How big can a file be?

Up to 200 KB. A larger one is refused outright rather than silently truncated, so a file is either stored whole or not stored.

Can I read part of a file?

Yes. A read takes a character limit and an offset, so an agent can page through a long document instead of pulling all of it into context.

Yes. JSON content is validated on write, so a malformed brand kit fails at the point of saving rather than at the point of reading.

Project files

Store long-form project context an agent can load later: a brand kit, a rules registry, templates or a topic queue, keyed, versioned and listable.

Updated October 20, 2026

Files hold the documents a project needs and memory should not: a brand kit, a rules registry, content templates, a topic queue. They are keyed, so a later session reads the current version by name without searching for it.

What a file is for

Long-form, reusable context an agent will load again. The test is whether it is a document rather than a fact.

A brand kit as structured JSON.
A rules registry: the things this project's content must and must not do.
A content template or an outline shape to follow.
A topic queue: what is coming, in order.
Reference material the writing has to respect.

Keys and versions

A file is addressed by key, and writing to a key that exists updates it. That is what makes a key usable as a stable reference: an agent asks for brand-kit and gets whatever the current one is.

Earlier versions are kept, and a read can ask for a specific one, so an overwrite is recoverable. A write can carry a short reason, which is worth filling in: "tightened the banned-words list" tells the next reader what changed without a diff.

Kinds and content types

Each file carries a kind, which is how a list can be filtered to just the brand kit or just the templates, and a content type, so a JSON document and a markdown one are not read the same way.

Listing before reading

A list returns metadata only: key, kind, content type, size and when it was last updated. No content.

That is the right way for an agent to start: discover what context the project holds, then read only the file it needs. Pulling every document into a session wastes the context window on things the task does not touch.

Limits

200 KB per file, refused above rather than truncated. A document near that size is usually several documents.

Files, memory and context

Three stores, one rule for choosing:

A fact goes in memory.
A description of the brand goes in context.
A document goes in a file.

The common mistake is pasting a whole brief into memory, which makes it unretrievable and clutters every memory read. The other direction, a one-line decision saved as a file, is harmless and just means nobody finds it.

Tip

A rules registry is the file most worth writing first. One document saying what this project's content always does and never does removes the same correction from every future session, and an agent reads it before drafting rather than after you reject a draft.

Questions about project files

How big can a file be?: Up to 200 KB. A larger one is refused outright rather than silently truncated, so a file is either stored whole or not stored.
Can I read part of a file?: Yes. A read takes a character limit and an offset, so an agent can page through a long document instead of pulling all of it into context.
Is JSON checked?: Yes. JSON content is validated on write, so a malformed brand kit fails at the point of saving rather than at the point of reading.

Was this page helpful?