visual-doc-workflow-pro

name

visual-doc-workflow-pro

description

Self-contained end-to-end workflow for official or public PDFs, scanned/image PDFs, reports, releases, notices, policies, charts, screenshots, and other visual documents. Embeds PDF inspection, extraction, rendering, OCR and form operations; adaptive direct/proxy visual reading; authorized-source locking; traceable fact gates; and integrated generation of Chinese official documents, public-account articles, PPTX presentations, news releases, guides, captions, updates, and review packages. Use whenever the user wants one installed skill to read, verify, transform, or produce deliverables from visual documents without separately installing pdf, official-document, ppt-generator, or weixin-article skills. Use whenever selected, tagged, or named.

Visual Document Workflow Pro

Version

Current version: v1.0.0.

Workflow baseline: visual-doc-workflow v0.9.0, derived from authoritative visual-public-notice-reader v1.1.0.

Integrated capability sources: pdf, official-document-skill 1.0.1, ppt-generator 1.0.2, and weixin-article-writer 3.0.11.

Update this version on every future change to SKILL.md, references, scripts, workflow gates, dependency routing, output behavior, or validation rules. Record the active version in self-check/process records so users can verify which skill version controlled the task.

Use this Skill to turn official and public visual documents into a verified fact layer first, then into trustworthy answers and multi-format products. Preserve the complete visual reading and self-check workflow of the baseline. PDF handling and downstream generation are internal modules of this package; they extend the verified fact layer and never replace the upstream gates.

This Skill replaces the runtime need for separate pdf, official-document-skill, ppt-generator, and weixin-article-writer installations. Invoke the corresponding internal modules and bundled scripts directly. Do not search for or load those external Skills. Third-party runtime libraries may still be required for particular file operations; follow references/runtime-dependencies.md and use recorded fallbacks without silently installing software.

The control flow is fixed:

lock authorized sources → bind current visual sources → inspect access and document integrity → determine text-layer and model-vision capability → choose direct-vision or proxy-vision path → reconstruct and verify full-page visual semantics → assign facts to exact regions and complete spans → set admitted/excluded fact states → reset transcript speaker state by question block and register visual relationship states → run fact gates → choose a fact-safe, output-appropriate treatment → generate the standalone product with controlled paraphrase → independently audit the saved product against source evidence → serialize the exact saved product and atomic claims into a release manifest → pass the deterministic draft release check → restore output quality and recheck after repairs → inventory all task-created temporary artifacts → ask the user whether to delete or retain them → execute the recorded choice → pass the final release and delivery checks → hand off for human review.

Do not place article, public-account, official-document, or PPT generation before visual verification and fact gates.

Hard Entry Contract

Manual selection is binding. If the user selects, tags, names, or invokes visual-doc-workflow-pro, treat the skill as already triggered. Do not reclassify the task away from this skill because another PDF reader, Python script, OCR tool, writing pattern, or model shortcut appears sufficient.

Automatic scene recognition is also binding. The user does not need to say "use this skill", "按 workflow", "可信事实层", or any skill name. If the prompt and current file match a public/official document workflow, this skill must trigger automatically.

Recognize these source-document signals in file names, visible text, or prompt text:

• Chinese signals: 公告, 通知, 政策, 办法, 细则, 指南, 办事指南, 申报指南, 项目申报, 资格认定, 补贴, 报名, 竞赛, 政府, 教育局, 学校, 新闻稿, 发布稿, 新闻发布会, 报告, 白皮书, 情况通报, 图表, 海报, 截图, 扫描件, 采访实录.
• English signals: notice, announcement, policy, service guide, application rule, qualification, subsidy, registration, government, press release, press conference, report, white paper, chart, poster, screenshot, scan, transcript.

Recognize these work-request signals:

• Reading and analysis: 阅读, 读取, 解读, 总结, 提取, 梳理, 回答问题, 判断, 核对, 整理.
• Downstream output: 新闻稿, 标题, 导语, 简讯, PPT, slides, 汇报, 公文, 情况通报, 办事指南, 公众号, 微信公众号, 文章, 推文, 图注, 更新稿, 勘误, 审校包, 草稿.

