Auto Heal External Skill

Universal-Skills

**Universal-Skills** is a collection of high-leverage "Meta-Skills" designed to upgrade AI Agents from simple chat bots to autonomous engineers.

Unlike standard prompts, these skills are **Agent-Agnostic** and **Architecture-Agnostic**. They function as plug-and-play modules that give an agent specific capabilities (e.g., "Review Code like a Sarcastic Senior Engineer", "Generate production-ready MCP servers", "Humanize AI text").

Table of Contents

• Universal-Skills
  • Table of Contents
  • Installation
    • Antigravity (Recommended)
    • Cursor
    • Claude (Projects)
  • Skill Registry
    • Core Meta-Skills
    • Development Skills
    • Content Skills
    • Office Skills
    • Verification Skills (VDD)
    • Agentic Workflows
    • Custom Commands Deployment
  • Licensing
  • Contributing
  • Documentation

Installation

First, clone this repository to your machine:

git clone https://github.com/MatrixFounder/Universal-skills

Then, choose your platform below.

Antigravity (Recommended)

**Where skills live**

Location	Scope
/.agent/skills//	Workspace-specific (Skills)
/.agent/workflows/	Workspace-specific (Workflows)
~/.gemini/antigravity/global_skills//	Global (Skills)

**Setup**

1. Copy the skills folder content into one of the locations above.
2. The agent will automatically index and load the skills.
3. See Antigravity Skills Docs.

Cursor

**Where skills live**

Skills are automatically loaded from these locations:

Location	Scope
.cursor/skills/	Project-level
.claude/skills/	Project-level (Claude compatibility)
.codex/skills/	Project-level (Codex compatibility)
~/.cursor/skills/	User-level (global)
~/.claude/skills/	User-level (global, Claude compatibility)
~/.codex/skills/	User-level (global, Codex compatibility)

**Setup**

1. Copy the skills folder content into one of the locations above.
2. Cursor will automatically recognize these as distinct capabilities.
3. See Cursor Skills Docs.

Claude Code

**Where skills live**

Location	Scope
/.claude/skills//	Workspace-specific (Project)
~/.claude/skills//	Global (Personal)

**Option A — install one skill at a time as a Claude Code plugin (recommended):**

This repo ships a Claude Code plugin marketplace at [`.claude-plugin/marketplace.json`](.claude-plugin/marketplace.json) listing each skill as an independently installable plugin. You can pick exactly the skills you need; Claude Code clones the whole repo once and exposes each plugin from there.

# 1. Add the marketplace once.
/plugin marketplace add MatrixFounder/Universal-skills

# 2. Install only the skills you want.
/plugin install docx@universal-skills
/plugin install pdf@universal-skills
/plugin install vdd-adversarial@universal-skills
# ... etc.

# 3. Browse the catalog any time.
/plugin

Available plugins (by category):

Category	Plugins
office	docx, xlsx, pptx, pdf, html, marp-slide
meta	skill-creator, skill-validator, skill-enhancer, skill-auto-improve
verification	vdd-adversarial, vdd-sarcastic
content	post-writing, text-humanizer, summarizing-meetings, transcript-fetcher
workflow	brainstorming
development	mcp-builder, hooks-creator, design-md

**Managing installed plugins:**

# Update a single plugin to the latest commit on origin.
/plugin update docx@universal-skills

# Refresh the marketplace catalog after we publish new plugins or update
# descriptions.
/plugin marketplace update universal-skills

# Uninstall when you no longer need a plugin.
/plugin uninstall docx@universal-skills

# Inspect what's installed and from which marketplace.
/plugin list

Plugins land in Claude Code's per-user cache at `~/.claude/plugins/cache/`; uninstall removes them cleanly. Install scope defaults to `--scope user`; pass `--scope project` to bind a plugin to a specific repo.

