How to turn a brand PDF into instructions AI can use

Why the PDF doesn’t work as it is

Models can read PDFs: Claude and OpenAI’s API use extracted text and page images; Gemini uses native vision. The conversion still needs to resolve what the guide leaves implicit:

Recording the guide and deciding unresolved rules are different jobs. Brand guidelines vs brand context explains what an agent needs for one task.

Gather what the PDF leaves out

Collect:

• The latest guide, date, addenda and recorded changes.
• Original logo variants, fonts and approved photography.
• Licences covering the tools and uses you plan.
• Renamed products, retired values and claims no longer approved.
• Recent approved work, rejected examples with reasons, and the person who settles questions.

For client intake, use the onboarding checklist. Without a guide, start from the website.

Inventory the guide before rewriting it

A summary can silently choose between conflicting values. Ask for a page-linked inventory first:

Go through the attached brand guide page by page and build an inventory, not a summary. For every rule, value and asset, record the page number and the exact wording or value. Label an item PICTURE ONLY if the page shows it but no text states it, ADJECTIVE if it describes without saying what to do, FILE NEEDED for any logo, font or image I’ll need as a real file, CONFLICT if another page or another attached document disagrees, and OUT OF DATE if it may no longer be true. End with a list of what the guide doesn’t cover. Don’t rewrite or fix anything yet.

Section and page:
Exact wording or value:
What the page shows without stating:
Original file still needed:
Conflict with another source:
Decision or open question:

Split a long guide into sections when needed, as Anthropic recommends. If layout matters, export slides or documents as PDF: OpenAI’s Responses API extracts only text from .pptx/.docx, while Gemini’s document vision meaningfully understands PDFs. Chat-app upload behaviour can differ.

Check exact values against selectable text or original design files. Vision interpretations need verification, especially small measurements and colour codes.

Write rules that stand alone

Keep observations separate from rules the owner confirms. For measurable text contrast, use the WCAG requirements.

Complete rules matter when tools retrieve excerpts. Claude projects can switch to retrieval; OpenAI vector stores chunk files. Anthropic’s contextual-retrieval example shows why a fragment without its subject loses meaning. Name the subject, value and condition in each rule.

Use the Markdown template for the finished structure and the voice-guide instructions for writing rules.

Swap pictures for files

Do not silently pick between conflicting colour conversions or crop low-resolution photos out of the guide when originals exist. Sharing brand assets covers making those files reachable.

Settle conflicts and retire old rules

1. Collect conflicting values, editions and decisions made after publication.
2. Resolve them against current files and recorded changes; ask the owner when those do not settle it.
3. List retired names and files beside their replacements. Keep campaign treatments separate from standing rules.
4. Leave unresolved gaps explicit, with who decides them.

Give the result a date, owner and version; versioning brand guidelines covers the record.

Choose where the result lives

Standards describes its Markdown export as a snapshot. Uploaded copies do not reflect later changes.

With Rainbrand, create an account, connect your agent, then add the PDF and original files. Use the file picker for local files in supported clients.

Add [your brand] to Rainbrand from the attached guide, current written decisions, logo artwork, font files and approved photos. Follow the attached source-priority note where they disagree. Wait until the brand is ready, then compare its guidance with the inventory and list anything missing or different.

Inspect the resulting identity, guidance and assets against your inventory. Correct mistakes through an edit: describe, inspect the proposal and apply it. Once ready, agents request a Brief for each task.

Products receiving customer PDFs can use the same sources through the create-brand API; customer-brand management covers the lifecycle.

Test it on real work

Run recent tasks with the converted brand and compare with approved work. Diagnose each miss:

• Wrong guidance: correct the brand.
• Correct file or rule never delivered: fix the handoff.
• Correct context ignored: revise the output and check again.

Rainbrand Review checks work against the brand. Testing agents covers repeatable cases.