report-creator（报告创建者）

kai-report-creator

Generate single-file HTML reports from source notes or .report.md IR. Keep this file as a thin router: load only the references needed for the current path, and move detailed contracts into references/, scripts, tests, or templates.

Core Principles

1. Zero Dependencies — generated reports are self-contained HTML, with CDN or bundled assets only when needed.
2. User Provides Data, AI Provides Structure — never fabricate facts or numbers; use [数据待填写] / [INSERT VALUE] when data is missing.
3. Plan Before Generate — complex reports should become .report.md IR first, then HTML.
4. Progressive Disclosure for AI — output keeps report-summary, section annotations, and component data machine-readable.
5. Thin Routing Over Prompt Growth — SKILL.md routes work and names hard gates; detailed rules live in references.
6. Contracts and Gates Beat Prompt Soup — prefer IR, guard validation, shell tests, and post-render review over adding more prose to this hot path.

Command Routing

When invoked as /report [flags] [content], parse flags first:

Default output filename: report-<YYYY-MM-DD>-<slug>.html. Slug: lowercase ASCII, non-alphanumeric to hyphens, collapse hyphens, trim, max 30 chars.

Reference Loading

Load references by route; do not read every reference by default.

Load references/spec-loading-matrix.md before --plan and --generate as a silent classifier. It covers optional archetypes: brief, research, comparison, update.

Always load references/anti-patterns.md before --generate. Load references/diagram-decision-rules.md whenever a diagram or diagram-like structure is being considered.

IR Quick Contract

.report.md has three parts:

1. YAML frontmatter between --- delimiters.
2. Markdown prose with ## / ### headings.
3. Component fences: :::tag [param=value] ... :::.

Minimal frontmatter:

---
title: Report Title
theme: corporate-blue                  # Optional. Default: corporate-blue
date: YYYY-MM-DD
lang: zh
report_class: mixed
archetype: research                    # Optional lightweight archetype hint for silent classification.
audience: "Busy decision-maker"
decision_goal: "Decide next move"
must_include:
  - Source truth that must survive compression
must_avoid:
  - Decorative placeholder chart
charts: cdn
toc: true
animations: true
abstract: "One-sentence summary"
poster_title: "Optional stronger poster headline"
poster_subtitle: "Optional poster subtitle"
poster_note: "Optional short closing sentence"
template: ./my-template.html
theme_overrides:
  primary_color: "#E63946"
custom_blocks:
  my-tag: |
    <div class="my-class">{{content}}</div>
---

For trivial reports, omit optional fields. For high-stakes or complex reports, keep report_class, audience, decision_goal, must_include, and must_avoid so review/evals can detect drift.

Poster summary mode is opt-in. Do not infer poster_title or poster_subtitle from punctuation in title.

IR validity terms: invalid_syntax, invalid_semantics, contract_conflict, auto_downgrade_target.

Canonical component routing lives in references/rendering-rules.md; component details live in references/rendering/*.md. Compatibility anchors that must remain discoverable here:

• :::kpi canonical body uses items:.
• Timeline Allowed Date tokens: YYYY-MM-DD, YYYY-MM, YYYY, Q[1-4] YYYY, Day N, Week N, Month N.
• Use ECharts for ALL charts.
• Badges are optional visual enhancements, not a first-class IR tag.

Language, Theme, And Class

Auto-detect lang unless frontmatter sets it: use zh when CJK is material or appears in the title/topic; otherwise use en. Apply language to placeholders, TOC labels, date display, and shell labels.

If no theme is provided, pick by intent, first match wins:

Classify content by numeric density: narrative < 5%, mixed 5-20%, data > 20%; short topics default to mixed.

--plan Flow

1. Detect language and suggest theme.
2. Classify report_class; optionally add archetype only when the report clearly matches brief, research, comparison, or update.
3. For regular-lumen or periodic keywords, load references/regular-report-content-rules.md.
4. Generate .report.md with complete frontmatter, 3-5 useful sections, source-faithful structure, and placeholders only where data is missing.
5. Use real visual anchors only. Never use placeholder-only KPI/chart blocks; prefer callout, timeline, diagram, table, or prose scan anchors.
6. KPI values must be short, real quantitative values. If the source has no actual metric, do not render a KPI card; explanations and status words belong in prose, badges, callouts, or tables.
7. Use theme_overrides only for a small content-tone color hint; do not create a new design system in the IR.
8. Save as report-<slug>.report.md.
9. Tell the user the IR path, placeholder fields, suggested theme, and render command.
10. Stop. Do not generate HTML in --plan mode.

