# BrandDocs: AI On-Brand Document Generator for Word, PowerPoint & Excel
**BrandDocs is a set of agent skills that learn your existing Word, PowerPoint and Excel templates and generate new on-brand documents from them.** Unlike generic AI document generators, it preserves **brand, structure, styles and formulas by construction**. Built for Claude Code, Codex and compatible AI agents.
[](LICENSE)
[](https://www.python.org/)
[](https://github.com/ferdinandobons/brand-docs/actions/workflows/ci.yml)
[](https://ferdinandobons.github.io/brand-docs/)
[](https://github.com/ferdinandobons/brand-docs/releases/latest)
[](#the-three-skills)
[](#project-status)
---
## What is BrandDocs?
**BrandDocs** is an open-source **agent-skill bundle** that learns a company's existing Office templates and generates new on-brand documents from them. Point it at one branded `.docx`, `.pptx`, or `.xlsx`; it **extracts** the brand (theme colors and fonts, named styles, the document's *structure*, layouts, cover anchors, logos and tables) into a portable **Brand Profile**. From then on, every document it **generates** is built *from the original template shell* and uses *only* the artifacts the template actually defines. Each format stays in its own lane: there is no cross-format conversion.
> **The core guarantee: off-brand output is impossible by construction.** No generator ever writes a literal style name, hex color, or font: those live only in the Brand Profile, and `verify` refuses a profile that points at anything the template doesn't contain.
### At a glance
| Question | Answer |
|---|---|
| **Input** | Existing company `.docx`, `.pptx`, or `.xlsx` templates |
| **Output** | Same-format on-brand Word documents, PowerPoint decks, and Excel workbooks |
| **Works with** | Claude Code, Codex, compatible AI agents, or the direct Python CLI |
| **Best for** | Repeatable reports, decks, workbooks, proposals, memos, briefs, and internal document workflows |
| **Privacy model** | Local-first; no cloud service is required, and real templates are git-ignored |
| **Speed** | First run on a new template: up to ~15 min end to end (extract + model comprehension + visual QA); every later document from the saved profile: seconds |
| **Current release** | [v0.10.1](https://github.com/ferdinandobons/brand-docs/releases/tag/v0.10.1) alpha |
---
## The three skills
| Skill | Format | Generates |
|---|---|---|
| **`brand-docx`** | Word `.docx` | reports, letters and memos in the template's structural order |
| **`brand-pptx`** | PowerPoint `.pptx` | decks from the template's real masters & layouts, with native charts, diagrams & merged tables |
| **`brand-xlsx`** | Excel `.xlsx` | workbooks: named-region fills with **formulas preserved** and brand number formats |
All three share one engine and expose the same verbs: **`extract` → `comprehend` *(optional, model-driven)* → `verify` → `generate`**, plus the learning verbs **`learn` / `propose-overrides` / `refine`** that fold QA findings and user feedback back into the profile. Details → [documentation/SKILLS.md](https://github.com/ferdinandobons/brand-docs/blob/HEAD/documentation/SKILLS.md).
**Several templates, one brand?** `extract --blend` folds a second same-format template into a saved profile at the value-fact level: it fills gaps and corroborates agreements, the primary template wins every conflict, and the brand guarantee is untouched because artifact pointers never cross templates. `compare-profiles` reports brand drift (theme colors, fonts, off-theme usage) between any two saved profiles and exits non-zero on drift, so it can gate brand coherence in CI.
**Two-phase by design:** the deterministic engine works with **no model at all** (extract / verify / generate, fully offline); the model-assisted verbs sit ON TOP and can only NAME captured facts - every proposal is validated fail-closed, so the brand guarantee never depends on a model being right.
**Published only after proof:** generation preflights the profile and shell, writes
to a same-directory temporary file, and atomically publishes the requested output
only after deterministic QA passes. XLSX also preserves formulas and raw worksheet
extensions such as sparkline groups; extension loss is a blocking QA error.
---
## Prerequisites
BrandDocs runs locally and needs a few things installed **before first use**:
- **Python 3.10-3.13** plus the packages in [`requirements.txt`](https://github.com/ferdinandobons/brand-docs/blob/HEAD/requirements.txt) (`python-docx`, `python-pptx`, `openpyxl`, `lxml`, `Pillow`):
```bash
pip install -r requirements.txt
```
- **Visual QA tools (keep the visual gate on):** the render-based **visual QA gate runs by default** and catches layout problems the deterministic checks can't (text overflow, blank pages, clipping, stale demo text). It needs LibreOffice + Poppler (Tesseract is optional, for OCR). Install them with one auto-detecting command:
```bash
bash scripts/setup_visual_qa.sh
```
Generation still runs without them (it degrades gracefully to deterministic-only QA, level L0), but install them so the visual gate stays on.
Check what's present at any time with `python scripts/brandkit/cli.py doctor`. Per-OS commands and the full setup → **[documentation/INSTALLATION.md](https://github.com/ferdinandobons/brand-docs/blob/HEAD/documentation/INSTALLATION.md)**.
---
## Installation
The three skills share one Python engine (`scripts/brandkit/`), so install the **whole repository** (copying a single skill folder on its own won't work). After either install below, set up the [prerequisites](#prerequisites) so the engine can run, then verify with `python scripts/brandkit/cli.py doctor`.
### Claude Code
The plugin loads all three skills plus the shared engine together:
```text
/plugin marketplace add ferdinandobons/brand-docs
/plugin install brand-docs@brand-docs
```
### Codex (and other agents)
Clone the repo and symlink the three skills into your Codex skills directory, so each skill's engine in `scripts/brandkit/` travels with it:
```bash
git clone https://github.com/ferdinandobons/brand-docs.git ~/.codex/brand-docs
cd ~/.codex/brand-docs && python3 -m venv .venv && . .venv/bin/activate && pip install -r requirements.txt
mkdir -p ~/.codex/skills
for s in brand-docx brand-pptx brand-xlsx; do ln -s ~/.codex/brand-docs/skills/$s ~/.codex/skills/$s; done
```
Restart or reload the agent if the skills don't appear immediately.
> Git-submodule install and updating instructions are in **[documentation/INSTALLATION.md](https://github.com/ferdinandobons/brand-docs/blob/HEAD/documentation/INSTALLATION.md)**.
---
## Quick start
**With an AI agent** (the intended experience). Describe what you want and attach a template:
> "Use this company Word template and write a report on the history of Napoleon."
The agent activates `brand-docx`, extracts (or reuses) a Brand Profile, fills the template shell in its structural order, runs QA, and returns the file. PowerPoint (`brand-pptx`) and Excel (`brand-xlsx`) work the same way.
**How long does it take?** The FIRST run on a new template is the slow one: extraction, the optional model comprehension, content authoring and the visual QA gate (plus any repair round) can take up to ~15 minutes end to end with an AI agent. Every later document from the saved profile takes seconds. Still a fraction of formatting the document by hand, and you get a faithful file instead of an approximate one.
**Direct CLI** (the engine, for tests & debugging). No template at hand? Try the
shipped synthetic example: `examples/templates/branddocs_template.docx` (also
`.pptx` / `.xlsx`).
```bash
# 1) Extract the brand from a template into a reusable Brand Profile
python scripts/brandkit/cli.py extract --name