01 / Knowledge layer for agents
Books become queryable skills. Answers stay sourced, token bills stay small.
Skill pass
OriginScattered docs
DestinationMaster skill
02 / What ATOM builds
One concept per note, each with its source. Cross-checked concepts merge into one consensus note. Disagreements stay visible as contested, with both sides and their sources, instead of being silently picked for you.
The index routes a question straight to the right note. Measured on a 118-concept master skill: 71 tokens through the taxonomy against 2,201 scanning flat. That gap is the product.
Concepts are linked by typed edges, so the agent can follow a relationship instead of guessing one. The measured master skill carries 124 edges across its 118 concepts.
Facets slice the same brain by angle, and the registry keeps every master skill findable. One field, one master skill. Ask in one field and only that field loads.
Photographed and scanned PDFs go through real OCR instead of producing empty skills. Vietnamese and English measured at 0.3% character error, with a cache that makes a re-run far faster.
Ask inside exactly one document. A bound view answers from that text alone and never blends in the rest of the field. Built for regulations, where the exact document is the answer.
03 / In the beginning
Every turn, the agent re-reads the PDF, re-finds the chapter, re-pays the token bill, and still guesses when the answer is not in front of it. Your agent does not have a reading problem. It has a memory problem.
Search the PDF and you get a list of pages, not answers. Ask about the book and the agent answers from vibes. Take notes as you read and you end up with a 200-line doc nobody opens again.
ATOM pays the structuring cost once, at conversion. Every question after that costs only the answer, not the book. The book becomes part of your workflow, and part of your project’s brain.
04 / How it works
01
Point it at a file, folder, or glob. The deterministic extractor pulls clean text and metadata, picks the right tool for the book, and OCRs the scans that have no text layer at all.
Formats PDF, EPUB, DOCX, MD, HTML, RTF, MOBI
02
The generator turns the text into atomic notes and builds the skill: frameworks, decision rules, anti-patterns, per-chapter files. Structure, not a summary. Conflicts stay contested.
Built notes, taxonomy, typed graph, facets, registry
03
Your agent loads the skill on demand and answers from the real content. One field, one master skill. A bound view answers from one exact document and nothing else.
Ask /my-book replication, routed in 71 tokens
05 / Token math
Measured on a real 118-concept master skill, not borrowed. Multiply by questions, by turns, by users. A project that asks a thousand questions does not pay a thousand book-readings. It pays a thousand lookups.
100 questions
The honest floor. Below roughly 5,000 tokens, about 10 A4 pages, dumping the document is still cheaper. ATOM earns its keep on everything longer, and the longer the document, the bigger the win.
06 / Experience the trilogy
Point it at a file, folder, or glob. Flexible chapter detection splits the book, and the generator distills structure, not a summary: frameworks, decision rules, anti-patterns, per-chapter files.
How it worksMany sources, one field. Atomic notes, taxonomy, typed graph, facets, and a registry build a single queryable brain. Fold new material in as it lands, without rebuilding from scratch.
Master skillsBuilt for projects, not casual readers. The layer other agents and assistants plug into: feed it regulations, procedures, and domain docs, then query it on demand. Bound views answer from one exact document.
For assistants07 / Install
One skill, read by every host that supports the open Agent Skills standard. No account, no upload. Extraction runs locally.
Works with Copilot CLI, Amp, Claude Code, Codex, Hermes Agent, OpenCode, OpenClaw
Needs Python 3.9+ for the extractor. Plain text needs no extra tools
Install the skill
# any host, via the skills CLI npx skills add jcarus-vn/ATOM
Convert your first book
$ /atom ./my-book.pdf # then ask: /my-book replication
Check your extractors
python3 scripts/extract.py --check