AgentKit
DocsKitsCLI ReferenceDesktop App

Skills

Create and edit office artifacts with ak:document-skills

Route PDF, spreadsheet, Word, and PowerPoint work through the bundled format guidance with explicit intermediate files and visual or formula verification.

Use ak:document-skills to read, create, edit, or analyze PDF, XLSX, DOCX, and PPTX artifacts. The parent Skill routes the task into four bundled format subskills and their scripts, references, schemas, and validation workflows.

Choose ak:document-skills for file artifacts

Use ak:document-skills when

  • You need a PDF generated, merged, split, extracted, rendered, or filled.
  • You need an XLSX workbook or CSV/TSV analysis with formulas, formatting, charts, or recalculated values.
  • You need a DOCX document created, analyzed, edited, commented, or redlined while preserving reviewable change history.
  • You need a PPTX deck created, inspected, adapted from a template, edited, or rendered for visual review.
  • One deliverable spans formats, such as a workbook feeding a reviewed slide deck and a final PDF export.

Choose another workflow when

  • You only need prose in Markdown. Use ak:docs or ak:interview-docs and avoid an Office container.
  • You need application code that imports or exports documents. Use the relevant development Skill; this workflow focuses on the artifacts themselves.
  • You need live control of a desktop Office application or cloud document. Choose a runtime capability that explicitly provides that session.
  • You cannot inspect a rendered result. Do not claim layout fidelity from XML, extracted text, or formulas alone.

Understand the nested format boundary

ak:document-skills physically includes ak:pdf, ak:xlsx, ak:docx, and ak:pptx. They are nested implementation guidance, not individually registered top-level Engineer exports. Invoke the parent Skill and name the format and outcome; do not assume /ak:pdf, /ak:xlsx, /ak:docx, or /ak:pptx is a separately installed command.

ArtifactBundled workflowCore verification
PDFText and table extraction, creation, page operations, rendering, and fillable or annotation-based formsPage count, extracted content, field validation, bounding boxes, and rendered pages
XLSXData analysis, workbook creation or editing, formulas, formatting, charts, and recalculationFormula scan, cached results, sample references, data types, and preserved template conventions
DOCXMarkdown extraction, new document generation, OOXML editing, comments, and tracked changesRe-extracted text or changes, valid package, preserved structure, and rendered pages
PPTXText extraction, new deck generation, template reuse, raw OOXML editing, and thumbnailsSlide inventory, package validation, text overflow, thumbnails, and rendered slides

Prepare inputs, tools, and output boundaries

Before you start:

  • Complete Onboarding, and confirm Engineer Kit is installed for the current runtime and scope.
  • Provide source files, expected output format and path, intended audience, required content, template or brand constraints, and whether originals may be overwritten. Prefer a new output file.
  • State whether comments, tracked changes, formulas, external links, macros, metadata, speaker notes, animations, accessibility tags, or form fields must be preserved.
  • Supply authoritative data and content. Identify confidential, personal, regulated, licensed, or password-protected inputs before processing.
  • Make the required tools available or approve installation separately. Depending on the route, these can include Python, Node.js, pandas, openpyxl, pypdf, pdfplumber, reportlab, Pandoc, MarkItDown, docx, PptxGenJS, Playwright, Sharp, LibreOffice, Poppler, qpdf, OCR tools, and defusedxml.
RuntimeInvocationAvailability boundary
Claude Code/ak:document-skills ...Native delivery can read and write project files and run available converters; Office applications, fonts, browser rendering, and system packages remain environment-specific.
Cursor/ak:document-skills ...Slash invocation is user-verified; equivalent converters, filesystem access, and visual inspection are not established.
Codex$ak:document-skills ...Native Skill discovery is supported; bundled document libraries and local rendering tools still depend on the active environment.

The parent Skill declares no positional arguments, flags, or default output directory. Give it the exact source, target, requested operation, preservation requirements, and verification standard.

Run a format-aware request

/ak:document-skills "Create board-report.xlsx from approved-metrics.csv with formulas for totals and growth, preserve source precision, add a Sources sheet, recalculate with LibreOffice, scan every formula result for errors, and save to ./deliverables without changing the CSV."

For a redline, name the review boundary explicitly: preserve the original DOCX, make only the approved textual changes as tracked changes, retain comments and formatting, render every page, and return both the reviewed file and a change ledger. For a deck, include aspect ratio, template, slide count, audience, speaker-note requirements, approved assets, and the thumbnail or rendered-slide review you expect.

Observe the artifact workflow

  1. The Skill classifies the task. It selects PDF, spreadsheet, Word, or PowerPoint guidance and identifies whether the work is read-only, creation, editing, redlining, template reuse, or form filling.
  2. It inspects before changing. It extracts text or data, inventories pages, sheets, slides, fields, relationships, and formatting, and records what must survive the transformation.
  3. It loads the required nested guidance. DOCX and PPTX creation or OOXML editing require their full referenced guides before implementation. PDF form work branches only after checking whether real fields exist.
  4. It creates a separate working artifact. The workflow uses structured source, unpacked OOXML, HTML slides, JSON inventories, rendered pages, or other intermediates appropriate to the format.
  5. It validates semantics and structure. Formulas, references, fields, tracked changes, slide shapes, package relationships, and requested content are checked before visual judgment.
  6. It renders and inspects when layout matters. Pages, slides, forms, overflow, contrast, clipping, and unintended shifts are reviewed and fixed.
  7. It delivers evidence. The final report names outputs, intermediates, checks, preserved features, mutations, dependencies, and remaining gaps.

