distill

Essence Extraction via Progressive Compression. Takes a large artifact and runs it through a chain of compression agents where each must compress the previous output by ~50% while preserving the most important information. The key insight: the DROPS at each compression layer — what each agent chose to cut — reveal the priority hierarchy. The waste product IS the signal.

Arguments

$ARGUMENTS — format: [path/to/artifact.md or inline text]

Parse Arguments

Extract:

• artifact_source: a file path or inline text

If the argument looks like a file path (contains / or ends in a common extension), treat it as a path and read the file. Otherwise, treat the entire argument as inline text.

If no argument is provided, ask the user what artifact they want to distill.

Phase 1: Ingest the Artifact

1. Read the file or parse inline text
2. Measure its size (word count, section count)
3. If it is very short (< 500 words), tell the user it is already concise enough — distill works best on substantial artifacts. Offer to proceed anyway if they insist.
4. Present a summary of what will be compressed:
   • Source (file path or "inline text")
   • Word count
   • Number of sections/headers
   • A 2-3 sentence description of the artifact's apparent purpose
5. Confirm with the user before proceeding. Adjust if the user gives feedback.

Phase 2: Run Compression Chain

Run 4 sequential compression agents. Each one:

1. Receives the previous agent's output (or the original artifact for agent 1)
2. Must compress it to roughly 50% of its length (note: code-heavy artifacts may compress non-linearly -- if the input is dominated by code blocks, summarize them rather than stripping entirely, and aim for 40-60% as an acceptable range)
3. Must explicitly list what it DROPPED and why
4. Must preserve the most important information in its judgment

Agent prompt template for each stage:

You are a compression agent. Your job is to compress the following text to roughly 50% of its current length while preserving the most important information.

## Input ({word_count} words)
{previous_output}

## Instructions
1. Read the input carefully
2. Identify what is MOST important (load-bearing claims, decisions, data, actionable items)
3. Identify what is LEAST important (context that can be inferred, repetition, hedging, examples that illustrate already-clear points)
4. Produce a compressed version at roughly {target_word_count} words
5. List EXPLICITLY what you dropped and why

## Output Format

### Compressed ({target_word_count} words target)
[Your compressed version]

### Dropped
| What was cut | Why | Importance (1-5) |
|-------------|-----|-----------------|
| [specific content] | [reason] | [how important was it really] |

### Compression Decisions
- Hardest cut: [what was most painful to remove and why]
- Easiest cut: [what was clearly noise]
- What I'd restore first if given 25% more space: [...]