Narrative rhythm reminders: lead-block, section-quote, and action-grid are optional prose upgrades. claim -> explanation -> scan anchor is a cadence, not a quota. These are optional prose upgrades, not default required blocks. If uncertain, keep normal paragraphs and add one clearer scan anchor instead of forcing a cadence block. Do not add more than one of lead-block / section-quote / action-grid by default inside the same section unless the source material clearly warrants it.

--generate Flow

1. Read IR input only. With no file given, extract exactly one valid IR block from context. Treat this as IR from context, not chat history. If zero or multiple are present, stop and ask for an explicit file or single IR block. Never render the surrounding conversation.
2. Load reference files minimally but reliably; load only the references that materially help the current render path. Standard HTML shell generation always loads the shell entry plus all references/html-shell/*.md; component rules load by IR inventory via references/rendering-rules.md.
3. Parse frontmatter and resolve lang, theme, report_class: mixed default, archetype, date display, chart mode, TOC, animation, template, and theme overrides.
4. Run guard validation before rendering:

- Use scripts/guard_validate.py with IR text from file or extracted context. - If fatal metadata is missing, stop and report the error. - If a block is invalid, apply its auto_downgrade_target (kpi -> callout, chart -> table, timeline -> list, diagram -> callout) and mention the downgrade.

1. Render components using references/rendering-rules.md, references/design-quality.md, and the path-specific references/rendering/*.md files selected from the IR.
2. Build the standard shell from references/html-shell-template.md plus all references/html-shell/*.md; follow Shell metadata, version/theme metadata, export completeness, and the duplicate-date guard.
3. Compute and embed <meta name="ir-hash" content="sha256:[ir-hash]"> from the exact IR text, not the file path.
4. Assemble CSS through references/theme-css.md: theme before-marker, shared CSS, theme post-shared override, TOC/shell CSS, frontmatter overrides.
5. Run pre-write validation and fix all violations:

- no raw ::: in HTML - valid ir-hash - no generic/template h2 headings - short, real quantitative .kpi-value and report-summary KPI values; no placeholder or status-only KPI values - .number body numerals use tabular lining numerals - badges clarify status/category/entity, never quota-fill - timeline dates are real time markers - no U+FE0F - no text-align: justify, black-background flood, body letter-spacing > 0.05em, or mobile-hidden critical controls

1. Run the final HTML quality gate with scripts/html_quality_gate.py on the rendered HTML. It must pass standard shell IDs, theme fidelity, and KPI value checks. If it fails, fix the HTML and rerun it before reporting success.
2. Run L2 shell checks. Required: data-template="kai-report-creator", data-version, data-theme, id="toc-toggle-btn", id="toc-sidebar", id="card-mode-btn", id="sc-overlay", id="export-btn", id="export-menu", id="export-print", id="export-png-desktop", id="export-png-mobile", id="export-im-share", id="report-summary", plus the JS bindings for print/desktop/mobile/IM export. If any export item or binding is missing, rebuild the whole export block from references/html-shell/export.md.
3. Run the silent final review pass from references/review-checklist.md, then write the HTML and report the path.

When the report is explicitly comparing named vendors, models, or tools, set data-report-mode="comparison" on the outer report container and use .badge--entity-a/.badge--entity-b/.badge--entity-c only for entity identity.

--review Flow

When the user runs /report --review [file]:

1. Read the HTML file.
2. Load references/review-checklist.md.
3. Apply hard rules automatically.
4. Apply AI-advised rules only when confidence is high and factual accuracy is preserved.
5. This is one-pass automatic refinement; no confirmation window.
6. Use references/review-report-template.md when the user wants a structured summary.
7. Write back unless the user asked for diagnosis only.
8. Tell the user what changed and what was intentionally left untouched.

--themes And --export-image

For --themes, read the theme preview template and write report-themes-preview.html verbatim.

For --export-image, after HTML generation run:

python <skill-dir>/scripts/export-image.py <output.html> --mode <mode>

If Playwright is unavailable, print install instructions and skip image export without failing HTML generation.

Shell And Template Boundary

Generate complete self-contained HTML. The shell entry contract lives in references/html-shell-template.md; full shell structure, inline JS, export behavior, summary card, edit mode, TOC, print rules, and footer/watermark degradation rules live in references/html-shell/*.md.

All scripts are inline in the shell template. Never load nonexistent files such as templates/scripts/*.js.

For custom templates and TOC slug rules, use references/toc-and-template.md.

Final Output

Always end with the file path and a one-sentence summary. If a validation, guard, or export step could not run, say exactly which step was skipped and why.