This skill must take control before any file-reading or generation action when any condition is true:

• The user explicitly selects or mentions this skill.
• The user provides an official/public-document PDF and asks for reading, analysis, extraction, answering, summary, PPT/slides, briefing, report, guide, public account article, or another document-grounded artifact.
• The user uses a short natural prompt such as "阅读这个 pdf 然后做一个 PPT", "帮我把这个公告做成汇报材料", "根据这个通知写公众号文章", or "读这个办事指南给我整理流程".

Before calling OCR, Python/pdfplumber/PyPDF/PyMuPDF, web tools, presentation libraries, document tools, or bundled scripts, record the routing decision:

• visual-doc-workflow-pro triggered: yes
• active skill version: v1.0.0
• trigger reason: manual selection, explicit mention, or automatic public-document workflow match
• current file name/path if available
• requested output type
• planned reading path and selected internal production module
• integrated PDF module entry status and the available runtime path

Prohibited bypasses:

• Do not start with a generic "I will read the PDF" path outside this workflow.
• Do not use a direct Read-tool attempt, Python PDF library, OCR engine, or presentation generator outside the integrated control flow.
• Do not directly use Python/pdfplumber/PyPDF/PyMuPDF/Tesseract/EasyOCR as the primary plan before Gate 0 and integrated Gate 1 are recorded.
• Do not generate PPT, article, briefing, guide, or report before the trusted fact layer and fact gates are complete.
• Do not skip this skill because the prompt is short, the user asks directly for a PPT/article, the PDF looks readable, or the file is encrypted/scanned/large.
• If reading fails, stay inside this skill and produce a failure report rather than switching to unsupported guesses or historical artifacts.
• Do not search the user's other projects, directories, prior outputs, caches, or similarly named files to replace, unlock, complete, or corroborate an unreadable current source unless the user first authorizes each added source.

Pre-Processing Fact Package Retention

Before any PDF reading, OCR, page rendering, Python extraction, web lookup, sub-agent delegation, or downstream generation, ask the user:

是否需要保留用以核对信息的可信事实包？

Explain the choices briefly:

• 是: save and deliver the trusted fact package together with the final PPT, article, briefing, guide, or other requested product in the user-specified output directory so the user can verify key facts and evidence.
• 否: build and use the trusted fact package only as an intermediate self-check layer for the agent; do not save it to the user-specified output directory as a user-facing deliverable.

This confirmation is mandatory. Do not infer the answer from the requested product type, output directory, deadline, or prior similar tasks. If the user has not already given an explicit retention choice in the current request, stop and ask this question before processing the document.

Record the decision:

• retention question asked
• user answer
• final trusted fact package output requested: yes/no
• trusted fact package output path if yes
• internal-only fact package retained for self-check if no

Authorized Source And Integrity Contract

Before reading source content, create an authorized_source_set. Include only files, URLs, images, notes, or prior artifacts that the user uploaded or explicitly identified for the current task. Record exact paths or URLs and stable Source IDs. Skill files, tool documentation, output paths, and task-created temporary files are operational resources, not evidence sources.

Use a default-deny source boundary:

• Do not run recursive or similarity-based discovery across the workspace, home directory, other projects, historical outputs, or caches to find substitute source content.
• Do not treat an unreadable, encrypted, damaged, incomplete, or missing source as permission to expand the source set.
• Before reading any candidate outside the authorized set, show its exact path/URL, proposed evidentiary role, and risk; wait for explicit current-task authorization, then add a new Source ID.
• If unauthorized source content is accessed, record source contamination, discard every derived fact, block downstream production, and report the safest recovery action. Do not sanitize the audit history by omitting the access.
• Run scripts/check_source_scope.py or an equivalent exact-path check before Gate 4 and final handoff when local source paths are available.

Inspect document integrity separately from readability. Record PDF container pages and visible/printed page numbers as different fields. Mark leading fragments, trailing fragments, observed page-number gaps, interrupted tables, missing attachments, encryption, and unknown total extent. Never calculate a missing-page percentage without an explicit total-page source. Never invent the subject, heading, section, or likely content of a missing span. Route only complete, independently attributable spans into authoritative copy.

Load references/source-scope-and-integrity.md whenever a source is unreadable, encrypted, incomplete, missing pages, fragmentary, duplicated, similarly named, or potentially contaminated by prior-task material.

