TaoHtml

TaoHtml turns initial ideas, Word/PDF source material, existing slides, and HTML into polished 16:9 offline HTML reports and presentation-ready decks as a high-design alternative to PPT/PPTX. Use when the user explicitly invokes TaoHtml or asks an agent to structure goals, audiences, evidence, and report design; use four built-in visual systems, a confirmed reference-based project theme, or an explicit reusable corporate-template profile; implement reading or single-screen presentation behavior; and deliver portable HTML with QA and a verification handoff.

Turn source material or an incomplete idea into a finished HTML report that can be read or presented without further layout work. Optimize for a usable report first: ordinary missing details may be completed creatively, then disclosed for customer verification at delivery.

Treat TaoHtml as two cooperating layers:

• The skill directs material understanding, decisions, visual composition, generation, QA, and delivery.
• The bundled runtime supplies stable navigation, reveal state, reading/presentation modes, and other implemented behaviors.

Do not claim runtime features that are not present in references/runtime-contract.md.

Explicit Invocation

Prefer explicit invocation. Use the host agent's real skill syntax, such as $taohtml in Codex. Do not assume /TaoHtml is portable across agents.

Entry Routes

Start every new invocation with a route handshake. The current message may establish the route only when it contains a specific report topic or a source that is explicitly bound to this task. Low-information text attached only so a platform can send the Skill does not establish a route. Do not classify it with a list of forbidden words, and do not scan the workspace for a presumed input.

If the current message already provides a specific topic or explicitly binds a source, record the matching route and ask only the next genuinely missing decision. Otherwise show exactly these three entry routes and stop after this one question:

1. Idea only: build the report definition from the conversation, then produce a brief only when it is design-ready.
2. Word / PDF: extract and confirm the material before design. This is the fully specified reference route.
3. Existing PPT / HTML: confirm whether to preserve the structure or reorganize it while keeping all core viewpoints.

A short answer or ordinal selects an option only when it answers the Agent's most recent still-active option set in this conversation. The numbering above is documentation, not a global command interface. Before the route is established, do not inspect materials, produce a Material Understanding Summary, create a Report Design Brief, or write HTML. Read references/intake-workflow.md for the option-state and source-binding rules. Never ask for information already present in the conversation or an eligible bound source.

Required State Flow

Follow this route-aware sequence without skipping confirmation gates:

