---

name: research-paper-figure-skill-factory license: MIT-0 description: "Use when the user wants a research-paper figure Skill Factory: build, patch, package, or use reusable specialized paper-figure-making skills from lawful literature/corpus evidence. Generated skills must use a specialized-skill-first workflow, full-feasible local PDF coverage where available, startup-plan-only first replies, target-paper candidate/final image isolation, mandatory image-embedded visual-structure explanation, mandatory non-target concept/modeling example display for abstract visual decisions, saved subtype/style illustration atlases, ChatGPT web Create image / ChatGPT Images 2.0 rendering, Codex $imagegen-first rendering, sample-image transfer rules, all-step/current-position state footers, and a mandatory first-round diverse candidate board followed by P6b/P6c paper-local best-practice optimization before final prompt construction." metadata: display_name: Research Paper Figure Skill Factory version: "2.0.5" author: OpenAI tags: research-figure, paper-figure, figure-skill-builder, figure-skill-factory, scientific-illustration, figure-taxonomy, meta-skill, image-prompt, imagegen, chatgpt-images-2, clawhub, openclaw, visual-style, figure-studio compatibility: ChatGPT web, Codex, OpenClaw, ClawHub marketplace. Requires image-generation capability for rendering. openclaw: skillKey: research-paper-figure-skill-factory

Research Paper Figure Skill Factory

This skill is a two-layer research-paper figure Skill Factory.

1. Skill Builder layer: build or patch a reusable specialized figure-making skill for one paper-figure class by acquiring lawful source material, extracting figure evidence, building a taxonomy, generating the skill package, testing it, and locking it.
2. Figure Production layer: after a specialized skill is locked, use that generated skill to design, compare, render, review, and integrate concrete figures for arbitrary target papers of the same figure class.

Version 2.0.5 adds a stricter visual-structure-as-image gate for generated specialized skills. When a generated skill explains or defines visual structure, layout skeleton, panel choreography, module topology, arrow grammar, candidate-board structure, second-round optimization geometry, or final content architecture in a text turn, it must show that structure with an embedded saved reference image or non-target concept/modeling example image. It must not substitute a prose-only or bullet-only visual-structure description. The existing hard gates remain: abstract visual decisions require inline reference/concept images, and after P6 selects the strongest first-round direction, P6b/P6b-IMAGE/P6c must run a paper-local best-practice optimization round before P7 final prompt construction. Target-paper candidate images, draft figures, final figures, and revisions still remain isolated in dedicated IMAGE_ONLY turns.

Non-Negotiable Contract

First Trigger

On first trigger, output only a startup plan. Do not analyze a paper, build a taxonomy, create candidate schemes, draft prompts, or generate images. The first reply is STARTUP_PLAN_ONLY (TEXT_ONLY).

If the first user message asks for images, record the request as pending only. The first reply must not call Create image, $imagegen, an image API, or include image artifacts.

Specialized-Skill-First Builder Rule

The normal route is:

figure-class goal -> corpus plan -> lawful acquisition/local corpus -> evidence extraction -> taxonomy -> specialized skill blueprint -> generated specialized skill -> tests/patches -> locked skill -> target-paper production.

Do not jump from source papers directly to one concrete figure unless the user explicitly chooses a full production fast-track. If fast-tracking, record the skipped builder steps and fallback skill/taxonomy.

Full-Feasible Corpus Rule

When local PDFs, a paper index, or retrieval manifests exist, enumerate the full relevant candidate set and process as many accessible relevant PDFs as feasible. A small sample can support only a limited/pilot/fallback lock unless the user explicitly accepts that limitation. Representative rendered pages are audit aids only, not the corpus size.

Mandatory Candidate-Image Bridge

Every generated specialized figure-making skill must include a hard workflow bridge after any multi-option text decision:

1. TEXT_ONLY candidate text turn: present 4-6 text candidates, normally 6.
2. TEXT_ONLY visual candidate setup turn: define candidate count, varied axis, fixed elements, rendering route, and what the user should compare.
3. IMAGE_ONLY candidate-board turn: generate/display 4-6 candidate images or schematic candidates, normally 6.
4. TEXT_ONLY candidate-review turn: record the previous image batch, compare candidates, recommend one direction, and ask the user to select, revise, or request another board.

This bridge is mandatory after candidate schemes, subtype choices, layout choices, style choices, metaphor choices, density choices, and prompt alternatives. The generated skill must not move directly from 4-6 text candidates to final prompt construction, final image generation, caption writing, or text-only locking unless the user explicitly says to skip image candidates and stay text-only. If skipped, record visual_candidate_board_skipped_by_user: true.

Generated skill lock/test must fail if:

• the workflow lacks a dedicated visual candidate setup step;
• the workflow lacks a dedicated IMAGE_ONLY candidate-board step before direction lock;
• examples show text candidates followed directly by final prompt or final image generation;
• the state footer cannot record visual_candidate_board_status, candidate_image_batch_id, and selected_visual_candidate;
• multi-option next prompts do not ask the user to generate/display multiple candidate images or schematic candidates, normally 6.