Mandatory Integrated PDF Module Entry

After the trusted fact package retention choice is recorded, enter the internal PDF module before any PDF extraction, rendering, OCR, form, merge, split, rotation, watermark, encryption, or creation work.

Required order:

1. Load references/integrated-pdf-operations.md; load references/pdf-forms.md for form tasks and references/pdf-advanced-reference.md only for advanced operations.
2. Inspect container integrity, encryption, page count, text-layer coverage, embedded images, scan status, renderability, and requested operation.
3. Choose an available internal path: bundled scripts/pdf/, Python PDF libraries, Poppler/qpdf, OCR engine, or another local runtime described in references/runtime-dependencies.md.
4. Record the selected runtime, exact operation, result, and any unavailable dependency. Do not install packages or system software without user authorization.
5. Keep every path inside this Skill's authorized-source, visual-verification, cleanup, and delivery gates.

A failed direct Read-tool attempt does not satisfy integrated PDF entry. Gate 1 passes only when the internal module, selected runtime, inspection result, and visual-reading path are recorded. Missing runtime libraries produce a recoverable dependency report; they do not permit unsupported guesses.

Adaptive Visual Reading Contract

The purpose of this Skill is to make visual documents readable and auditable regardless of whether the active model natively accepts images. Detect the usable path after integrated PDF inspection:

• direct-vision: use when the active model can actually open and understand full-page renders. Inspect pages and evidence regions directly, optionally corroborating high-risk fields with OCR and the text layer.
• proxy-vision: use when the active model cannot accept or understand images. Render pages, extract evidence-bearing images, run region-aware OCR such as EasyOCR or another available OCR engine, retain bounding boxes and reading order, reconstruct semantic panels, and cross-check OCR with the PDF text layer.

Native image understanding is an optional acceleration and corroboration path, not a prerequisite. model cannot read images must trigger proxy-vision; it must never by itself trigger a block or a request for a text-only exception.

Raw OCR text, image counts, object dimensions, coordinates, or a self-check tick alone do not pass visual verification. A complete proxy path passes only after each task-relevant evidence region has a traceable region record, OCR or equivalent reconstructed wording, representative visual facts, comparison status (text-visual match, visual addition, text-visual conflict, or uncertain), qualifiers, units, and OCR risk through references/proxy-visual-reading.md and references/visual-semantic-inspection.md. Visual-only facts must enter the trusted fact layer even when they are omitted from the downstream product.

Do not claim no additional visual information, all visuals match the text layer, or no conflict from confirmation of text-layer facts alone. These conclusions require region-by-region comparison evidence. Use partial-proxy rather than pass-proxy when the proxy record lacks coordinates, reading order, region-level OCR evidence, or visual-addition checks for a non-critical region.

Block only when all applicable routes fail to recover evidence needed for the requested task, or when an unresolved critical visual field makes the requested authoritative claim unsafe. Record partial-proxy for non-critical unresolved regions and exclude only unsupported claims rather than stopping unrelated work.

Mandatory Post-Run Cleanup Choice

After every run that creates page renders, extracted images, OCR text, caches, conversion files, logs, debug files, or temporary directories, inventory those artifacts regardless of whether they are inside or outside the user-facing delivery directory. Before deleting any of them, ask:

本次生成了以下临时文件或目录：<exact absolute paths>。是否需要删除？

This question is mandatory for the current run. Do not infer consent from the artifacts being under /tmp, from their absence in the delivery directory, from a prior run, or from a general expectation that temporary files are disposable. Wait for the user's explicit current-run answer.

• If the user chooses deletion, delete only the exact validated absolute paths, record the command result, and inspect the paths afterward.
• If the user chooses retention, leave them in place and record retained by user decision.
• If no temporary artifact was created, explicitly report no task-created temporary artifacts found; do not invent a cleanup action or user confirmation.
• Never write user confirmed, cleanup approved, or deleted successfully unless the corresponding user reply and successful command both occurred in the current task.

Activation

Trigger this skill immediately when it is manually selected, tagged, named, or invoked.

Automatically trigger this skill even without an explicit skill mention when the task contains both:

• A public or official visual source: PDF, scanned PDF, image-based PDF, release, report, press material, chart, screenshot, transcript, notice, announcement, policy, service guide, application rule, or similar official/public file.
• A request to read, self-check, answer, summarize, extract, compare, make a PPT, prepare a briefing, write a news or public-account article, draft an official document, create a guide, caption media, update a story, or produce another source-grounded deliverable.