**Option B — copy the skills folder manually:**

1. Copy the skills folder content into one of the locations above.
2. See Claude Code Skills Docs.

Note for marketplace developers: both the per-plugin `plugins//` tree AND `.claude-plugin/marketplace.json` are generated by `bash scripts/sync-plugins.sh`. Re-run it after adding/removing/renaming a skill. The script:

• Reads skills/*/SKILL.md to enumerate plugins.
• Symlinks plugins//skills/ → ../../../skills/ so the source of truth stays in skills/ (no duplication, no breakage of the byte-identity replication protocol in CLAUDE.md §2).
• Pulls per-plugin description (curated tagline) and keywords from a curated table inside the script — NOT from SKILL.md description:, which is optimized for Claude's skill-routing trigger matching, not for marketplace browsing.
• Runs a drift check at the end: every plugin on disk must appear in the catalog, and every catalog entry's source must resolve. The script exits non-zero if either invariant is violated.
• Only deletes a plugins// folder when no corresponding skills// exists, so manually-curated plugins outside the skills/-tracked set survive a re-sync.

Skill Registry

Core Meta-Skills

Skill	Description	Tier
Brainstorming	Explores user intent, clarifies requirements, and designs solutions with domain-specific research and self-correction. Use before writing code.	2
Skill Enhancer	A meta-skill to audit, fix, and improve other skills. Enforces "Gold Standard" compliance (TDD, CSO, Script-First). Implements rules from Skill Execution Policy.	2
Skill Creator	Authoritative guidelines for creating NEW skills. Ensures compliant directory structure and philosophy. Implements rules from Skill Execution Policy.	2
Skill Auto-Improve	Vendor-agnostic meta-skill that automatically improves any artifact (skill, prompt, command/workflow, eval dataset) against a measurable metric. An orchestrator runs Proposer + Evaluator subagents in a propose→evaluate→keep/revert loop (immutable eval harness, git-isolated, revertible). Works across Anthropic / OpenAI / Gemini / OpenAI-compatible gateways.	2
Summarizing Meetings	Universal, model-agnostic summarization meta-skill. Summarizes meeting transcripts and documents (articles / papers / threads), auto-detecting the content class. Emits a two-level pyramid Markdown note (default) or opt-in structured note-JSON (--emit note-json) for a knowledge-base / wiki import — with known_concepts reconciliation, verbatim-quote + clean-name discipline, and an explicit no-silent-translation policy. v1.0 meeting→pyramid path is byte-for-byte back-compatible.	2
Wiki Ingest	Hybrid maintenance layer for an Obsidian-style llm-wiki (Karpathy pattern). Four modes: ingest (delegates to Summarizing Meetings, then upserts concept/entity pages additively, attaches [^src-...] footnotes, flags contradictions, updates index.md + log.md), lint (orphans / dangling links / missing pages / open contradictions), reindex (rebuild index.md from disk, preserve custom sections, merge duplicates), query (keyword-find + synthesise with citations). Filesystem-safe (path-traversal guard, Unicode-aware slugs, case-collision detection). Composes with Transcript Fetcher + Summarizing Meetings: fetch → summarize → ingest.	2
Obsidian CLI	Drives the running Obsidian desktop app via the official obsidian CLI (1.12+): link-safe rename/move (app rewrites inbound [[wikilinks]]), typed frontmatter properties, task toggles, daily-note capture, template insertion, Base queries, file-history recovery. Three-tier safety model (T1 read-only / T2 mutating / T3 banned-by-default incl. eval and template-exec surfaces), injection-resistant (note content is data, never commands). Pairs with Wiki Ingest: live-app ops here, wiki layer there, wiki-ingest reindex as the single coherence step after structural changes — see the pairing guide.	2

