JSON to LLM Context Rules

Goal

Turn JSON trees into short readable summaries that are easier for LLMs to consume than pretty JSON.

Output Modes

• sectioned: default; best for LLM prompting because sections are easier to reference
• flat: simpler bullet list; useful when section headers feel too verbose for tiny payloads

Safety Controls

• strict: keep more literal structure and stop dropping default-false flags
• preserve: force-keep keys or dotted paths such as status, profile.email, or orders
• expand collections: add sample sub-bullets under list summaries
• expand details: add nested sub-bullets under detail summaries
• expand all: expand both details and collections
• show-paths: add [@path] markers so the model can trace each rendered line back to source locations

Core Heuristics

• Group root output into:
  • Summary for status, contact, timestamps, and other headline facts
  • Details for nested objects or extra attributes
  • Collections for arrays and list-like fields
• Remove noise:
  • null
  • empty strings
  • empty arrays
  • empty objects
  • selected default-false flags such as deleted=false
• Preserve high-signal fields:
  • id
  • name, title, label, username
  • status
  • timestamps
  • relationship IDs
  • contact fields such as email
• Prefer readable phrases:
  • active=true → active
  • email + email_verified=true → email x (verified)
• Summarize arrays:
  • small scalar arrays → report line with total and values
  • object arrays → report line with count, top status distribution, and a few examples
• Summarize nested objects:
  • prefer report-style fragments when possible
  • avoid dumping raw nested JSON unless the structure is too irregular

Entity Header Heuristics

Prefer:

Type[id]: name

Type selection order:

1. explicit type
2. parent key name
3. conservative built-in hints such as User, Order, Event
4. fallback Record

Name selection order:

1. name
2. title
3. label
4. username
5. email

Array Examples

Input:

{"orders":[
  {"id":"A12","status":"paid","total":42},
  {"id":"A13","status":"pending","total":18},
  {"id":"A14","status":"paid","total":9}
]}

Readable:

Context

Details
- Meta: count: 3.

Collections
- Orders: 3 total; statuses: paid 2, pending 1; examples: Order[A12] with status: paid and total: 42, and Order[A13] with status: pending and total: 18.

Tuning

• Lower --max-samples to reduce token usage further
• Lower --max-depth for deeply nested payloads
• Lower --max-string-len when text blobs dominate output
• Raise --max-depth only when nested relationship details are important