Short prompts still trigger the skill. For example, a prompt equivalent to "read this PDF and make a PPT for tomorrow's briefing" must enter this workflow. The user does not need to mention this skill, a workflow name, a fact layer, or pdf.

Examples that must trigger this skill, without requiring the user to say the skill name:

• "阅读这个 PDF，然后做一个明天会议汇报用的 PPT。"
• "根据这个公告写一篇微信公众号文章。"
• "把这个办事指南整理成领导汇报材料。"
• "解读这个补贴申领指南，帮我做一份办事流程。"
• "读这个项目申报规则，提取关键时间、条件和材料清单。"
• "Read this government notice PDF and make a PPT for tomorrow's meeting."
• "Read this qualification announcement and write a public account article."
• "Use visual-doc-workflow-pro to process this PDF and create the requested deliverable."

Do not trigger this skill for ordinary non-public PDFs, blank templates, pure text polishing with no source document, coding/deployment work, finance analysis, or spreadsheet calculations.

Mandatory Workflow

Follow this order for every triggered task:

1. Lock the skill entry, authorized source set, and current file. Record that visual-doc-workflow-pro is controlling the task, the trigger reason, whether the trigger was automatic or manual, every authorized source path/URL and Source ID, current PDF/file name, page count, printed-page sequence when visible, text-layer status and how it was tested, scan/image status, requested output type, and whether any historical artifact is authorized. Do not declare no text layer from one extractor's empty result when another extraction path or selectable text may exist. Do not discover substitute content outside the authorized set.
2. Ask the trusted fact package retention question. Before processing the PDF, ask whether the user needs to keep a user-facing trusted fact package for verification. Do not continue until the answer is recorded, unless the user already gave an explicit yes/no retention choice in the same request.
3. Enter the integrated PDF module before PDF processing. Load references/integrated-pdf-operations.md, inspect the file, choose a bundled or locally available runtime, and record the result. Use scripts/pdf/ where applicable. For PDF forms, load references/pdf-forms.md. For advanced manipulation, load references/pdf-advanced-reference.md. No external PDF Skill is required or expected.
4. Select and complete a visual reading path. Prefer direct-vision when the model can understand images; otherwise immediately enter proxy-vision. In proxy mode, render full pages, extract evidence-bearing images, run region-aware OCR, retain coordinates, reconstruct panels and reading order, and cross-check each region against the text layer. Classify each region as text-visual match, visual addition, text-visual conflict, or uncertain; add visual-only facts to the trusted fact layer with OCR risk. Inventory meaningful regions and assign every parenthetical, date, unit, footnote, and fact to the smallest supported region. Image extraction, dimensions, raw OCR, or an unsupported statement that everything matches does not pass; the completed proxy evidence workflow can pass. Follow references/proxy-visual-reading.md and references/visual-semantic-inspection.md.
5. Build a trusted fact and integrity layer with finite admission states. Extract only authorized current-source facts with page, paragraph, timestamp, slide, chart, or image-region evidence. Preserve exact source wording in the evidence layer and keep it separate from normalized or editorial wording. Give every candidate fact exactly one state: or . A fact may be admitted only when its source is authorized, span is complete or independently cross-page supported, evidence is traceable, speaker/identity state is explicit or not applicable, and any visual relationship is explicit or not applicable. Leading fragments, trailing fragments, unresolved spans, unverified speakers, inferred identities, and unverified visual relationships are and cannot enter authoritative copy. Record observed page gaps without guessing missing contents or total extent. Include applicable names, titles, organizations, event and publication times, figures and units, quotations, eligibility, limits, required actions, process steps, materials, channels, contacts, visible rights notices, internal source conflicts, and explicitly unknown items. Keep document version separate from reporting or coverage period. Follow for the machine-checkable schema.

If the current PDF cannot be read and the user did not authorize another source, keep the authorized set unchanged, stop, and produce a failure report. Do not search for a substitute or generate a PPT, article, official document, guide, or recommendation from unread content.

Reading Strategy

Use question-driven reading rather than blindly stuffing all pages into one model context.