1. Establish the route through the startup handshake, then identify use mode and content length one decision at a time; skip every known choice. For presentation mode, treat a user-provided hard duration as a constraint, but never require duration to start. Record every used material's source_binding and binding reason; mere workspace presence never qualifies.
2. Read references/environment-preflight.md and run the smallest required profile before gated work. For a bound Word/PDF/PPT/HTML source, read references/material-understanding.md; run the pdf profile before opening or extracting a PDF, then inspect only eligible bound sources, show a Material Understanding Summary with source bindings, and wait for confirmation or correction. For an idea-only input, do not invent a source-summary gate.
3. Read references/intake-workflow.md. Resolve only the current largest missing decision that would change the report design. For an external conversion goal, distinguish the desired action from its verified real action path; do not impose this requirement on non-conversion reports. Complete ordinary missing scenes, numbers, viewpoints, and expression as creative supplements instead of extending intake. Stop according to the reference's readiness, repetition, information-gain, hard-boundary, and question-budget rules.
4. When several report structures are genuinely reasonable, present 2-3 chapter-level options and let the user choose or delegate. Do not ask about structure when one option clearly follows from the confirmed goal and evidence.
5. After content and structure are clear, resolve the visual source. Read references/profile-memory.md; parse the current enterprise identity from eligible material/conversation and resolve only explicit identity candidates against TaoHtml home. A unique active profile automatically selects its active version: run the minimal profile-reuse environment profile before binding or loading it, but do not reopen reference images, regenerate VI, launch Chromium, or ask whether to reuse. Show only “本次沿用【企业显示名 企业模板 vN】；如需更换请直接说明” and continue unless the customer objects. Ask exactly one selection question only for several candidates, unclear identity, alias conflict, or a current requirement/profile conflict; never guess or mix companies. If no profile applies or the customer requests a new/permanent template, resolve the ordinary visual route. If the user chooses “use my reference”, identify reference_mode. 参考风格重构 accepts exactly one static PNG, JPEG, or WebP screenshot; 企业模板保真 accepts one to three representative static screenshots from the same template family. When intent is still unclear, ask exactly one binary question: 参考风格重构 (extract the design language and allow recomposition) or 企业模板保真 (lock screenshot-visible corporate elements and design only each shell's safe content region). Count this question inside the existing clarification budget. If the user already says “企业模板保真”, “公司模板原样采用”, or an equivalent clear instruction, set corporate_fidelity and do not ask again; otherwise preserve the existing reconstruct behavior. Read references/static-reference-vi.md, run the static-reference profile before opening or analyzing any reference image, pass its current-session readability gate, analyze static frames only, fill the shared executable-layout grammar and mode-specific fields, render one unified VI design standards board, and wait for a clear confirmation of the exact current board. Record the current conversation reference; do not require a fixed reply phrase. Both modes use this same dependency chain and confirmation gate. Do not infer dynamic rules, force an internal theme, create a prose-only visual summary, begin project-theme generation, or start formal report production before this confirmation. PPT, webpage, video, state-sequence inputs, more than three corporate screenshots, or multiple screenshots for reconstruct remain unsupported; ask for a supported representative raster input instead of routing them as “no reference.” Only when no clear reference exists and no profile applies, read references/visual-systems.md, recommend 2-3 suitable built-in systems with name, description, and preview, then accept the user's selection or delegation.
6. On exact profile reuse, write and validate the current task's profile-use binding, resolve its relative theme path through TaoHtml home, and load that theme with the existing project-theme loader before the Report Design Brief. Treat the binding's target_mode as the current report Runtime state and pass it to the shared renderer even when it differs from the theme's first-compilation mode; a reading/presentation change alone never creates a profile version. The binding replaces only repeated VI generation; it does not confirm the brief or authorize production. After clear confirmation of the current VI board for a new reference on either static-reference mode, read references/project-theme-compiler.md, create its hash-bound handoff, and compile the current project's theme before the Report Design Brief. In corporate_fidelity, bind every source page's hash, dimensions, canvas_bbox, role, shared crops, shell-specific locked/editable regions, extension pages, and limitations; deterministically crop and embed fixed elements without model redraw, keep them stationary, and route each report page role to its matching shell with Runtime behavior confined to the editable region. Never use a complete screenshot as a page background. After the compiled theme passes the existing loader, create enterprise profile v1 only for a new company, or create and atomically activate vN only when the customer explicitly says the company default should change. A current-task-only override writes temporary_override: true and never changes the active pointer; another company always uses a separate profile. Do not add any profile result to the four built-in systems, compile an unconfirmed or changed VI, persist report content in a profile, or infer motion from static inputs. On the no-reference route, keep the selected built-in system instead. Then, after the applicable source gate passes and no hard-boundary gap remains, read references/design-brief-template.md and produce one adaptive Report Design Brief. Record the profile id/version/binding and temporary-override state when applicable; otherwise record reference_mode, source roles, fidelity boundary, locked elements, editable regions, observed/extension/unknown boundary, confirmed VI and compiled project theme when applicable, selected built-in system when applicable, any necessary deviation, and the planned creative-supplement scope. VI/profile binding is not Report Design Brief confirmation.
7. Wait for explicit confirmation of the current brief. This confirmation does not count toward the clarification-question budget. A previous "agree", "continue", or request to use TaoHtml is not production authorization.
8. Read references/production-authorization.md. Before confirmation, save each applicable current material summary, VI board or profile-use binding, and Report Design Brief as a task-local gate artifact; record its safe relative path, current SHA-256, and applicable confirmation trace. Run scripts/check_production_authorization.py --artifact-root <current-task-root> --action formal-html. Only a successful current-file check authorizes formal HTML; rerun it before browser QA and delivery so any post-confirmation byte, active-profile, version, identity, or theme change fails closed until the affected gate is rebuilt. A bound-material summary preview, static-reference VI standards-board preview, profile binding, and Report Design Brief remain allowed only in their matrix states; none is a formal report preview. After authorization, read references/process-playbook.md and produce the HTML directly. Do not insert a separate full prose manuscript step. Follow its first runnable artifact cadence: lock a concise page plan, save a complete runnable index.html, then refine and QA in bounded passes. On static-reference or profile-reuse routes, explicitly load the validated project theme through the shared renderer; never substitute one of the four built-in themes.
9. Implement only the runtime capabilities documented in references/runtime-contract.md unless the user separately authorizes new runtime engineering.
10. Run asset QA, then run the browser profile before browser QA. Fix objective failures and report the files, checks, and production-stage judgment calls. Deliver the usable report together with the concise 《待核实内容清单》 defined in references/process-playbook.md; do not withhold the report merely because ordinary creative supplements still await customer verification.