Know the files each route can create

RouteTypical intermediate effectsFinal artifact boundary
PDFExtracted text or tables, page images, field-info JSON, field-values JSON, bounding-box JSON, validation images, split pages, or repaired copiesA new PDF plus requested extracted data; the original remains unchanged unless overwrite is explicit
XLSXAnalysis tables, source notes, formulas, charts, and LibreOffice recalculation; first use can create a LibreOffice user macro in the local profileA recalculated workbook with no reported formula errors and documented hardcodes
DOCXExtracted Markdown, unpacked OOXML directories, editing scripts, repacked working copies, PDF renders, and page imagesA new DOCX with requested comments or tracked changes plus text and visual verification
PPTXExtracted Markdown, thumbnail grids, template inventory, outline, working deck, text inventory JSON, replacement JSON, HTML or JavaScript slide sources, unpacked OOXML, and rendered imagesA new PPTX with validated content, structure, overflow, and visual review

PDF forms have two distinct routes. Fillable forms extract declared fields and validate field IDs and values before writing a new PDF. Non-fillable forms use page images, label and entry bounding boxes, automated intersection checks, manual validation-image review, and text annotations. A visual guess without those checks is not a verified filled form.

For spreadsheets, formulas must remain formulas rather than Python-calculated hardcodes. The packaged recalc.py uses LibreOffice to recalculate the workbook, then scans all cells for #VALUE!, #DIV/0!, #REF!, #NAME?, #NULL!, #NUM!, and #N/A. Its JSON reports status, error count, bounded locations, and formula count. On first use it may initialize LibreOffice and write Module1.xba into the user's LibreOffice profile.

Keep document mutations and sensitive data controlled

Office files are executable workflows, not plain text

Conversion, unpacking, recalculation, OCR, browser rendering, and repacking can create many files, change cached values, drop unsupported features, expose hidden content, or execute installed tooling. Work on copies and define the output directory before starting.

  • Do not remove encryption, signatures, protection, macros, external links, or metadata unless the user owns the file and explicitly authorizes that change.
  • Treat comments, tracked changes, hidden sheets, speaker notes, embedded media, formulas, links, and document properties as potentially sensitive even when they are not visible in the main view.
  • Do not send files to OCR, conversion, font, image, or cloud services without approval. The bundled guidance is local-first but does not make every dependency local or already installed.
  • Template slide replacement can clear every inventoried text shape that lacks replacement paragraphs. Review the complete inventory and work on a duplicate.
  • Opening an XLSX with data_only=True and saving it can replace formulas with cached values. Keep formula-preserving and value-reading passes separate.
  • Never overwrite the only copy of a source artifact. Commit, upload, email, publish, print, or share the deliverable only through a separately authorized action.

Verify outputs and evidence

A complete result should include:

  • Exact source and final artifact paths, formats, sizes, page, sheet, or slide counts, and a checksum when integrity matters.
  • All intermediate directories and files that remain, plus any temporary files removed and local profile changes made by tools.
  • Content checks: extracted text, table or data reconciliation, formula counts and errors, field mappings, tracked-change ledger, slide inventory, and speaker-note or comment preservation as applicable.
  • Structural checks: package validation, relationships, formulas, named ranges, fields, references, and openability in the selected format.
  • Visual evidence from rendered pages, slides, thumbnails, or validation images, with clipping, overlap, overflow, contrast, pagination, and font substitution assessed where relevant.
  • Explicit gaps when Office applications, fonts, converters, OCR, browser engines, passwords, source assets, or human review were unavailable.

The pinned stable and beta releases contain the same parent and nested document trees. The parent description does not itself guarantee that all dependencies are installed or that every advanced Office feature round-trips losslessly.

Troubleshoot and interpret limits

SymptomSafe next step
The parent Skill does not pick the right routeName the source extension, exact operation, output extension, preservation needs, and required evidence.
A nested ak:pdf or similar command is not foundInvoke ak:document-skills; the four format subskills are bundled guidance, not registered top-level exports.
PDF text or tables are incompleteDetermine whether the source is scanned, structured unusually, encrypted, or corrupt; render pages and use approved OCR only when needed.
A PDF form has no fieldsUse the non-fillable annotation workflow with validation images and bounding-box checks; do not force the fillable-field script.
Workbook formulas show stale valuesRun the packaged recalculation path with LibreOffice, inspect its JSON, fix every reported error, and reopen without saving a data_only=True copy.
DOCX edits lose formatting or changes are hard to reviewUse raw OOXML and the redlining workflow, make minimal tracked edits in small batches, repack, re-extract, and render.
PPTX text disappears after replacementReinspect the complete text inventory. Shapes omitted from replacement paragraphs are intentionally cleared by the packaged workflow.
The artifact opens but layout is wrongRender with the target-compatible tool and inspect every affected page or slide; successful packaging is not visual proof.
A converter or library is unavailableReport the blocked stage and missing dependency. Do not install system packages or claim verification without approval.
The runtime does not recognize the SkillConfirm target and scope, restart the runtime session, then follow Runtime cannot find a Skill or Agent.

Use ak:docs for Markdown-first documentation and review Projects, artifacts, and checkpoints before a large or destructive artifact transformation.