Give your agents company knowledge they can reason over.
SIGN is a plain-text notation for the definitions, rules and relationships your agents work from. It is typed, so an agent can tell a rule from a property. It is compact, so a whole index fits in a system prompt. And every fact records where it came from.
Two ways in, depending on what you need today.
Validate a SIGN file in five minutes, then write your own.
- Clone the spec and run the validator on the examples
- Write and compile your first document
- Load the index into an agent and serve documents on demand
Learn what SIGN changes for your agents, and how to try it on one body of knowledge.
- The problem SIGN solves, in plain terms
- A five-step adoption plan you can hand to an engineer
- What it costs, who maintains it, and what stays under your control
The common formats were built for a different reader.
JSON was built for programs, markdown for people, and RDF for semantic-web tooling. SIGN was designed with an agent's context window as the reader.
| Approach | Where it falls short for agents |
|---|---|
| JSON | Much of each payload is brackets, quotes and keys. It has no way to say how two things relate, what can be inferred, or what is forbidden. |
| Markdown | Compact, but untyped. An agent cannot tell a rule from a property or a constraint. |
| RDF and OWL | Rigorous, but the syntax is heavy, the toolchain is large, and models have seen little of it in training. |
| Prompt text | Fast to write, but unversioned, hard to audit, and copied from prompt to prompt. |
| SIGN | Typed blocks for definitions, relationships, rules and constraints; provenance on every fact; versions and review dates on every document. |
Six kinds of content, each in its own typed block.
Structure
Entities, definitions, properties and the taxonomy they sit in.
Relationships
Typed edges between documents and entities, with every predicate defined in a vocabulary block.
Inference rules
Governed rules that derive new facts, each with a condition, a result, a confidence and a source.
Constraints
Rules an agent must enforce: what is mutually exclusive, what is required, what is forbidden.
Provenance
Whether each fact was asserted or inferred, where it came from, and how it was validated.
Governance
Status, version, review cycle and owner on every document, so agents skip retired rules.
Agents load a short index, then fetch only the documents they need.
Layer 1 is an index of every document: its ID, status, links and a one-line summary. It goes into every system prompt. Layer 2 is the full text of each document, which the agent retrieves by ID when a task needs it.
Everything is open and on GitHub.
sign-lang
The specification, the quick reference card, worked examples, the validator and compiler, and VS Code syntax highlighting.
sign-lang on GitHubsign-tools
Reference tools: sign-ingest converts documents to SIGN, and Claude Code skills draft new documents and encode a code repository.
sign-tools on GitHubCitation
Shepherd, J. and Career Highways (2026). SIGN: Sigil Intelligence Graph Notation. DOI 10.17605/OSF.IO/6YZJ7.
The paper on OSFSee SIGN behind Charter, from Five Deep.
Charter delivers a company's knowledge and rules to its agents in SIGN. Charter requires the area owner's approval. A document the owner adds is approved as it's added. Anyone else's waits for the owner. Test Drive Charter is free for 30 days.
Run the validator on the examples. It takes about five minutes.
You need Git and Node.js 18 or later. The engineer guide walks through each command and what you should see.