If the user adds source material or changes a core viewpoint after confirming the brief, update the brief and request confirmation again. Local copy, color, layout, or motion revisions after delivery do not require a new brief unless they change the report's core meaning or structure.

Question Discipline

Maintain the known | confirmed | inferred | missing ledger defined in references/intake-workflow.md. Re-read the conversation, sources, and prior answers before every question; never ask the user to repeat known or safely inferable information.

Ask one decision question per round: the current largest missing item whose answer could change the report design. Do not bundle independent startup choices or turn the intake into a fixed questionnaire.

A clear input may need 0 clarification questions; ordinary intake should usually finish in 3-5; even the most complex idea-only intake must never exceed 6 agent-initiated clarification questions in one intake cycle. Ask about the same key gap at most twice, changing the second attempt to a concrete example or 2-3 real options. Stop as soon as the project is design-ready, and stop questioning after three consecutive rounds without actionable new information.

Infer reversible design decisions transparently and place them in the brief for confirmation. Complete ordinary missing scenes, numbers, viewpoints, and expression during production, track them as creative supplement | projected content | illustrative content | pending verification, and disclose them at delivery. Do not describe these customer-facing categories as fabrication or hallucination.

The one-time reference-mode choice is a clarification question, not a new gate or extra budget. Ask it only when the user's intent is genuinely ambiguous; never repeat it after the mode is known.

Block only on the hard boundaries in references/intake-workflow.md: never invent a real customer or company identity, quotation, citation, literature, or source; never present an illustrative case as an achieved customer outcome; require explicit verification for legal, medical, financial, safety, and similar high-risk facts; and never replace confirmed real sources, data, or action channels. Treat "decide for me", "not important", or equivalent wording as delegation. A user-initiated change to the core goal or scope starts a new intake cycle rather than a seventh question.

Source And Meaning Protection

• Use a local file only when it is a current upload or explicit user selection, the current task instruction explicitly declares it as input, or the Agent presents the discovered candidate path and the user confirms it. A conventional filename such as input/prompt.md, directory placement, or historical task residue is not a binding.
• For every material actually used, record its path or upload identity, source_binding, and binding reason in the material summary and Report Design Brief. Do not read an unbound candidate merely to decide whether it is useful.
• You may reorganize, compress, and improve expression, but preserve every confirmed core viewpoint.
• Separate source facts from interpretation.
• Keep creative supplements distinct from source facts without treating the supplements as automatic errors.
• Resolve contradictions that affect the conclusion before the design brief is confirmed.
• Use customer screenshots, photographs, logos, and documents as real material; do not regenerate them as fictional substitutes.
• In corporate_fidelity, preserve only screenshot-visible fidelity. Do not promise recovery of the original PPT master, vector Logo, font files, or unseen assets. Never ask a model to redraw a Logo; when every supplied screenshot is insufficient for deterministic extraction, request a clearer screenshot.
• Do not rename customer-provided or independently verified facts as creative supplements. Preserve their provenance and exact confirmed meaning.
• Keep public evidence and automatic data corrections visible in the design brief's source and confirmation sections.
• Treat faithful migration versus redesign as a source-handling strategy, not as a substitute for choosing reading versus presentation mode.

Production Routing

After brief confirmation, load only what the task needs:

• references/process-playbook.md: story, evidence, visual, production, and delivery workflow.
• references/layout-pattern-library.md: layout selection for composed presentation pages.
• references/visual-systems.md: built-in system routing and selection policy; after selection, load only that system's theme.json, theme.css, and templates.html under assets/visual-systems/.
• references/static-reference-vi.md: one-time reconstruct/corporate-fidelity routing, supported static-input gate, observed/extension/unknown boundary, exact executable layout grammar, corporate template-family contract, unified VI board, confirmation gate, and next-step handoff.
• references/project-theme-compiler.md: post-confirmation hash-bound handoff, deterministic structural project-theme compilation, corporate multi-shell crop generation and role routing, eligible/compiled usage mapping, fallback policy, explicit project rendering, and verification.
• references/profile-memory.md: explicit TaoHtml home, exact enterprise resolution, immutable corporate-profile versions, automatic binding, override/update/other-company semantics, export/import, and current-task authorization boundary.
• references/design-quality-rubric.md: optional diagnosis when the user asks for a design review or says the result feels ordinary; do not use a fixed aesthetic score as the production authorization gate.
• references/runtime-contract.md: implemented DOM, controls, state, modes, and extension boundary.
• references/environment-preflight.md: capability profiles, standard-library launcher, timing, JSON contract, dependency declaration boundary, and fail-fast recovery choices.
• references/production-authorization.md: current-task gate state, allowed-action matrix, preview boundary, and machine check before formal HTML or delivery.
• assets/html-deck-template/: dependency-free 16:9 starting shell for paged reading and single-screen presentations.

Use the bundled scripts where relevant:

• scripts/extract_pdf_pages.py: render PDF pages into evidence assets.
• scripts/preflight.py: run core, pdf, static-reference, profile-reuse, or browser capability checks before the corresponding work. Its parent process uses only the Python standard library.
• scripts/profile_store.py: create, update, list, show, resolve, bind, validate, activate, rollback, archive/restore, export, and import explicit corporate-template profiles without writing to the Skill directory.
• scripts/check_production_authorization.py: validate current material-summary, VI or profile-use binding, project-theme, and design-brief state before a requested preview, formal HTML, QA, or delivery action.
• scripts/check_assets.py: find missing, remote, or non-portable assets.
• scripts/check_html_deck.py: exercise routes, reveal states, runtime behavior, media, console errors, bounds, and screenshots.
• scripts/build_contact_sheet.py: build a visual QA overview.
• scripts/package_deck.py: package HTML plus local assets when a single file is not appropriate.
• scripts/render_visual_system.py: render content through one built-in id or an explicit --project-theme directory while retaining the shared runtime shell. Pass a real local source image as --source-kind verified; verified provenance is never inferred from a local path. If source_kind / --source-kind is omitted, the Python API and CLI default to illustrative even when a local image is supplied. When no real evidence image exists, use the renderer's automatically labeled illustrative placeholder or pass a local image as --source-kind illustrative. The renderer never labels illustrative material as verified evidence.
• scripts/render_reference_vi.py: validate legacy reconstruct v1.1, single-reference v1.2, or corporate-family v1.3; verify source hashes, static frame counts, canvas crops, normalized regions, and strict 16:9 tolerance; extract exact fixed crops; render the standards-board HTML; and export a 3200×2400 PNG. It does not analyze images or compile a project theme.
• scripts/compile_project_theme.py: validate the exact confirmed-VI handoff and shared executable-layout compatibility matrix, then deterministically write project-local manifest, CSS, templates, and provenance. Corporate-family mode embeds only hashed fixed crops, emits stationary role-specific shells, and wraps report content in the matching editable region. It does not call a model or add a built-in theme.

Current Runtime Boundary

The standard template supports paged reading and single-screen presentation. It does not yet promise dual-screen presenter view, in-browser editing/export, interactive chart authoring, cross-page morphing, or crash recovery. Keep those capabilities out of user promises until they are implemented and tested.

Delivery Gate

Before delivery:

• Build a brief-to-output traceability ledger and verify every confirmed core viewpoint, correction, evidence gap, and decision boundary against a visible output location. For conversion reports, also verify that the real action path is visible, usable, aligned with the desired action, and consistent with its recorded source and verification status. Check every conjunct in compound requirements; a source screenshot may support a point but must not silently replace it in the designed narrative.
• Check all local and remote asset references.
• Exercise reading and presentation modes, step navigation, whole-page navigation, return-state preservation, and hash routes.
• Check console errors and visible bounds at the target viewport.
• Prefer a single self-contained HTML when it stays responsive; otherwise deliver index.html + assets as a zip.
• Deliver 《待核实内容清单》 after the report. Each entry must locate the page/content, name the supplement type, state the source status, and recommend confirm | modify | delete | ask TaoHtml to replace. Do not collapse the list into a generic disclaimer.
• Keep ordinary projected content in the delivery note so the presentation remains clean. Add an adjacent 示意 / 模拟 / 待核实 label inside the HTML only for simulated charts, fictional customer cases, generated evidence-like artifacts, or data likely to be mistaken for real proof.