System constraints · tool policies · task detail · repeated examples · historical turns
12,480 tokensWIP LAB 001 / BENCHMARK 001
Context Diet
Compression with receipts. A staged prompt-analysis lab for reducing context while preserving the instructions that matter.
PROMPT ANALYZER / CONCEPT
See the context before you cut it.
Context Diet will separate signal from repetition, expose the proposed reduction, and retain an audit trail before any export is made.
Preserved constraints · normalized tools · compressed examples · cited decisions
4,990 tokens
BENCHMARK SIGNALS
What Benchmark 001 will report
Token reduction
Fewer tokens, explicit delta.
Faithfulness retained
Meaning checked against the source.
Latency saved
Measure the response-time change.
Cost saved
Estimate input-token savings.
Export validity
Validate the generated SKILL.md.
COMPRESSION MOTIF
Reduce the window. Keep the warrant.
The objective is not shorter prompts at any cost. It is a smaller, verifiable context package that still explains its operating constraints.
SKILL.md EXPORT
From analysis to an inspectable skill.
Once an analyst approves a diet, Context Diet will package the retained instructions, assumptions, benchmark notes, and provenance into a GAIA-compatible SKILL.md export. The export remains a proposal until its evidence is reviewed.
# Context Diet export
## Retained constraints
- Preserve task intent
- Cite benchmark evidence
## Provenance
status: proposedSUPPORT STATES
Built for the moments between results.
Empty
No prompt loaded yet.
Analyzing
Evidence pass in progress.
Blocked
Export needs another pass.