LLLin000/dashscope-patent-flowchart

Biological process-step figures — compound playbook

Generate accurate standalone figures for biological, biomaterials, tissue-engineering, and laboratory process steps.

Trust boundary: Only source-verified mapping cards may introduce scientific claims. Render inspection verifies pixel conformance to the mapping card; it cannot upgrade or create source claims. Style and reference images never decide science.

Structure: atoms → molecules → this playbook

This skill is a compound (playbook). It does not contain the procedure itself — it orchestrates molecules (composites), which explicitly chain atoms (capabilities).

compounds (this playbook): choose mode, run molecules, hold judgment
  └─ composites/: scoped workflows, explicit atom order (2-6 atoms each)
       └─ capabilities/: single-purpose primitives, near-deterministic

Depth is capped at two levels on purpose: molecules call atoms directly, atoms never call other atoms, the playbook never reaches past molecules. A deeper graph is where agent reliability falls apart.

• Atoms (capabilities/): one goal, one artifact, mechanical. No judgement calls about orchestration.
• Molecules (composites/): a scoped task with explicit "call atom X, then atom Y" instructions. Minimizes runtime decision-making.
• This playbook: mode selection, ordering, escalation, and the invariants that overrule everything.

Molecule map

Molecule	File	Chains atoms
analyze-project	composites/analyze-project.md	extract-source-text → analyze-entities → build-material-bible → extract-style-sheet (if reference)
plan-step	composites/plan-step.md	write-mapping-card → check-mapping-card → Gate 1
render-step	composites/render-step.md	assemble-prompt → render-image → run-probe (iterate within budget)
finalize-step	composites/finalize-step.md	fit-labels (+ build-evidence-package on request)

Intake — ask before starting

At task start, clarify the user-set options in one round with the question tool (ask), before touching files:

1. Task type — routes the workflow (table below).
2. Task mode — quick (exploration, stays DRAFT) or verified (final deliverable).
3. Gate 1 mode — strict_verified (human source check) or agent_draft (no human gate).
4. Source — patent / paper / protocol / DOCX / PDF / pasted text; path or file.
5. Reference image — provided (style only) or none.
6. Iteration budget per step (default 5 renders).
7. Deliverables — PNG only, PNG + labels, or PNG + labels + evidence package.
8. Model — default qwen-image-3.0-pro, or fetch the catalog (capabilities/list-image-models.md, scripts/list_image_models.py) and let the user pick; record choice in model_choice.json.

Do not guess options the user can answer in one round. Reuse prior answers when continuing an existing project.

Orchestration

Task type → workflow routing

Task type	Molecules to run	Notes
Single new step figure	plan-step → render-step (+ finalize-step if verified)	run analyze-project only if no project bible exists yet
Full figure set (multi-step)	analyze-project → per step: plan-step → render-step → finalize-step	full pipeline
Revise an existing step render	render-step only	reuse the verified card; re-run Gate 1 only if the card changes
Relabel an accepted render	finalize-step (fit-labels only)	no re-render
Evidence / audit package	finalize-step (build-evidence-package only)	indexes existing artifacts
Style exploration / sketch	plan-step (agent_draft, minimal) → render-step	output stays DRAFT/UNVERIFIED

Modes

• quick — exploration: minimal mapping, agent_draft, no human gates, no labels, no evidence package. Output DRAFT/UNVERIFIED.
• verified — final deliverable: full analysis, strict_verified Gate 1, labeling, Gate 3, optional evidence package.

Each molecule ends with a named artifact and a verification state; never skip a molecule's required output. On any DRAFT/UNVERIFIED output: state it explicitly; never present it as verified.

Required inputs

1. Scientific source (patent / paper / protocol / DOCX / PDF text or user description) — the authority for operation, identity, and state.
2. Optional reference image (style only).
3. Task mode and Gate 1 mode.

Required outputs

Per generated step the workflow MUST produce:

File	Molecule	Content
project_material_bible.yaml	analyze-project	Stage 0 analysis (once per project)
stepN_mapping.yaml	plan-step	mapping card + verification status
stepN_prompt.txt	render-step	final prompt verbatim
stepN_vXX.png + .json	render-step	one candidate per version + metadata
stepN_probe.yaml	render-step	Gate 2 probe results
stepN_labeled.png	finalize-step	after Gate 3 (verified mode)

Gates

Gate	Location	What it verifies
Gate 1	plan-step	mapping card ↔ open source (human, strict mode)
Gate 2	render-step	render ↔ mapping card (fixed probes)
Gate 3	finalize-step	labels ↔ shipped pixels

No other artifact in this workflow is a truth gate. Everything else is documentation, hygiene, or an executable substrate.

Required invariants

1. Source authoritative for operation, identity, state.
2. No operation or scientific entity invented for composition.
3. Repeated entities: one canonical descriptor per project; identity tokens stable unless a declared transition justifies drift.
4. Uncertain properties stay generic or OPEN; never silently resolved.
5. Style bible / reference image ≠ scientific evidence.
6. Topology comes from the verified mapping card.
7. Labels and shipped pixels agree.
8. API success ≠ visual or scientific acceptance.
9. A skipped source check produces DRAFT/UNVERIFIED, never a verified claim.

Repository layout

SKILL.md                  this playbook (compound)
composites/               molecules
capabilities/             atoms
schemas/                  JSON Schema for material bible / mapping card / render metadata / probe
adapters/                 renderer-specific contract (dashscope.md)
scripts/                  generic renderer (generate_image.py)
examples/                 worked artifacts (hydrogel_patent/)
.env.example              key template; copy to .env locally

Extending

• New single-purpose operation → add a capabilities/*.md, keep it mechanical, give it one schema or script if it touches one.
• New scoped task → add a composites/*.md that chains existing atoms; prefer reusing atoms over new ones.
• New top-level job → compose molecules in this playbook (or a new playbook).
• A capability that calls another capability is a molecule by definition; do not deepen the atom layer.