• First inspect text-layer availability, page count, scan quality, table/footnote density, and target questions or target product.
• For short readable PDFs, the main agent may read all pages.
• For long or scanned PDFs, build a page index first, then prioritize pages likely to contain requested facts.
• For OCR, use small page batches and record batch size, DPI, failures, and retries.
• For long PDFs, prefer candidate-page OCR and key-page review. Full OCR is allowed when needed for completeness, but record that it was full OCR and why.

Load references/reading_strategy.md for detailed reading steps and references/page_chunking_strategy.md when the PDF is long, image-heavy, or risks context overflow.

For any page containing or possibly containing visual material, load references/visual-semantic-inspection.md. When the active model lacks native image understanding, also load references/proxy-visual-reading.md. Do not pass visual review from object counts, extracted-image dimensions, or a raw OCR dump alone; do pass a complete proxy-vision evidence reconstruction when its coverage and risk gates succeed.

Key Field Review

Before final answers or downstream products, review high-risk fields:

• Identity or eligibility subjects: applicant, team member, author, representative, guardian, institution, certifying authority.
• Dates and deadlines: opening, closing, submission, review, publication, pickup, payment, appeal.
• Numbers and ratios: amounts, quotas, shares, scores, ages, periods, page counts.
• Restrictive words: only, must, shall, may, cannot, not accepted, not recognized, excluded, except, limited to, first/primary.
• Materials and originals: distinguish material categories, originals, copies, uploaded files, online-verified items, and items not stated.
• Process actors and authority: who submits, who reviews, who confirms, who can operate the system.
• Newsroom fields: speaker and title, event versus publication time, number scope and denominator, direct quotation fidelity, causal and superlative language, image identity, visible source/credit, copyright and reposting wording.
• Integrity fields: visible printed-page sequence, leading/trailing fragments, missing subject or predicate, observed gaps versus unknown total extent, and interrupted tables or attachments.
• Source-scope fields: exact authorized source set, candidate sources awaiting approval, actual content-bearing paths accessed, and contamination status.

Specific evidence beats general evidence. Footnotes, red text, captions, table notes, same-page notices, and adjacent-page continuations can override a broader summary. If evidence conflicts, create an evidence-conflict note. Prefer the more specific source only when it clearly governs the same fact; otherwise retain both forms, identify their common factual nucleus, and neutralize or escalate the conflicting field instead of declaring no conflict.

Load references/keyfield_review_strategy.md for detailed card formats and conflict handling.

Not-Stated Protection

Never fill gaps with common sense when the source does not state the information.

Use clear labels such as:

• 原文未说明
• 材料未载明
• 仅凭本文无法判断
• 当前 OCR 结果无法确认

This protection applies especially to fees, amounts, payment timing, application channels, materials, originals, processing time, contacts, appeal routes, eligibility exceptions, and required operators. Practical advice is allowed only when labeled as advice and separated from source facts.

Sub-Agent Strategy

Do not use a fixed multi-agent plan by default. Choose dynamically:

• Short material or focused question: 0 sub-agents.
• Medium complexity or multiple independent sections: 1-2 sub-agents.
• Long, high-risk, cross-section material: up to 4 sub-agents.

Never delegate final responsibility for truth, evidence conflicts, or final conclusions to a sub-agent. The main agent must merge evidence, apply fact gates, and perform final self-check.

Load references/subagent_strategy.md when deciding whether to use sub-agents.

Downstream Products

When the user asks for a downstream artifact, first build the trusted fact layer and a product input package. Do not invoke separately installed downstream Skills. The downstream artifact must not use unverified OCR text, old scripts, old PPTs, historical fact packs, external data, or similarly named local files unless each content source is explicitly authorized, registered, and labeled.

The trusted fact package is always built for agent self-check. Whether it is delivered to the user depends on the mandatory pre-processing retention choice:

• If the user chose 是, save the trusted fact package as a separate user-facing file in the specified output directory and deliver it with the final artifact.
• If the user chose 否, do not save the trusted fact package as a user-facing file in the specified output directory; retain it only for internal verification and final self-check.

