dashscope-patent-flowchart

Source-anchored workflow for generating accurate standalone biological process-step figures with DashScope Qwen Image.

Structured as atoms → molecules → compounds:

• capabilities/ — atomic, single-purpose, near-deterministic operations
• composites/ — scoped workflows that chain atoms in explicit order
• SKILL.md — the compound playbook: mode selection, ordering, invariants

Design

A source-verified mapping card is the only artifact that may introduce scientific claims.

SOURCE
  ↓  PROJECT ANALYSIS: entity map + Material Bible + style sheet
  ↓  MAPPING CARD  ── Gate 1: human check vs open source (strict) or DRAFT (agent)
  ↓  PROMPT ASSEMBLY
  ↓  VERSIONED RENDER  ── Gate 2: fixed content/topology probes
  ↓  ACCEPTED RENDER
  ↓  LABELS  ── Gate 3: label-to-shipped-pixel fit
LABELED STANDALONE FIGURE

• Pixels show schematic instances; labels and metadata carry source-level claims.
• Material Bible and style sheet are separated; style and reference images never decide science.
• Renderer details live in an adapter (adapters/), keeping the core workflow renderer-agnostic.

Repository layout

Path	Role
SKILL.md	compound playbook (orchestration + invariants)
composites/	molecules: analyze-project, plan-step, render-step, finalize-step
capabilities/	atoms: extract-source-text … build-evidence-package
schemas/	JSON Schema for material bible / mapping card / render metadata / probe
adapters/dashscope.md	renderer contract (DashScope Qwen Image)
scripts/generate_image.py	generic versioned renderer
examples/hydrogel_patent/	worked example artifacts

Requirements

• Python 3.9+
• pip install dashscope

Quick start

# 1. local secret (never committed)
cp .env.example .env        # fill in DASHSCOPE_API_KEY

# 2. render one step figure from a prompt file
python scripts/generate_image.py \
  --prompt-file step4_prompt.txt \
  --step-id step4 \
  --verification-status DRAFT/UNVERIFIED \
  --output step4_v01.png \
  --meta step4_meta.json

The renderer refuses to overwrite existing files and writes a sibling metadata JSON that satisfies schemas/render_metadata.schema.json. For verified output, run the playbook molecules in order: analyze-project → plan-step (Gate 1) → render-step (Gate 2) → finalize-step (Gate 3). See SKILL.md and the composites/ docs for the full protocol.

Status

This repo is an agent skill / workflow spec with a working renderer script — the Markdown capabilities (entity analysis, mapping, probes, label fit) are executed by an agent following the playbook. It is not yet a one-command CLI orchestrator (plan / render / verify subcommands); that automation layer is a planned extension.

Examples

examples/hydrogel_patent/ ships the committed text artifacts of one complete step-4 walkthrough: material_bible.yaml, step4_mapping.yaml, step4_prompt.txt, step4_probe.yaml. The PNG and its metadata are generated artifacts — regenerate them with the renderer (requires a DashScope key), then run the probe and record the real verdicts.

Security

DASHSCOPE_API_KEY is never committed. Only .env.example is tracked; the real key lives in a local .env, which is gitignored.

License

MIT — see LICENSE.