Development Skills

Skill	Description	Tier
MCP Builder	Comprehensive guide for building Model Context Protocol (MCP) servers in Python or TypeScript. Includes best practices for tool design.	-
Hooks Creator	Generates secure, compliant Gemini CLI hooks (Bash/Node) with automated config and VDD-verified security gates.	2
DESIGN.md Design Systems	Authors, audits and exports DESIGN.md — the open design-system format (YAML design tokens + a markdown rationale body). Three entry routes: from a written brief, from a screenshot (colours measured by pixel quantization, never estimated by eye), or from an existing codebase. Ships four starter systems that lint with 0 errors / 0 warnings (skeleton, product-saas, editorial, and a mandatory cyrillic one whose families are verified to actually carry Cyrillic), a worked end-to-end example, and known-good / known-bad fixtures. Three stdlib scripts: lint (human-readable wrapper over the pinned @google/design.md@0.4.0, with a remedy line per rule), extract-palette (Pillow or a built-in PNG decoder, so screenshots work with no install), and check-contrast (every plausible text-on-surface pair at AA 4.5 / AA-large 3.0 / AAA 7.0 plus the 3.0 non-text gate — a check the upstream linter does not provide). Carries an anti-slop reference of sixteen prohibitions, each with a detection command, because a file that lints clean is not thereby good.	2

Content Skills

Skill	Description	Tier
Post Writing	4-step editorial pipeline for LinkedIn/Telegram posts: analyze → hook → draft → review. Includes brand voice guides (RU/EN), 7 hook formulas, platform formatting, and review checklist.	2
Text Humanizer	Rewrites text to remove "AI slop" or generates prompt tailored to specific genres (Wiki, Crypto, etc.).	2
Marp Slide Creator	Creates professional Marp presentation slides with 7 pre-designed themes. Supports image layouts, custom CSS, and "make it look good" requests with automatic theme inference.	2
Transcript Fetcher	Pulls clean plain-text transcripts from video / classroom URLs. Four sources: YouTube + Vimeo + X.com/Twitter (via yt-dlp) and Skool classroom lessons (stdlib HTML scrape; cookies optional — public communities work without auth, private ones accept Netscape cookies.txt). X (native video + Broadcasts/Spaces) is captions-first and falls back automatically to ASR through a pluggable backend chain (MacWhisper → Whisper CLI → whisper.cpp → opt-in OpenAI/compatible cloud); ffmpeg is required for the HLS ASR path (fails fast with exit 7 + remediation when absent). Skool lessons delegate embedded YouTube/Vimeo to those adapters automatically. Manual→auto language fallback ladder (ru → ru-orig → en with quality-flag), rolling-caption dedup with cue-grouping + suffix-prefix splice, anchored >> speaker turns, JSON stat sidecar (with transcript_origin provenance). Opt-in --with-description emits a .description.md (YAML frontmatter + Markdown body, rendered from ProseMirror for Skool, from yt-dlp info.json for YouTube/Vimeo/X). Skill-local .env config (secrets-safe) + install_components.py component installer. Composes with Summarizing Meetings: fetch → summarize.	2

Office Skills

**Proprietary — All Rights Reserved.** Effective 2026-04-25, the four office skills (`docx` / `xlsx` / `pptx` / `pdf`) are licensed under their per-skill `LICENSE` (e.g. [skills/docx/LICENSE](skills/docx/LICENSE)). Source is available for audit and code review only; **any use, execution, copying, modification, or distribution requires prior written permission**. For licensing inquiries: . The rest of this repository remains Apache-2.0; see the [Licensing](#licensing) section below.

Independent re-implementations of the `docx` / `xlsx` / `pptx` / `pdf` capability set, built from public specifications (ECMA-376, Microsoft Open Specification Promise, W3C, PDF ISO 32000-2) and open-source libraries. Third-party components keep their original licenses; attribution is preserved in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and in the per-skill `NOTICE` files.