The final artifact and the control layer follow a mandatory separation contract:

• Standalone product file: always create one separate file for each requested business product. It contains the usable PPT, public-account article, news release, official document, guide, or other target content only.
• Trusted fact package: create a separate user-facing file only when the user chose 是.
• Review/audit package: when required, create a separate file for the source registry, evidence map, verification queue, media-rights record, gate results, process record, and human final-review list.
• Do not create a combined file such as 新闻稿与审校包, 文章与事实包, or PPT与审核记录.
• Put rendered pages, OCR output, extracted text, and other scratch material in a temporary or intermediate directory and remove it from the delivery directory before handoff.

Before drafting news or public-account copy, load references/editorial-quality-gate.md. Factual fidelity limits claims but does not require a generic headline, source-order structure, missing subheadings, or dense report-style paragraphs.

After drafting, load references/claim-fidelity-gate.md and audit the saved product itself. Treat title fragments and section headings as claims. Keep exact source wording in the trusted fact layer. In the standalone product, preserve restrictive, legal, procedural, identity, actor, time, unit, denominator, and scope language exactly when a rewrite could mislead; allow recorded newsroom compression or stylistic paraphrase when it remains materially equivalent. Do not label a deliberate product-level paraphrase as OCR error unless the extraction or trusted fact layer also misread the source. Then load references/attribution-and-scope-gate.md to prevent measurement, quotation, responsibility, publication attribution, visual-region ownership, or source-framework categories from spilling into adjacent claims and to preserve scope-critical words in official names.

Before handoff, create the temporary machine-readable release manifest defined by references/release-manifest.md and run scripts/check_release_manifest.py against the exact saved product. A narrow rewrite such as X数据显示该指标为Y can pass from a verified same-region data_source_label only when it makes no publisher, report-release, recency, causal, or broader institutional claim. X发布该海报/报告 and X最新数据显示 require explicit publication and time/recency evidence respectively.

After the mandatory cleanup question is answered and all authorized cleanup actions complete or retention is recorded, load references/delivery-verification.md and run scripts/check_delivery_contract.py when applicable. Use --length-mode soft for requests such as 约800字, 800字左右, or a non-strict range; use hard only for explicit limits such as 不得超过, 必须在, or 严格控制在. A user-retained temporary artifact is recorded as retained by user decision, not as an agent or Skill-quality failure; do not claim the artifact was deleted. The location of a temporary file outside the delivery directory never waives the cleanup question.

Use these internal defaults:

• PPT/slides/presentation: load references/integrated-ppt-generation.md, the selected ppt-style-*.md reference, and use scripts/ppt/ or an available presentation library. Never force web enrichment in source-only mode.
• Public account article: load references/integrated-weixin-article.md, references/weixin-compliance-guide.md, and optionally references/weixin-credible-sources.md when external research is authorized.
• Official document, leadership briefing, work report, or service guide: load references/integrated-official-document.md.
• News release, caption, update, or correction: load references/newsroom_production_strategy.md and applicable editorial gates.

Fidelity is enforced by the trusted fact layer, product input package, page/region evidence mapping, release manifest, and final self-check. Internal production modules handle structure, style, and file generation.

For PPT tasks, output or retain:

• generated PPT/PPTX as the standalone product file
• PPT input package as an internal artifact or a separate control-layer file
• fact/evidence and self-check record in the separate review package
• page-by-page mapping from PPT content to trusted fact entries and PDF pages in the separate review package

Load references/downstream_production_strategy.md for downstream product rules.

For news, public-account, official communication, caption, update, correction, and newsroom-review tasks, also load references/newsroom_production_strategy.md. The newsroom module begins only after the trusted fact layer and Gate 0-4 records exist. It may structure and phrase verified facts, but it may not bypass visual re-check, repair uncertain quotations, turn editorial suggestions into facts, or approve publication.

Formatting

Respect user format constraints. If the user forbids Markdown tables, do not use them in final outputs, process records, or auxiliary Markdown unless the user explicitly allows process tables. Run scripts/check_markdown_tables.py <file.md> when a no-table constraint matters.

Load references/format_compliance.md for detailed format checks.

Failure Behavior

After reasonable attempts, failure is acceptable; hidden guessing is not. A failure report must state:

• current file and binding details
• methods attempted
• what failed and why
• unread pages or uncertain regions
• what cannot be concluded
• safe next steps for the user

Load references/failure_report.md for the report template.

