Documents Orchestrator
The single entry skill for document work. It locates the task on the format × intent map and
delegates to one of 8 specialist spokes. The cross-cutting decision every document task shares —
whether you need the editable source (round-trippable, structure preserved) or a rendered
artifact (pixel-fidelity output you cannot edit back) — lives in documents-core; read it
before you pick a tool, because the choice locks which spoke can touch the file next.
Cluster map (routing by intent → spoke + role)
Office documents (create / edit / extract / convert)
- Word, PDF, PowerPoint, Excel — the full editable engine →
documents (bundles docx/, pdf/, pptx/, xlsx/ subskills + a Playwright HTML→PDF report pipeline)
- Quick natural-language PDF tweaks ("rotate page 3", "redact this") →
nano-pdf
Extract & ingest (document → data)
- Scanned/image PDF → text + structured data with confidence flags →
deepread-ocr
- Turn a corpus into a notebook, briefing, or podcast →
notebooklm
Publish & narrate the docs themselves
- Build/maintain a documentation site →
mintlify
- Generate user-facing release notes from git history →
changelog-generator
Diagrams & visual documentation
- Professional draw.io diagrams (flowcharts, BPMN, WBS/Gantt/RACI, network, UML) →
drawio-diagrams-enhanced
- Animate a Mermaid diagram into a GIF →
mermaid-to-gif
Picked-up spokes
Additional documents spokes folded in from the antigravity-awesome-skills library (MIT). Route to these the same way — load on demand by reading their SKILL.md.
Author project docs
- Write an absurdly thorough
README.md after deep codebase exploration (tech stack, getting-started, architecture, env vars, scripts, testing, deployment, troubleshooting) → readme
Rendered marketing artifacts
- Capture true-retina (2x) marketing screenshots of an app with Playwright — discover routes/features from the codebase, handle login, then shoot each view (Product Hunt / social / landing / docs) →
screenshots (one-way render gate: produces PNGs you cannot edit back into the app)
Standard Operating Flow
- Classify the task: which format (DOCX/PDF/PPTX/XLSX/diagram/site) and which intent (create → edit → extract → convert → publish).
- Decide editable vs rendered (the model in
documents-core). If the user needs to keep editing, stay in the native format and refuse a one-way render until they confirm.
- Delegate to the spoke(s). Multi-step asks fan out in pipeline order (e.g. "scan this contract and summarize" →
deepread-ocr → notebooklm; "report from the assessment data" → documents ConsultingReport → Playwright PDF).
- Return: chosen spoke(s), whether output is editable or rendered, the format, and the next action.
Guardrails
See documents-core. In short: preserve the source — prefer the editable, round-trippable
path; treat any render-to-PDF/OCR/GIF as a one-way gate and say so before crossing it. Never
hardcode a value a formula should compute (XLSX), never silently flatten tracked changes (DOCX),
and never trust OCR text without surfacing its confidence (deepread). Keep the original file
intact; write derived artifacts alongside it, not over it.
Loading spokes on demand
To keep CLI startup context lean, this cluster's spokes are not separately registered as skills — only this orchestrator and its *-core are enumerated. When you route to a spoke named above, load it on demand by reading its file:
~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md (or skills/<spoke-name>/SKILL.md inside the skill-clusters repo).
1---2name: documents-orchestrator3description: Route a document task to the right skill among the documents specialists — office formats (DOCX/PDF/PPTX/XLSX), docs sites, changelogs, OCR extraction, natural-language PDF edits, NotebookLM, and diagram authoring. USE WHEN a user wants to create, edit, extract from, convert, or publish a document but hasn't named the specific tool or format.4---56# Documents Orchestrator78The single entry skill for document work. It locates the task on the **format × intent** map and9delegates to one of 8 specialist spokes. The cross-cutting decision every document task shares —10whether you need the **editable source** (round-trippable, structure preserved) or a **rendered11artifact** (pixel-fidelity output you cannot edit back) — lives in `documents-core`; read it12before you pick a tool, because the choice locks which spoke can touch the file next.1314## Cluster map (routing by intent → spoke + role)1516**Office documents (create / edit / extract / convert)**17- Word, PDF, PowerPoint, Excel — the full editable engine → `documents` *(bundles `docx/`, `pdf/`, `pptx/`, `xlsx/` subskills + a Playwright HTML→PDF report pipeline)*18- Quick natural-language PDF tweaks ("rotate page 3", "redact this") → `nano-pdf`1920**Extract & ingest (document → data)**21- Scanned/image PDF → text + structured data with confidence flags → `deepread-ocr`22- Turn a corpus into a notebook, briefing, or podcast → `notebooklm`2324**Publish & narrate the docs themselves**25- Build/maintain a documentation site → `mintlify`26- Generate user-facing release notes from git history → `changelog-generator`2728**Diagrams & visual documentation**29- Professional draw.io diagrams (flowcharts, BPMN, WBS/Gantt/RACI, network, UML) → `drawio-diagrams-enhanced`30- Animate a Mermaid diagram into a GIF → `mermaid-to-gif`3132## Picked-up spokes3334Additional documents spokes folded in from the antigravity-awesome-skills library (MIT). Route to these the same way — load on demand by reading their `SKILL.md`.3536**Author project docs**37- Write an absurdly thorough `README.md` after deep codebase exploration (tech stack, getting-started, architecture, env vars, scripts, testing, deployment, troubleshooting) → `readme`3839**Rendered marketing artifacts**40- Capture true-retina (2x) marketing screenshots of an app with Playwright — discover routes/features from the codebase, handle login, then shoot each view (Product Hunt / social / landing / docs) → `screenshots` *(one-way render gate: produces PNGs you cannot edit back into the app)*4142## Standard Operating Flow43441. Classify the task: which **format** (DOCX/PDF/PPTX/XLSX/diagram/site) and which **intent** (create → edit → extract → convert → publish).452. Decide **editable vs rendered** (the model in `documents-core`). If the user needs to keep editing, stay in the native format and refuse a one-way render until they confirm.463. Delegate to the spoke(s). Multi-step asks fan out in pipeline order (e.g. "scan this contract and summarize" → `deepread-ocr` → `notebooklm`; "report from the assessment data" → `documents` ConsultingReport → Playwright PDF).474. Return: chosen spoke(s), whether output is editable or rendered, the format, and the next action.4849## Guardrails5051See `documents-core`. In short: **preserve the source** — prefer the editable, round-trippable52path; treat any render-to-PDF/OCR/GIF as a one-way gate and say so before crossing it. Never53hardcode a value a formula should compute (XLSX), never silently flatten tracked changes (DOCX),54and never trust OCR text without surfacing its confidence (deepread). Keep the original file55intact; write derived artifacts alongside it, not over it.5657## Loading spokes on demand5859To keep CLI startup context lean, this cluster's spokes are **not** separately registered as skills — only this orchestrator and its `*-core` are enumerated. When you route to a spoke named above, **load it on demand** by reading its file:6061`~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md` (or `skills/<spoke-name>/SKILL.md` inside the skill-clusters repo).
Sheshiyer