Skip to content

Latest commit

 

History

History
70 lines (60 loc) · 4.95 KB

File metadata and controls

70 lines (60 loc) · 4.95 KB

The three skills & project status

The three skills

Skill Format Generates
brand-docx Word .docx reports, letters, memos: cover, headings, paragraphs, callouts, quotes, captions, lists, tables, in the template's structural order
brand-pptx PowerPoint .pptx decks: title / section / content slides from the template's real masters & layouts, with real bullet levels and long-text splitting
brand-xlsx Excel .xlsx workbooks: fills named cells & regions while preserving formulas and workbook structure

All three expose the same core verbs: extractverifygenerate. Each skill is self-contained and same-format (a Word template makes Word documents, never a deck or a sheet). They share one engine: a single profile schema, resolver, OOXML layer and QA gate underpin all three formats.

On top of the deterministic core sit the model-assisted verbs, each one fail-closed (the model only NAMES captured facts; the engine validates every proposal and authors every value):

  • comprehend - the model reads a bounded summary of the template and records what each structure is for: cover slots, demo-vs-real regions, derived-index conventions, brand-color naming (plus optional palette aliases for off-theme accents), faked-heading adjudication (promote_appearance), and recurring layouts proposed as reusable component/section fragments. Frozen into the Brand Profile; generation works with no comprehension at all (the deterministic path).
  • learn - deterministically distills RECURRING QA findings from the cross-run generation history into a shell-frozen overrides lesson (advisory until --accept).
  • propose-overrides - the model proposes corrections for the ambiguous recurring remainder learn could not bind, through the same fail-closed sink.
  • refine - turns end-of-generation user feedback (text or a screenshot) into a validated comprehension delta, improving FUTURE generations.

Project status

Alpha, maturing. The Word vertical (brand-docx) is the reference implementation, verified end-to-end on real templates; PowerPoint and Excel share the engine and are catching up.

Area Status
Shared engine (profile schema, resolver, OOXML, CLI, dual store) ✅ working
brand-docx: extract → verify → generate ✅ working
Document structure extraction & order-aware generation ✅ working
Brand-guarantee enforcement (verify fails on missing artifacts) ✅ working
Deterministic QA (L0: styles, palette, residual text, tables, formula and XLSX extension preservation, language) ✅ working
brand-pptx: roles from real layouts, native charts / SmartArt / merged tables ✅ working (fidelity still catching up to docx)
brand-xlsx: named-region fills, formula-preserving, native charts ✅ working (fidelity still catching up to docx)
Visual QA (LibreOffice render + manifest-driven repair loop) 🚧 implemented with graceful degraded mode
Native charts (DOCX / PPTX / XLSX), SmartArt diagrams (DOCX / PPTX), merged tables ✅ working
Native Word toc (authored field, or deferral to a preserved outline TOC) with a Word-faithful refreshed cache (authored bookmarks, per-level styles, PAGEREF; template demo entries never survive) ✅ working
Excel semantic number formats (number.<family> resolved to the template's mask) ✅ working
Model-driven reusable-fragment population (comprehendcomponents / sections, with {{slot}} substitution) ✅ working
Brand appearance capture: typography (all 3 formats) + paragraph geometry / table conditional formats / list numbering (docx), dominant-sampled and shell-verified ✅ working
Learn-from-errors loop (cross-run generation report → learn / propose-overrides, advisory until --accept) ✅ working
Faked-heading detection + model adjudication (pseudo_headingspromote_appearance) ✅ working (docx-first)
Cover synthesis when the template has no cover anchor (anchors.cover.kind == NONE) ✅ working (docx + pptx; xlsx N/A by design)
PyMuPDF PDF raster fallback ✅ working
Optional OCR rendered-text residual scan ✅ working when Tesseract is installed
Template-based skill eval set (DOCX/PPTX/XLSX) ✅ working in CI
Strict visual mode (--qa strict) ✅ working
Multi-template profile blending (extract --blend, same-format, value-facts only, pointers never cross shells) ✅ working
Cross-template brand-drift report (compare-profiles, read-only, exit 1 on drift) ✅ working
Role-first authoring surface (PROFILE.md role / palette-role / hints sections + per-skill authoring guidance) ✅ working
Local-corpus fidelity benchmark (scripts/corpus_benchmark.py, real templates outside the repo) ✅ working
Richer image analysis 🔭 planned

Visual Word overflow needs LibreOffice, since Word lays out at render time.