TOON — Token-Oriented Object Notation

The CLI

# encode JSON -> TOON
cat data.json | ~/.local/bin/toon

# decode TOON -> JSON
cat data.toon | ~/.local/bin/toon --decode

# non-JSON passes through unchanged (safe to pipe anything)

Requires ~/.local/bin/toon (Node wrapper around @toon-format/toon).

WHEN to auto-encode (you should pipe | toon)

Trigger conditions — at least 2 must match:

• Tool output is JSON and > 2KB
• Contains uniform array of objects (5+ rows, identical keys)
• Will be summarized/analyzed rather than written verbatim
• Mentioned "tokens", "context", "compress" in the task
• Repeated structure (logs, records, time-series)

WHEN to NOT encode

• Output goes to file/DB/other code (not LLM context)
• Small JSON (< 1KB) — overhead exceeds savings
• Deeply nested non-uniform — JSON-compact wins
• Downstream tool expects JSON (round-trip cost)
• Single record, not tabular

Self-check before piping

# count rows + keys homogeneity
cat X.json | jq 'if type == "array" then length else 0 end'
cat X.json | jq '[.[] | keys | join(",")] | unique | length'
# uniform = length matches array length, unique = 1

If uniform + ≥5 rows + >2KB → encode. Otherwise skip.

Format reminder

users[3]{id,name,email}:
  1,Alice,a@x
  2,Bob,b@x
  3,Carol,c@x

[N] = array length, {fields} = column schema, CSV rows below.

Round-trip verification

echo '{"a":1}' | ~/.local/bin/toon | ~/.local/bin/toon --decode
# must equal '{"a":1}'

Honest caveat

This skill is descriptive — I must remember to apply it. True auto-firing needs: (a) an AGENTS.md rule in the workdir, or (b) a Hermes tool-result hook. Without one of those, I default to JSON unless explicitly told.