Stop making AI re-read the same .pdf
The card is written on a read you were already doing. It stays inside the file: what it is, the facts that matter, where to look. The next model skips the parse. PDF first.
Free · no account · nothing stored · spec CC0
The race
Same question. Same file.
An 8-page supply agreement. One side reads every page. The other reads the card inside. Token counts are measured from the real file.
“What is the liability cap, and on which page?”
i
Plain PDF
Eight pages, no card. Every read pays the full parse.
ii
With the card inside
The same file, plus a 1 KB card named open-matter.yaml.
Watch both sides. The card side answers first. Writing the card cost one full read — this is the second.
Estimate, paced so you can watch it. Tokens are measured from the sample (loading…), four characters to a token. Tools that do not look for the card still read the whole PDF.
Now run it on your own PDFHow it works
A document should only have to be understood once.
Every time someone asks an AI about a file, the model starts from nothing. You wait. You pay. The next person does the same work to the same file — which has not changed. The understanding never travels with it.
i
Your agent was going to read it anyway
On the way out it writes a small card: what this file is, the facts that matter, which page they live on. Marginal cost: a few hundred tokens. Not a second paid parse.
ii
We check the card against the pages
The questions come from the document, not from the card. If a number is wrong or missing, we rewrite. A card that fails is not attached.
iii
You get the same file back
Nothing on the page looks different. The next model — Claude, Grok, something on a laptop — opens the card first, then only the page it needs.
What the card enables
Not a cache. A layer other software can trust.
The same pattern as ID3, EXIF, and package.json: a tiny embedded map.
Small models can finish the job
Write the card once with a frontier model. A 3B model given that card has beaten a frontier model still hunting through raw pages.
Lost in the Middle · SLM survey · DeRetSyn
Open the right page first
Filter before you guess. Metadata-filtered retrieval lifts precision by about 15%, and quality nearly 2×. The card is that filter.
Deasy Labs · Two-Step RAG
Decide before you load
One kilobyte to know whether to open the file. Claude Skills route on ~100-token descriptions. Eight skills cost ~500 tokens at startup instead of 70,000.
SKILL.md / AGENTS.md
Know when memory is stale
A content hash is cache-validity for agents. If the pages changed, the understanding is void. Signed cards are next.
Mem0 staleness · C2PA 2.3
A folder becomes a graph
Entities and sections, already named. The expensive GraphRAG step ships pre-computed. Documents as nodes, shared names as edges.
Coming — exploration, not guaranteed truth
Who it’s for
Three desks. Same card.
The file leaves the building. The understanding has to leave with it.
You write agents
Stop handing your agent the haystack.
Claude, Codex, Cursor, Grok already load skills. Drop ours in. The agent opens the card, then one page.
Install for ClaudeYou issue the file
Mark it for other people’s machines.
You wrote it. You card it. Same shape as schema.org. Readers who want the digest can take it. Readers who don’t still have a normal file.
Read the specYou run small models
A 3B can finish what the 70B was hired to start.
Frontier writes the digest. Your on-device stack answers from it.
Why this matters locallyInstall
A Claude plugin. Skills and MCP if you want the pieces.
Claude first — skill plus hands in one install. The same folders work in Cursor, Codex, Grok, and Copilot.
Claude
ChatGPT
Cursor
Grok
Copilot
/plugin marketplace add reisierx/open-matter /plugin install open-matter@open-matter
Spec CC0 · code MIT · github.com/reisierx/open-matter
Later
Cards on files you are already ingesting
Not a prepaid pass over a cold archive. If you already run a pipeline, the card can be exhaust there too. Leave an email.
Your numbers
What a second read costs
Writing each card costs one full read. Value starts at the second read. Time is wall-clock ingest, not billed tokens.
Assumptions, stated: 220 tokens/page, 220-token card, 0.3s/page, 0.2s/card, $3 per million input tokens. Writing the card is included as one full read per document. Inputs stay in this browser.
7.5× fewer tokens · 87% cheaper · $137 / year · 18 hours of model time
Without cards: $158 and 20 hours. With cards (including the write): $21 and 2.1 hours.
Leave an email if you already ingest files