compress-prompt

Compress Prompt

Goal

Transform a prompt into the minimal instruction needed for the model to succeed. Not "preserve everything densely"—instead, "what's the least I need to say?"

Output: Display compressed result + stats. Optionally write to file with --output <path>.

Input

$ARGUMENTS = prompt (file path or inline text) [--output path]

If file path: read content. If inline: use directly. If ambiguous: try as file first.

Principles

1. Trust capability, enforce discipline - Models know HOW to do tasks. But they cut corners, forget context, skip verification, declare victory early. Drop capability instructions, keep discipline guardrails.

2. Goal over process - State WHAT to achieve, not HOW. Let the model choose its approach.

3. Training filter - "Would a competent person need to be told this?" If no → drop it. Models are trained on millions of examples.

4. Maximize action space - Fewer constraints = more freedom = better results. Each constraint should earn its place.

5. Inline-typable brevity - Short enough you could type it verbally to a capable colleague.

6. Avoid arbitrary values - "Max 4 rounds" or "2-3 examples" become rigid rules. State the principle, not the number. Constrain productively while giving flexibility.

What to Keep vs Drop

Execution discipline examples (KEEP these):

• "Write findings to file BEFORE proceeding" — prevents context rot
• "Don't finalize until X confirmed" — prevents premature completion
• "Read full log before synthesis" — restores lost context

Training-redundant examples (DROP these):

• "Be thorough", "Handle errors gracefully", "Ask clarifying questions"
• "Consider edge cases", "Use professional tone"

Constraints

Create todo list - Track: input validation, compression, verification iterations, output.

Verify with agent - Launch prompt-compression-verifier to check goal clarity, novel constraints preserved, no over-specification. Iterate until verification passes.

Single paragraph output - The compressed prompt must be one dense paragraph, not reformatted sections or bullets.

Non-destructive - Original file untouched. Display output + optional file save.

Output Format

Compressed: {source}

Original: {tokens} tokens
Compressed: {tokens} tokens ({percentage}% reduction)

---
{compressed paragraph}
---

Verification: PASSED/INCOMPLETE ({iterations} iteration(s))

Example

Before (1,247 tokens): Full code reviewer prompt with phases, edge cases, examples...

After (67 tokens):

Review code for bugs, security issues, performance problems; success = all critical issues identified with actionable fixes. Output JSON {file, line, issue, severity, fix}. Never approve code with critical issues.

Kept: Goal, acceptance criteria, output format, novel constraint (never approve with critical issues). Dropped: Process phases, edge case handling, examples, obvious constraints.