Skill	Description	Tier
docx	Create / edit / convert / validate .docx. Markdown ↔ DOCX (with A4 / landscape / custom-margin page geometry — --page-size A4), template fill ({{placeholders}}), accept tracked changes (LibreOffice), unpack/pack OOXML, redlining validator (--compare-to ORIGINAL.docx) catches "editor forgot Track Changes" scenarios.	2
xlsx	CSV → styled .xlsx (bold header, frozen row, auto-filter, leading-zero preservation), force formula recalculation via LibreOffice, scan for #REF!/#DIV/0! errors, attach bar/line/pie charts to a range (xlsx_add_chart.py).	2
pptx	Markdown → PPTX (built-in pptxgenjs renderer with auto-pagination, mermaid diagrams via bundled Cyrillic-capable config, accent stripes — OR --via-marp delegation to marp-slide for editorial polish), pptx2md.py (PPTX → structured Markdown: per-slide sections, GFM tables, sidecar images with sha1-dedup, speaker notes, opt-in per-image --ocr via system tesseract, plus opt-in --ocr-denoise to filter noise from decorative images), outline2pptx.js (heading-only outline → slide skeleton with TODO placeholders), pptx → PDF, slide thumbnail grids, pptx_clean.py to drop orphan slides/media after manual editing.	2
pdf	Markdown → PDF (weasyprint, with optional ```mermaid → PNG via mmdc and bundled Cyrillic-capable mermaid-config.json), PDF merge / split (by ranges, per-page, fixed chunks), AcroForm form fill / inspect / flatten (pdf_fill_form.py), TOC bookmarks preserved.	2
html (formerly html2md)	Web/HTML → Markdown web-clipper + universal agent step. Input: a live URL or a saved .html/.mhtml/.webarchive. Hybrid Python+Node reusing the docx turndown core + pdf web_clean cleaner (two-master, fork-free). html-owned conversion for doc-site SPAs (GitBook/Mintlify/Fern): ARIA-role → GFM tables, chrome/empty-heading tidy, multi-line-link collapse, arXiv/LaTeXML ltx_listing → fenced code, and data:-URI blob stripping. YAML frontmatter, image download (default on, -aware, sha1-deduped _attachments/, --max-images), dual-output (.md + .reader.md), --stdout + --json-errors. Lite (httpx+trafilatura) fetch with per-hop SSRF guard + PDF/binary rejection; opt-in Chrome (--with-chrome) for JS pages. Resilient fallback ladder (--engine lite|chrome|auto|jina|remote): auto local-first with remote last-resort, jina/remote remote-first with automatic local fallback — a Jina outage never kills a run (one typed all_engines_failed only when every tier is exhausted). Vendor-agnostic remote reader (HTML_READER_URL/_PROVIDERS → self-hosted Jina or any reader; --no-remote kill-switch; private targets never sent externally). Web search (--search "QUERY", s.jina.ai default, per-result fallback). --remote-format markdown (trust the reader's Markdown) + --target-selector. Authenticated (login-gated) Chrome: replay a human-minted session (html login URL --save-state s.json → --chrome-storage-state/--chrome-cookies-file/--chrome-user-data-dir) to read X Articles/threads, paywalled, or private pages; --chrome-scroll for lazy replies. The Chrome tier is SSRF-gated (private + off-target-public redirects refused, non-public sub-resources aborted); storage_state is server-deployable (read-only, concurrency-safe); secrets file/env-only (0600). Auth is opt-in — with none set, behaviour is unchanged. Committed regression battery.	2

The three OOXML skills (docx/xlsx/pptx) share an identical `scripts/office/` module + `_soffice.py` LibreOffice wrapper. The `docx` skill is the **MASTER** — modifications must follow the strict replication protocol in [CONTRIBUTING.md §3](docs/CONTRIBUTING.md#3-office-skills-modification-protocol-strict).

`html` (formerly `html2md`) is the repo's first **two-master** skill: its turndown core is byte-identical to docx's (`html2md_core.js` ← `docx2md.js`) and its `web_clean/` HTML-cleaning cluster is byte-identical to pdf's `html2pdf_lib/` — both fork-free and `diff -q`-gated in CI. It therefore joins the proprietary set (derived work embedding docx/pdf code). Same protocol; see [CONTRIBUTING.md §3](docs/CONTRIBUTING.md#3-office-skills-modification-protocol-strict).

