TOON Format Skill

TOON Format Skill

TOON (Token-Oriented Object Notation) is a compact, human-readable, schema-aware encoding of JSON designed to be easy for LLMs to read and (with good prompting) to generate. It is especially effective for uniform arrays of objects (tabular data), where keys are declared once and rows carry values.

Quick decision: should I use TOON?

Use TOON when:

• You are passing large structured context (especially arrays of objects) to an LLM and want fewer tokens.
• You want structure guardrails ([N] counts + {fields} headers) to reduce model drift and enable validation.
• You need a format that stays close to CSV’s readability but is a lossless representation of JSON.

Prefer JSON (or JSONL/​CSV) when:

• You need interoperability with systems that expect standard JSON.
• Data is extremely irregular (ragged objects) and the TOON representation won’t be more compact or readable.
• The task is primarily “machine ↔ machine” with no human/​LLM in the loop.

For more detail: see reference/​when-to-use.md.

Core patterns you must follow

1) Recognize TOON in the wild

TOON commonly looks like:

• Scalars/​objects as key: value / indented blocks
• Arrays with lengths: users[3]: ...
• Tabular arrays with field headers: users[3]{id,name,role}:
• Rows are delimiter-separated (comma by default; tab often best)

2) Converting JSON ↔ TOON (recommended paths)

A) CLI (fastest, good for pipelines)

Use the official CLI for round-tripping and stats.

Encode JSON → TOON

npx @toon-format/​cli input.json -o output.toon

Decode TOON → JSON

npx @toon-format/​cli input.toon -o output.json

Token stats + maximum efficiency

npx @toon-format/​cli input.json --delimiter "\t" --stats -o output.toon

Optional structure compression

• Key folding (encode): --keyFolding safe
• Path expansion (decode): --expandPaths safe

For full CLI options and examples: see reference/​cli.md.

B) TypeScript/​JavaScript library

Use @toon-format/​toon for in-app encode/​decode. See reference/​library.md.

3) Prompting patterns for LLMs (critical)

When sending TOON as input:

• Show, don’t explain.
• Wrap data in a fenced code block labeled toon (or yaml if needed):

  • (TOON data)

When asking for TOON output:

• Provide the expected header template (e.g., users[N]{id,name,role}:)
• Explicit rules:
  • 2-space indent
  • [N] must match row count
  • no extra commentary; output only the code block (if you need strict parsing)

For proven prompt patterns and pitfalls: see reference/​llm-prompting.md.

4) Always validate model-generated TOON

If the model outputs TOON, you MUST validate it (strict mode by default in official tooling). If validation fails:

• Ask for re-output with the same header template
• Or fall back to JSON output

Minimal working examples

Example: filter rows (LLM output expected as TOON)

Input

users[3]{id,name,role,lastLogin}:
  1,Alice,admin,2025-01-15T10:30:00Z
  2,Bob,user,2025-01-14T15:22:00Z
  3,Charlie,user,2025-01-13T09:45:00Z

Instruction Return only users with role "user" as TOON. Use the same header format. Set [N] correctly. Output only the code block.

---

Utilities included in this Skill

• scripts/​toon_convert.py: A convenience wrapper that:
  • detects TOON code fences in Markdown and extracts them
  • calls the official CLI via toon (if installed) or npx @toon-format/​cli
  • supports --encode/​--decode, --delimiter, --stats, and folding/​expansion flags

See: scripts/​README.md