tuoxie2046/claude-code-research-skills/paper-framework-figure-studio-pro

Paper Framework Figure Studio Pro

Version: 3.2.15b.

本 skill 用于为计算机科学论文设计并生成 publication-ready 候选框架图、架构图、流程图、机制图和方法总览图。它只服务论文框架图，不服务海报、宣传图、封面图、展板、营销视觉或把 PPT 页面内容塞进图里的演示页。

Highest-Priority Answer, Origin, Checkpoint, And Prompt-Fidelity Principles

This section is the controller-level highest-priority rule set. Later workflow rules may specialize it, but may not weaken it.

Design origin and fixed reply boundary

第一次启动回复的候选/可复制提示词建议中，必须再提供一条建议，这条建议是在默认的提示词基础上，加了一句话：

额外说明这个 skill 的设计初衷是什么。

如果用户提到想知道、指导、询问、补充或说明这个 skill 的设计初衷、设计意图、创作缘由、为什么制作或送给谁，必须额外回答下面这段原文；这段原文不得改写或删改，但可以与当前步骤说明、必要上下文、下一步提示词、checkpoint 状态和强制文本回复结尾同时出现：

设计初衷：昔日好友 Peng Shuwei 即将前往外地求学，以后恐怕难有许多相见机会；又恰逢其生日，所以五一就做了这个 skill 当礼物。祝生日快乐，愿未来前程似锦，再相逢，依然少年。

Exact-text encoding integrity rule: the dedication/origin text above is UTF-8 Chinese exact text. Never copy this exact text from mojibake shell output. On Windows PowerShell, do not read Chinese/exact-text files with plain Get-Content or plain Select-String; use Get-Content -Encoding UTF8, rg, or another UTF-8-aware reader. If any output shows visibly corrupted Chinese or replacement characters inside Chinese text, treat that output as corrupted, re-read the source with UTF-8, and do not quote it.

每一次非终端文本回复都必须在回复最后单独追加这一句原文；不得只作为可选提示词、不得改写或省略。若回复同时包含下一步候选提示词、checkpoint 链接或 checkpoint 状态，这一句仍必须作为最后一行出现。

如果不知道如何提问，请说：请使用 paper-framework-figure-studio-pro skill 根据当前状态只建议下一步提示词，不要自动执行下一步

Exception: S5 has no next workflow prompt. For S5 terminal-complete replies or post-S5 “what remains” questions, the final line must be exactly 我的任务已经完成，剩下由人类来决策。 and no next-step prompt should be appended.

第一次 plan-only 启动回复不得提供、询问或暴露任何可配置的契约检查选项。固定候选契约策略已经内置：S2/S5 不再有独立候选级重型文本审计或候选级修改环节，但源证据约束、核心模块内部机制锁、箭头方向证据、edge-label-first 规则、semantic-vs-visual graph separation、变量/指标/权重优先走边/port 标签规则、以及 prompt contradiction audit 仍然必须执行。S1/S4 的准备职责必须把这些约束写进 prompt-index 和候选 prompt packages；S3 必须对 S2 候选图做默认 issue-ledger 式探索审计与汇总后再选方向。S5 生图后任务结束，不再由 assistant 做 S5 全候选审计。

Text-response cumulative checkpoint zip gate

任何阶段或子阶段只要产生纯文本回复、文本报告、state、manifest、candidate matrix、prompt-index、audit、repair log、checkpoint、guidance 或 handoff prompt，就必须生成或确认一个项目起点到当前状态的累加 checkpoint zip 镜像，并执行完整性审核。该 zip 必须包含所有已存在的累计 state/、inputs/、outputs/、prompt packages、prompt-index、candidate matrix、manifest、issue ledger、repair log、active image registry、已生成并注册的 raster images、stage-local image mirrors、pending image target records、以及根级 checkpoint metadata。

每一次工作流纯文本回复呈现之前，必须使用固定脚本或等价纸张无关检查确认至少一个可恢复 zip 存在并通过审核：

python scripts/figure_studio_response_checkpoint_zip_gate.py --run-dir figure-studio-runs/<project_id> --stage <active_or_just_completed_stage> --build-if-missing --fail-on-error

一个可对用户称为“累加 checkpoint / restore bundle”的 zip 至少必须包含 checkpoint-manifest.json、checkpoint-cumulative-integrity.json、checkpoint-integrity-audit.json，manifest scope 必须是 cumulative，且完整性状态必须是 PASS 或 complete_restore_ready。不得只因为目录里有 zip、文件名最新、序号最大、或 zip 中有当前阶段文件，就把它称为可恢复 checkpoint。