**Cross-cutting safeguards** in the shared `office/` module:

• office/_encryption.py — every reader script calls assert_not_encrypted() before opening, so password-protected files AND legacy .doc/.xls/.ppt (CFB containers) fail fast with a clear remediation hint instead of a BadZipFile traceback.
• office/_macros.py — detects .docm/.xlsm/.pptm (vbaProject.bin or macroEnabled content-type); writers warn loudly when the chosen output extension would silently drop the macros at open time. Read-only support: bytes survive through unpack/pack, but Office apps respect the content-type, so renaming an output to a non-macro extension is treated as data loss.
• office/validate.py package-layout allow-list (2026-05-12) — every ZIP entry must live under the canonical OOXML prefix set ([Content_Types].xml, _rels/, word//xl//ppt/, docProps/, customXml/ per ECMA-376 §11.3.10). Non-canonical entries surface as warnings; with --strict, they fail validation (exit 1). Catches scratch-file leaks (e.g. temp files accidentally packed into the container) that Microsoft Word silently refuses to open while LibreOffice tolerates.

**Cross-skill helpers** (byte-identical at `scripts/` level across all four office skills, including pdf):

• _errors.py — --json-errors flag on every Python CLI (including argparse usage errors via a transparent parser.error patch) emits a single-line envelope on stderr: {"v":1,"error":"…","code":N,"type":"…","details":{…}}. The schema version v is always present; bumped only on breaking changes. Domain exits use code≥1; usage errors keep argparse's code=2 with type:"UsageError".
• preview.py — universal INPUT → PNG-grid renderer; takes any of .pdf/.docx/.xlsx/.pptx (incl. .docm/.xlsm/.pptm) and produces a single labelled JPEG. PDF goes straight through Poppler; OOXML routes via LibreOffice → Poppler. Tunable via --soffice-timeout (default 240s, OOXML→PDF) and --pdftoppm-timeout (default 60s, PDF→JPEG); auto-creates the output's parent directory; CLI bounds (--cols ≥1, --dpi ≥1, etc.) validated up-front so bad inputs return an InvalidArgument envelope instead of a Python traceback.

**Mermaid diagrams** are rendered to PNG by `mmdc` (`@mermaid-js/mermaid-cli` v11) inside `md2pdf.py`, `md2pptx.js`, and `marp-slide`. Office skills ship a byte-identical Cyrillic-friendly `mermaid-config.json` (Arial Unicode MS → Noto Sans → DejaVu Sans fallback chain) and pass it to mmdc by default; override with `--mermaid-config PATH` or opt out with `--no-mermaid-config`. The Chromium binary mmdc drives lives in `~/.cache/puppeteer/` (~1 GB, **shared user-wide** across every skill that uses mmdc) — it's NOT in any per-skill `node_modules/`.

**End-to-end smoke tests** (per skill `scripts/tests/test_e2e.sh` + top-level [`tests/run_all_e2e.sh`](tests/run_all_e2e.sh)) run every user-facing CLI on a real fixture and validate the output. The full suite covers all four skills with **201 assertions** (66 docx + 40 xlsx

• 48 pptx + 47 pdf) — including parameterized --json-errors envelope checks for every plumbed CLI, three rounds of VDD adversarial regression guards (false-positive macro detection, parser.error envelope routing, subprocess hygiene), structural-semantic validation (slide/sheet/layout/master chains, shared-string + style index bounds, orphan parts, ECMA-376 ID range rules) backed by 18 unit tests in office/tests/test_{pptx,xlsx}_validator.py, mermaid edge-case fixtures (cyrillic / sequence / gantt / large-mindmap / broken-with- strict), and visual regression (`tests/visual/`) of every output PDF against a committed first-page golden. A separate hypothesis fuzz suite (`tests/property/`) black-box-tests md2pdf / md2docx / csv2xlsx for crash-on-edge- input. Together they form the primary regression gate before each release; the `office-skills` GitHub Actions workflow runs all of the above on every PR.