Target-Paper Image Isolation And Required Inline Reference Display

Every response must distinguish target-paper figure production from explanatory reference display:

• TEXT_ONLY: planning, intake, diagnosis, candidate text, candidate-board setup, prompt writing, critique, status, next prompts, and inline display of allowed reference images.
• IMAGE_ONLY: target-paper candidate-board generation, draft/formal figure generation, final figure generation, and target-paper revision image generation only. No prose, captions, critique, prompt text, or state footer.

Allowed reference images inside a TEXT_ONLY reply:

• already-saved package-local subtype/style atlas or reference images;
• non-target concept diagrams used to explain an abstract visual grammar, workflow, taxonomy axis, or modeling pattern;
• non-target example images created while building or demonstrating the specialized skill itself, as long as they do not represent the user's target paper, are not offered as selectable candidates, and are not treated as draft/final paper figures.

Required abstract-decision trigger:

• If a TEXT_ONLY step explains or compares figure subtype, layout grammar, visual style, density, metaphor, modeling pattern, candidate scheme differences, or final content architecture, the generated skill must display at least one relevant saved atlas/reference image or non-target concept_example / non_target_reference image with Markdown image syntax.
• This applies especially to P2, P3, P4, P6b, and P7; it also applies to P1, P6, or P9 when those steps contain abstract visual comparison or final content-architecture reasoning.
• Pure state synchronization, caption/body text drafting, simple confirmation, and ordinary restatement of an already registered image batch do not require a new concept/example image.
• If no suitable saved reference exists and live inline generation is unavailable, record concept_example_required: true, concept_example_status: generation_pending or missing_recorded, concept_example_role, and concept_example_trigger_reason, then make the repair action explicit.

Required visual-structure-as-image trigger:

• If a TEXT_ONLY step explains, compares, or defines visual structure, layout skeleton, panel choreography, module topology, arrow grammar, content architecture, candidate-board structure, second-round optimization geometry, or final image-brief structure, the generated skill must display a structure image with Markdown image syntax in that same text reply when technically possible.
• The displayed image must be an already-saved atlas/reference image or a non-target concept_example / non_target_reference generated for explanation. It may use generic placeholders, but it must not contain the user's target-paper-specific modules, claims, data, or final labels unless the user explicitly supplied them as detached generic examples.
• Do not use prose, tables, bullets, ASCII diagrams, Mermaid, SVG, or code-rendered sketches as the only representation of a visual structure. Text may name the structure role, fixed elements, and varied axes, but the structural form itself must be shown as an embedded image.
• If the only useful structure preview would be paper-specific, defer that preview to the next target-paper IMAGE_ONLY step (P5, P6b-IMAGE, or P8) and embed only a generic structure/reference image in the text reply. Do not embed paper-specific candidate, second-round, formal, final, or revision images in prose.
• State must record visual_structure_image_required, visual_structure_image_status, visual_structure_image_role, and visual_structure_image_trigger_reason in addition to the concept_example_* fields. Missing status is temporary only; production lock fails until an available saved reference or generated non-target structure image is embedded or a host-rendering block is explicitly recorded with repair action.

Target-paper images must not be embedded in text replies. If an image is meant for choosing or locking a visual direction for the target paper, refining a paper-specific figure, or producing a formal/final paper figure, the generated skill must use the P5/P8-style IMAGE_ONLY boundary. A concept/example image embedded in text must be labeled in state as non_target_reference, must not set or reuse candidate_image_batch_id, and must not be used as evidence that the candidate-image bridge has been satisfied.

If the host cannot generate and embed a non-target concept/example image in the same text response, generate/save it first and embed it in a later TEXT_ONLY reply. Do not relax the IMAGE_ONLY boundary for target-paper candidate or final outputs.

Mandatory Best-Practice Divergence After P6

The first target-paper candidate-board round should be deliberately diverse. P4/P5 should vary high-level direction-setting axes such as subtype, layout grammar, metaphor, density, panel rhythm, or style family so the user can choose a promising direction.

P6 records the first-round candidate_image_batch_id, compares candidates, and selects the strongest current direction. P6 is not allowed to jump directly to P7. After P6, generated skills must run a paper-local best-practice optimization round:

1. P6b TEXT_ONLY: propose 4-6 optimization axes, normally 6, based on best practices and the selected paper-local details: local module relationships, evidence/case anchors, label economy, panel transitions, color semantics, callout placement, and reviewer-facing readability. State exactly which elements stay fixed from the first-round winner and which local details vary.
2. P6b-IMAGE IMAGE_ONLY: generate/display 4-6 second-round target-paper variant images, normally 6. This uses a new second_round_candidate_batch_id; it must not reuse candidate_image_batch_id or any concept/example image id.
3. P6c TEXT_ONLY: record the second-round batch, compare variants, select or combine the final direction, and only then allow P7 final image brief construction.

Generated skill l