Skip to main content
Agent prompt
Paste that into Claude Code, Cursor, Codex, or any agent that can fetch a URL. It installs Charter, adds the skill that teaches the agent to write packs, and points it at these docs. The instructions are served as plain text, so you can read exactly what your agent will read before it runs anything. The instructions live at that URL rather than in the prompt, so a prompt you pasted last month still does the current thing. They install a skill file and, only if you ask for one, an MCP server. Both are inspectable before you run anything: the skill and the server.

​
Three things that say “agent”

The skill is the one to install first. It needs no server, no config file and no restart.

​
The pack-writing skill

Twelve packs will not cover the API you use. A pack is declarations, so an agent that knows the rules can write one for whatever you use, and the conformance suite gates the result. The skill is those rules, as one markdown file.
For a project-scoped install, put it in .claude/skills/ inside the repo, so it travels with the code and your teammates get it from a checkout. It is a document, not a package: you can also open the file and paste it into whatever your agent reads.
The skill is deliberately not delivered through Charter’s MCP server. The server serves tools at runtime; the skill is guidance at authoring time, often on a different machine and always at a different moment. Bundling them would put a server install, a config edit and a restart in front of a file you can curl.

​
What it makes the agent do

The skill is a checklist, in order, and the order is the point.
One endpoint routinely spans ten documentation pages: the request body references a resource, the resource references an enum, the enum’s values are documented somewhere else. Miss one enum value and the model will eventually emit it and earn a 400 you cannot reproduce.This is the step an agent skips if you do not tell it not to, and it is the one that decides whether the pack is right.
Each field’s description is the vendor’s own sentence, copied rather than summarised, so the declaration can be read against the page it came from.Anything the agent wants to add, a unit, a conversion, a value the API reads as something else, goes in a Gloss. It is appended to the description the model reads and absent from the wire schema, so the vendor’s text stays diffable against their page. Left in the description instead, the pack’s sentence and the vendor’s become one string, and the next regeneration from the docs deletes the help without anyone seeing it go.
Path(), Query(), Body(); see the wire contract. An unmarked top-level field is refused rather than guessed, so a pack that skips this fails on its first call rather than silently sending a field to the wrong place.
Not per field. The key case cascade has four levels, and picking the widest one that is true keeps the declaration honest as the pack grows.
Where the API wants a wire format and the model should speak meaning: MIME, base64url, protobuf Value. See transforms.
An Envelope on the factory, never logic inside a response handler. This is the rule with the sharpest consequence: a handler-level check is opt-in per tool, so the first tool somebody adds without one reports a failed write as a success.
Every HTTP interaction mocked. A pack whose tests reach the network is a pack whose tests will fail for reasons that have nothing to do with the pack.

​
Then check it mechanically

The skill gets the shape right. The conformance suite proves it: nineteen properties that must hold for every pack, checked without knowing anything about any particular API, and each one verified against a pack broken on purpose so the check cannot go vacuous. That pairing is the whole idea. Guidance makes a stranger’s pack come out in the right shape, and a gate that does not care whether the stranger read the guidance holds it there.

​
These docs as a source

Charter is new enough that a model’s training data is unreliable about it. Pointing an agent at the live docs costs one fetch and removes a whole class of confident wrong answer.
  • https://docs.r28.ai/charter/llms.txt lists every page with its description. Fetch it first to discover what exists.
  • Append .md to any URL for the page as markdown, with no navigation to strip: https://docs.r28.ai/charter/tools/wire-contract.md.
  • The menu beside every page title opens the page in ChatGPT or Claude, or connects this site as an MCP server to Cursor or VS Code.
Most questions are answered by one of these four: For authoring, the skill beats the docs. It is the rules in the order they have to be applied, rather than a reference an agent has to assemble an approach from.

​
A rule worth giving your agent

Charter describes one request. It ships no agent loop, no pagination loop and no retry policy, and it never obtains an OAuth grant. An agent that assumes otherwise will write plausible code against an API that does not exist. Limitations is the page that prevents it, and it is worth putting in front of a model before it writes anything substantial. A second rule, for an agent assembling a tool surface rather than writing a pack: a pack tool models its whole endpoint, and a few of them are wide. gsheets.spreadsheets_batch_update is 52,168 tokens of schema, because Sheets puts 74 kinds of edit behind one method. Narrow it with Tool.derived instead of dropping the tool or writing a thinner one by hand.