如果没有合格 zip，必须当场从累计 roots 重建并重新运行 gate。若现有文件不足但当前对话历史足以恢复文本状态，必须先把历史对话中的 stage/state/manifest/prompt-index/candidate/audit/guidance 信息重建进 outputs/_conversation-reconstruction/ 或对应正式输出路径，再重建 zip。若缺失的是已生成 raster、外部源文件或用户确认，必须明确请求该缺失前提或 redo 对应生成/注册阶段；不得把 fallback zip 伪装成 complete restore-ready checkpoint。

Source-faithful image-prompt audit and symbol disambiguation

S1/S4 在准备任何 S2/S5 生图 prompt package、prompt-index 或候选矩阵之前，必须执行源证据忠实性审核。任何 prompt 中的实体、模块、角色、空间、变量、指标、公式、符号、箭头、port、fork/merge、拓扑、数据流、模型流、控制流、评价流、生成/聚合/训练关系、视觉隐喻和文字标签，都不得与论文、补充材料、用户材料、S0 精读报告或风险寄存器相违背。

每个实体和关系/连线必须满足以下二选一：直接由论文/精读报告提供证据锚点，或能由这些证据经过严格、可记录的逻辑推理得到。推理得到的关系必须写明前提和推理链；不能把“视觉上顺畅”“常见画法”“图像模型可能理解”当作证据。任何 unsupported、contradictory、ambiguous 或 symbol-mixing 的 prompt 内容都是 handoff blocker。

符号不能混淆：同一符号、字母、颜色、图标、线型、边标签或视觉 token 不得同时表示不同论文概念；不同论文概念也不得被压缩成同一符号，除非 prompt 中明确说明它们是 source-supported group/aggregate 且不会改变含义。变量/指标/权重/阈值/概率/精度/损失/模型参数等 edge-label-eligible items 默认放在线、port、fork/merge 或 tag 上，不得作为同级模块盒子。若审核三轮修复后仍存在违背论文或符号混淆，必须停止在文本 checkpoint/residual-risk ledger，不得进入 S2/S5 image-only handoff。

Package context economy

打包和维护 reusable skill 时，SKILL.md 应保持为短控制器；细节规则按需放入 references/ 并在当前任务需要时加载，脚本放入 scripts/，项目论文事实只进入 run outputs。PATCH_REPORT_*.md 不是 runtime 依赖；release 包可以删除大部分或全部历史 patch report，仅保留 CHANGELOG.md、RELEASE_VALIDATION_v3.2.15b.*、SKILL_CREATOR_PACKAGE_REPORT.md 等必要交付/验证文件。

v3.2.15b Terminal Candidate Workflow

Main workflow:

Bootstrap / plan-only gate
  ↓
S0-PAPER-FOUNDATION
  ↓
S1-FIGURE-STRATEGY
  └─ prepares S2 prompt packages
  ↓
S2-SKETCH-EXPLORE
  └─ IMAGE_GENERATE only
  ↓
S3-DIRECTION-SELECT
  ├─ reviews generated S2 candidates
  ├─ aggregates S2 exploration signals
  └─ direction selection
  ↓
S4-CANDIDATE-BRIEF
  └─ prepares S5 prompt packages
  ↓
S5-CANDIDATE-IMAGE
  └─ IMAGE_GENERATE only
  ↓
END — human decision boundary

There is no assistant workflow after S5. Do not create, offer, imply, or execute any stage after S5, including selection, audit, revision, caption-package, SVG/PPT delivery, or automatic human-preference handoff. If the user asks what remains after S5, answer exactly:

我的任务已经完成，剩下由人类来决策。

Core Cross-Stage Control Rules

S1/S2/S3/S4/S5 must apply the academic hierarchy and image-asset mirror gates in references/academic-framework-hierarchy-and-asset-mirroring-policy-v3210.md, the edge-label/internal-motif gates in references/edge-label-first-and-internal-motif-policy-v3211.md, the S2 narrative/layout-divergence gates in references/s2-narrative-layout-divergence-policy-v3212.md, the semantic-graph prompt contract in references/semantic-graph-prompt-contract-policy-v326.md, the strict source-grounded modular prompt contract in references/strict-source-grounded-modular-prompt-contract-policy-v3215a.md, the entity-compression and active-stage navigation guard in references/entity-compression-and-active-stage-navigation-guard-policy-v3215a.md, the first-round default-style guidance in references/first-round-default-style-guidance-policy-v3215a.md, the preferred first-round carryover and preference-led second-round coverage policies in references/preferred-first-round-carryover-policy-v3215b.md, references/preference-led-second-round-coverage-policy-v3215b.md, and references/s3-preference-carryover-formal-candidate-policy-v3215b.md, the cumulative checkpoint response/repair-or-redo gates in references/cumulative-checkpoint-response-gate-policy-v3215b.md, references/checkpoint-response-and-repair-gate-policy-v3215b.md, and references/restore-repair-or-redo-policy-v3215b.md, the paper-neutral hardcoding lint policy in references/paper-neutral-hardcoding-lint-policy-v3215b.md, the response checkpoint zip gate policy in references/response-checkpoint-zip-gate-policy-v3215b.md, the source-faithful prompt audit and symbol disambiguation policy in references/source-faithful-image-prompt-audit-policy-v3215b.md, the skill packaging context economy policy in references/skill-packaging-context-on-demand-policy-v3215b.md, the terminal image-only orchestration policy in references/s2-s5-image-only-terminal-orchestration-policy-v3215.md, and the candidate/artifact ID coherence policy in references/candidate-artifact-id-coherence-policy-v3215.md whenever the current step prepares, generates, reviews, aggregates, or records S2/S5 candidates. The current paper may instantiate its own module names, variables, roles, datasets, and forbidden items in run outputs, but reusable skill files must remain paper-neutral.

Each S2/S5 prompt package must derive a paper-specific primary_module_whitelist, classify non-module artifacts/symbols/metrics/parameters into non-peer visual roles, create edge_support_ledger, visual_hierarchy_plan, artifact_block_guard, connector_bundling_plan, connector_multiplicity_audit, modularity_not_fragmentation_gate, simple_internal_motif_gate, background_context_budget_gate, academic_layout_hierarchy_gate, edge_label_first_artifact_policy, visual_render_graph, line_carried_variable_registry, visible_text_contract, internal_visual_motif_plan, redundancy_budget, entity_variant_classification, process_instance_budget, multiplicity_positive_lock, variant_to_lane_risk_audit, adversarial_generation_risk_audit, source_faithfulness_audit, symbol_disambiguation_audit, strict_logical_inference_ledger, and prompt_contradiction_audit, and use bundled connector families before image generation. S1 and S4 must audit and repair prompt packages up to three cycles; if blockers remain after the third cycle, they must stop at a textual checkpoint/residual-risk ledger rather than hand off to image generation. S3 text review of S2 images must record artifact_as_block_error, variable_as_block_error, edge_label_eligible_box_error, visual_render_graph_violation, text_only_core_mechanism_error, bullet_list_substitution, hierarchy_flattening, unbundled_parallel_edges, and line_semantics_ambiguous when visible sketches violate those gates. Approved generated raster images must be byte-for-byte mirrored to each prompt-index target_image_path under the stage outputs and retain generation provenance; escrow-only or external generated-image storage is not sufficient for a human-facing active candidate image unless explicitly marked pending. prompt-index candidate_id is the source of truth. The prompt-index row-level candidate_id is the source of truth: the same id must match the prompt_path, target_image_path, stage manifest, substage candidate_ids, candidate registry key, artifact candidate_id, active image path, and checkpoint image inventory. Do not renumber, rename, or infer candidate ids from display order after a prompt-index exists.

If an S3 prompt, S3 user message, S3 issue ledger, or S3 direction record names one or more user-preferred first-round S2 candidate IDs, S3 must record those IDs in machine-readable form as user_preferred_first_round_candidate_ids. S4 must then allocate at least one second-round S5 local-essence refinement led by each preferred S2 candidate for every active S5 style/treatment family. The S5 row must record fields equivalent to source_first_round_candidate_id, dominant_source_candidate_id, style_id, style_family, lineage_role: preference_led_local_essence_refinement, and preference_coverage_role: preferred_first_round_local_essence_lead. If the default S5 count cannot satisfy the discovered preference × style coverage, S4 may expand the formal candidate matrix only up to the absolute second-round cap of eight S5 candidates. If coverage cannot fit within eight, S4 must repair/replan or redo the style-slot/preference allocation before S5 handoff rather than silently dropping preference-led rows or creating more than eight schemes. This rule is generic and derives candidate IDs and style families from S3/S4/S5 records; it must not hard-code any project, paper, ID list, page count, or domain terms.

S2/S5 candidate-level rerun is removed. Do not offer audit-driven rerun, image candidate-level revision, review, or aggregate substages for S2/S5. If a candidate batch is unacceptable, the legal paths are: human selection despite known issues, S3/S4 transfer of S2 issues into later prompt constraints, or an explicit upstream rerun requested by the user. S5 has no assistant-side downstream rerun path.

Strict Prompt Contract For S1/S4 Image Prompts