For practical usage of all four skills — including the complete **package inventory** (what's installed globally vs per-skill, total disk footprint, cleanup recipes) — see the [Office Skills Manual](docs/Manuals/office-skills_manual.md). For the roadmap of upcoming additions and known limitations, see the [Office Skills Backlog](docs/office-skills-backlog.md).

Verification Skills (VDD)

Skill	Description	Tier
VDD Adversarial	Verification-Driven Development. Challenges assumptions, simulates failures, and rejects "happy path" thinking.	2
VDD Sarcastic	The "Sarcasmotron". Same rigor as VDD Adversarial, but with a provocative tone to force the agent/user to defend their logic.	2
Skill Validator	Automated security & compliance scanner for skills. Detects malware, obfuscation, Base64 payloads, and structural violations.	2

Agentic Workflows

Workflow	Description
Auto-Healing Skill	Automated "Doctor" loop. Runs skill-validator to find issues, then activates skill-enhancer to fix them using VDD-verified patterns.

Custom Commands Deployment

To make the **Auto-Healing Workflow** available as a native command (`/auto-heal-external-skill`), follow these steps:

Gemini CLI

Create a symlink to the custom command definition:

mkdir -p ~/.gemini/commands
ln -s $(pwd)/custom-commands/auto-heal-external-skill.toml ~/.gemini/commands/auto-heal-external-skill.toml

**Usage**: `/auto-heal-external-skill `

Claude

Create a symlink to the slash command definition:

mkdir -p ~/.claude/commands
ln -s $(pwd)/commands/auto-heal-external-skill.md ~/.claude/commands/auto-heal-external-skill.md

**Usage**: `/auto-heal-external-skill `

Licensing

This repository uses a **split licensing model**:

Scope	License	Source of truth
Repository root and all skills except the proprietary office skills	Apache-2.0	LICENSE
skills/docx/, skills/xlsx/, skills/pptx/, skills/pdf/ (effective 2026-04-25), and skills/html/ (derived work embedding docx/pdf code)	Proprietary — All Rights Reserved	skills/docx/LICENSE, skills/xlsx/LICENSE, skills/pptx/LICENSE, skills/pdf/LICENSE, skills/html/LICENSE
Third-party components used by any skill (XSD schemas, pip/npm runtime deps, external CLI tools)	Their original licenses, unchanged	THIRD_PARTY_NOTICES.md + per-skill NOTICE

**For the office skills (`docx` / `xlsx` / `pptx` / `pdf`):** the source is published for audit, security review, and personal study only. Any **use, execution, integration, copying, modification, or redistribution** — commercial or non-commercial — requires prior written permission from the copyright holder. Snapshots of these skills that were distributed under Apache-2.0 prior to 2026-04-25 remain available under their original Apache-2.0 terms in the public git history; that grant is not retroactively revoked.

**Licensing inquiries:** .

Contributing

Before opening a PR or modifying any skill, read [**docs/CONTRIBUTING.md**](docs/CONTRIBUTING.md). It covers:

• Project layout and per-skill structure conventions.
• Universal change checklist (tests, validator, examples, gitignore).
• STRICT Office-skills modification protocol — all changes to scripts/office/ or _soffice.py must be made in the docx master skill and then byte-replicated to xlsx and pptx in the same commit. diff -qr between the three is enforced.
• Adding a new skill via skill-creator.
• Commit hygiene (no node_modules/, .venv/, compiled binaries, ECMA schemas in tree).

Agent-facing rules (development environment behaviours, replication protocol enforcement) live in [CLAUDE.md](CLAUDE.md) and [GEMINI.md](GEMINI.md), which are kept in sync.

