actor labels, badges, or lane headers by default. Do not render them as ordinary system cards unless the user explicitly asks for that visual treatment.
• Legend must live in a bottom safe area and participate in canvas sizing. Never place the legend as a floating overlay in the top-right corner.
• Final SVG/HTML height must be derived from the bottom-most node, legend, summary cards, and footer plus padding. Do not use fixed-height wrappers or overflow: hidden that can clip the last row.
• Z-order: bg → grid → title → region → arrows → nodes → legend → cards → footer
• Component border: rx="8", stroke-width="2"
• Region border: rx="16", stroke-dasharray="8,4", opacity="0.4"
• Geometry-sensitive integrity checks must use the numeric thresholds from evals/export-integrity-thresholds.json, not prose heuristics.

Output

• Single HTML file: {blueprint_stem}.html alongside the blueprint JSON
• No external dependencies (except Google Fonts CDN for JetBrains Mono)
• Opens in any browser, printable to PDF

Error Handling

• If --validate returns errors: fix structural issues before proceeding to --export.
• If --validate returns only warnings: proceed but note the warnings in any handoff.
• If Python version < 3.12: the package will refuse to install. Use python3 -m business_blueprint.cli with system Python as fallback.
• If a specialized route fails integrity: fall back to its configured fallback route.
• If freeflow also fails integrity: export exits non-zero with a structural diagnostics payload instead of emitting a silently broken artifact.

Cross-Platform Scope

Phase 2 does not attempt full Windows terminal parity.

Known deferred cases:

• PowerShell pipe quirks beyond documented CLI contract tests
• console-default encoding issues outside explicit UTF-8 execution paths

Accepted workaround for encoding-sensitive runs:

• use python -m business_blueprint.cli
• set PYTHONIOENCODING=utf-8 where needed