figforge-plan - SKILL.md Agent Skill

name: figforge-plan description: > figforge-plan: use when the user wants to turn a source artifact into a publication-quality scientific figure plan, optimized image-generation prompt, or optional editable HTML/SVG schematic. Trigger for "make a main figure", "generate a prompt for this paper/repo", "visualize this system", "draw the architecture", "turn this into a diagram/poster", "map this repo", "paper figure prompt", "论文主图 prompt", "绘图 prompt", or any request converting a paper, GitHub repo, local codebase, diagram, algorithm, or system description into a scientific figure prompt package. Do not trigger for ordinary code review, debugging, or text-only architecture discussion unless the user asks for a visual/prompt output.

figforge-plan — Source → Scientific Figure

Turn technical material (code, paper, diagram, algorithm, or design request) into a truthful, publication-quality figure spec. Default output: an image-generation prompt package. Optional output: a self-contained HTML/SVG artifact when the user asks for deterministic editable vector output.

STEP 1 — CLASSIFY INPUT

Score every strong signal before routing. If two or more types score, route HYBRID; do not collapse mixed input to the first signal.

Type	Strong signals	Sub-skill
CODE_REPO	repo path, manifests (`package.json`, `pyproject.toml`, `Cargo.toml`...), file tree, source syntax (`import`, `class`, `def`)	`skills/repo_analyzer.md`
RESEARCH_PAPER	PDF, abstract/method/results sections, "Figure 1", DOI/arXiv, "we propose / we evaluate"	`skills/paper_to_poster.md`
DIAGRAM_IMAGE	image/SVG/screenshot with boxes, arrows, nodes, swimlanes	`skills/diagram_to_draft.md`
ALGO_TEXT	numbered steps, pseudocode, stage-by-stage prose, no source-code syntax	`skills/algo_to_draft.md`
HYBRID	≥ 2 strong types above	`skills/hybrid.md`
DESIGN_REQUEST	new design ask, no reference material	`skills/design_from_scratch.md`

If routing is ambiguous after scoring, see router/intent_parser.md for the full signal-scoring rules and ambiguous-case table. If the user does not want a visual or prompt artifact, do not use this skill.

STEP 2 — TRUTHFULNESS CONTRACT

Every output distinguishes evidence from inference. Read style/evidence_discipline.md before compiling. It is the single source for the evidence ledger format, metric discipline, and the rationalization table.

Quick self-check — STOP if you catch yourself thinking:

"This number probably exists in the source."
"Marking unknown looks ugly; let me put a placeholder."
"User wants polish; rough estimates are fine."
"This dependency is so common it must be present."
"The diagram needs symmetry, so I'll add one more module."

Each is a Contract violation. Mark unknown, omit the panel, or label as assumption. Full rationalization table in style/evidence_discipline.md.

STEP 3 — SELECT OUTPUT TARGET

IF user asks for a prompt, image model, Midjourney, GPT image, Stable Diffusion,
   "生图", "绘图 prompt"
  → image_prompt   (default)

ELSE IF user asks for editable HTML/SVG, deterministic labels, or a local artifact
  → html_artifact

ELSE
  → image_prompt

If the user wants both the prompt and a directly generated image, see router/image_handoff.md for the two-stage handoff and credential boundary rules.

STEP 4 — EXECUTE

Read style/visual_system.md (palette and diagram language) and style/evidence_discipline.md (evidence rules). If the source has a strong domain signal (paper venue, repo dependencies, diagram notation, algorithm field), also read style/domain_hints.md for matching Favor primitives and Avoid items.
Read the sub-skill from STEP 1.
Run its phases; produce the structured content block.
Read the chosen renderer:
- renderers/image_prompt.md (default)
- renderers/html_artifact.md (deterministic vector output)
Compile the final output. The renderer is always the last step; never compile before analysis is complete.

State briefly to the user: "Routing as [TYPE] → compiling [output_target]." Ask at most one clarifying question, only if the missing detail would change the route or output type.

For image_prompt, return the full prompt package in chat. For html_artifact, save the file in the current workspace if local filesystem access exists, unless the user requested a different destination.

Environment Notes

Chat: use uploaded PDFs, images, text, or zip contents directly. Do not claim local filesystem access unless the tool environment provides it.

Local coding: use fast file tools (rg --files, rg, package manifests) to scope the repo. Skip node_modules, .git, dist, build output, generated files.