Compact JSON with TOON for LLM prompts

Goal: Shrink a JSON document before pasting it into an LLM prompt, by writing it as TOON, which states each array’s field names once instead of repeating them on every row.

Prerequisites: None on the command line, because -F toon writes TOON and -I toon reads it. The built-in toon module does the same conversion from inside a query.

Query

$ mq -I json -F toon '.' users.json

Input (users.json)

{
  "users": [
    { "id": 1, "name": "Alice", "role": "admin" },
    { "id": 2, "name": "Bob", "role": "dev" },
    { "id": 3, "name": "Carol", "role": "dev" }
  ]
}

Output

users[3]{id,name,role}:
  1,Alice,admin
  2,Bob,dev
  3,Carol,dev

Minified, the JSON above is 122 characters and the TOON is 65.

Embed it in a prompt

Use toon_stringify to convert one part of the data and put it inside a larger string:

$ mq -I json 'import "toon" | "Users:\n" + toon::toon_stringify(get("users"))' users.json

Users:
[3]{id,name,role}:
  1,Alice,admin
  2,Bob,dev
  3,Carol,dev

Read TOON back

$ mq -I json -F toon '.' users.json | mq -I toon -F json 'get("users") | map(fn(u): u["name"];)'

[
  "Alice",
  "Bob",
  "Carol"
]

Notes

• The saving comes from arrays of objects that share the same keys, such as database rows or API list responses. A plain nested object still gets shorter, since TOON drops the braces and most quotes, but by less. {"name": "demo", "keywords": ["a", "b"]} becomes name: demo and keywords[2]: a,b.
• Strings that could be confused with another type are still quoted. "1.0.0" is written as "1.0.0", while tsc is written bare.
• The counts above are characters, not tokens. Token counts depend on the model’s tokenizer, so measure with your own before relying on a number.
• -I json is only the source format. YAML, TOML, XML and CSV can be the input too.
• To cut a Markdown document down to size instead, see Trim a document down to LLM-sized context.