Final Self-Check

Every final delivery should include or accompany a self-check record:

• Did this task trigger the skill, and why?
• Which visual-doc-workflow-pro version controlled the task?
• Was the trusted fact package retention choice asked before PDF processing?
• Was the integrated PDF module entered before direct parser/OCR tools?
• Was the current PDF/file independently read?
• Which reading path succeeded?
• Were fact gates passed?
• Were key fields reviewed?
• Were not-stated items protected?
• Which internal production module, runtime, and bundled scripts were used?
• Were the news claim ledger, quotation, headline/lead, media-rights, version, and release gates applied when relevant?
• Which visual reading mode controlled the task: direct-vision, proxy-vision, hybrid, partial-proxy, or blocked-all-paths?
• In direct mode, were full-page renders actually opened and semantically inspected?
• In proxy mode, were full pages rendered, evidence regions OCRed, coordinates retained, semantic panels reconstructed, and each region classified as text-visual match, visual addition, text-visual conflict, or uncertain?
• Did every pass-proxy claim include actual region evidence rather than a self-check assertion, and did visual-only facts enter the trusted fact layer?
• Did every atomic claim in the saved product pass the claim-fidelity gate, including title and heading fragments?
• Were exact source transcription, normalized newsroom wording, and intentional editorial paraphrase kept as separate fields?
• Were acceptable stylistic rewrites recorded as editorial choices rather than automatically labeled OCR errors or hallucinations?
• Did every candidate fact have exactly one admitted or excluded state, and did every used claim depend only on admitted facts?
• Were transcript speaker states reset at each question block, with named speakers anchored inside the current block?
• Was every visual relationship typed, and were visible_brand, data_source_label, publisher, repost, organizer, copyright, author, and source_site kept distinct?
• Did the release manifest match the exact saved-product SHA-256 and pass both draft and final deterministic checks?
• Does every attribution phrase govern only the clause supported by that actor or source?

Load references/self_check_record.md for a reusable checklist.

Reference Files

Read only the files needed for the current task:

• references/intent_routing.md - natural-language trigger and routing rules.
• references/source-scope-and-integrity.md - authorized-source whitelist, contamination handling, page-gap and fragment rules, and unreadable/encrypted-source recovery.
• references/fact_gate.md - Gate 0-4 fact gate checklist.
• references/release-manifest.md - v1.0.0 machine-readable fact admission, local attribution state, atomic-claim, and deterministic release contract.
• references/integrated-capability-router.md - single-entry routing and rule precedence for embedded PDF, official-document, public-account, PPT, and newsroom modules.
• references/integrated-pdf-operations.md - embedded PDF inspection, extraction, rendering, OCR, manipulation, creation, and validation workflow.
• references/pdf-advanced-reference.md - advanced PDF operations and runtime alternatives.
• references/pdf-forms.md - fillable and non-fillable PDF form workflow using bundled scripts.
• references/integrated-official-document.md - embedded Chinese official-document selection, drafting, anti-AI, format, and quality rules.
• references/integrated-weixin-article.md - embedded public-account article workflow, reasoning discipline, mobile structure, verification, and compliance routing.
• references/integrated-ppt-generation.md - embedded PPT planning, design, generation, rendering, and visual QA workflow.
• references/ppt-color-palettes.md and references/ppt-style-*.md - on-demand presentation style library.
• references/runtime-dependencies.md - local runtime selection, optional libraries, and recoverable dependency failures.
• references/reading_strategy.md - PDF reading and question-driven workflow.
• references/proxy-visual-reading.md - formal visual-document reading path for text-only models using rendering, region-aware OCR, layout reconstruction, and cross-checks.
• references/visual-semantic-inspection.md - mandatory full-page visual understanding and visual evidence cards.
• references/page_chunking_strategy.md - page batching and context overflow prevention.
• references/keyfield_review_strategy.md - key-field cards, strict evidence, and conflict handling.
• references/subagent_strategy.md - dynamic sub-agent selection and limits.

Publishing Boundary

Keep this skill generic. Do not hard-code answers, page numbers, policy facts, material names, test IDs, experiment results, or user-specific workspace paths. Store only transferable workflow rules, validation checks, and tool orchestration constraints.

Versioning boundary: before publishing any modified skill package, update Current version in this file and make the same version visible in process/self-check records. Do not leave the version unchanged after changing workflow rules, dependencies, gates, output behavior, references, or scripts.