Documentation

Detailed manuals for specific components can be found in `docs/Manuals`:

• Text Humanizer Manual: Deep dive into the Humanizer's taxonomy and patterns.
• Post Writing Manual: Prompt formulas, IDE/Agent examples, and automated content pipeline architecture.
• Skill Writing Manual: Detailed guide on the philosophy of "Rich Skills".
• Skill Auto-Improve Manual: Guide to the vendor-agnostic auto-improvement meta-skill — the propose → evaluate → keep/revert loop, the six artifact types and their edit formats, the absolute-delta vs debiased pairwise decision rules, text-quality mode (rubric + best-of-N), provider/gateway setup (incl. OpenRouter), the two token-budget axes, the immutability harness, and offline testing.
• Hooks Creator Manual: Guide to generating lifecycle hooks and security blockers.
• Skill Validator Manual: Security auditing guide — CLI reference, architecture, pattern catalog, and CI/CD integration.
• Skill Evals Guide (🇷🇺 RU): Tutorial on how skill evaluations work in skill-creator — the two eval types (trigger vs behavior), LLM-judge vs deterministic script-grader (with a top-level schema of its internals), how to author quality eval sets, when evals go stale, what they affect, and token/effort budgets. Distills advanced practices (reproducibility pinning, set diversification / the "mirage" lesson, A/B isolation, multi-rep) from the wiki-verify reference stand.
• Auto-Healing Workflow: Guide to using the automated repair loop for skills.
• Summarizing Meetings Manual (🇷🇺 RU): Guide to the universal summarization meta-skill — the two axes (transcript vs document × pyramid vs note-JSON), content-class & document-mode detection, the opt-in note-JSON contract (known_concepts reconciliation, verbatim quotes, clean names, neutral title/body fields + --contract wiki alias), the explicit translation policy, templates, tag taxonomy, and Obsidian integration.
• Wiki Ingest Manual: Guide to the llm-wiki maintenance meta-skill — vault layout (_sources/_concepts/_entities), the four modes (ingest/lint/reindex/query), citation pattern ([^src-...] footnotes), contradiction handling, idempotency invariants, the 11 wiki_ops.py subcommands (incl. classify-folder and register-summary), and the fetch → summarize → ingest pipeline.
• Marp Slide Creator Manual: Guide to creating Marp presentations — theme selection, image patterns, CSS customization, and quality checklist.
• Marp CLI Manual: Installation, dependencies, core commands, configuration, and troubleshooting for Marp CLI renderer.
• Office Skills Manual: Practical reference for docx / xlsx / pptx / pdf — install, common workflows, the redlining validator (--compare-to), the LD_PRELOAD AF_UNIX shim for sandboxed deployment, the bundled mermaid config, the test suite (E2E + visual regression + hypothesis fuzz + GHA CI), and a complete package inventory (global vs per-skill vs user-cache).
• html Manual: Practical reference for the Web/HTML → Markdown web-clipper — setup, full CLI + exit codes, lite vs Chrome engines, dual-output (whole vs reader), what the converter fixes (ARIA tables, doc-site chrome, split headings, arXiv/LaTeXML listings, data: stripping, image resolution), the SSRF / image-confinement / PDF-guard security model, honest-scope limits, and the two-master replication topology.
• Office Skills Troubleshooting: Symptom → Cause → Fix recipes for the recurring failures — pango/cairo missing, soffice timeout, mmdc fail, AF_UNIX sandbox, encrypted-input rejection, visual-regression drift, ImageMagick v6 vs v7, openpyxl recalc, AcroForm vs XFA.

Project-level guides:

• CONTRIBUTING.md: How to modify skills (universal checklist) and the strict office-skills replication protocol (docx → xlsx/pptx, byte-identical).
• Office Skills Refactoring Plan: Original design rationale for the open-source office skill architecture (historical reference; do not edit, supersede via CONTRIBUTING.md).
• SKILL_EXECUTION_POLICY.md: When to use script-first vs prompt-first, Tier definitions used by skill-validator.