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 ( ~50%

• Agent 2: ~50% -> ~25%
• Agent 3: ~25% -> ~12%
• Agent 4: ~12% -> ~6% (the "elevator pitch")

Track the full drop log from every agent for use in Phase 3.

Phase 3: Analyze the Priority Hierarchy

After all 4 compression stages complete, produce a full analysis with these sections:

1. The Essence The final ~6% compressed version — the irreducible core of the artifact.

2. Priority Hierarchy Classify every piece of content by how many compression rounds it survived:

• Tier 1 (Core): survived all 4 compressions — the absolute essentials
• Tier 2 (Important): survived 3 compressions — important but not irreducible
• Tier 3 (Supporting): survived 2 compressions — adds value but not critical
• Tier 4 (Context): survived 1 compression — background/context
• Tier 5 (Noise): dropped in round 1 — likely not load-bearing

3. Compression Difficulty Map What was hardest to cut at each stage. These are the areas where priority is ambiguous or contested — the interesting boundaries.

4. Restoration Order If you could add things back one at a time from the essence outward, what order would you restore? This is the true priority ranking of the artifact's content.

Save the full analysis to distill_{slugified_topic}.md in the working directory.

Phase 4: Present

Show the user:

• The final essence (Tier 1)
• The priority hierarchy (all 5 tiers)
• The hardest cuts (areas of ambiguity)
• The restoration order

Then ask: does this priority ranking match your intuition? Where does it diverge?

Rules

• Sequential by design: each agent must see only the previous agent's output, not the original. Compression judgments at each layer only work with the material at hand.
• Explicit drops: every agent must say what they cut and why. The drops are the point.
• No padding: compressed output should not add new information, hedging, or meta-commentary about the compression process.
• ~50% target is approximate: within 40-60% is fine. Do not pad to hit a word count.
• Preserve structure: when compressing, prefer keeping structure (headers, lists) and cutting content within sections, rather than flattening structure.
• No filtering: include all drop data in the analysis, even when agents disagree about importance. Tension between layers is signal.
• Be honest about ambiguity: if the priority hierarchy has unclear boundaries, say so. The difficulty map exists for this reason.

Pipeline Position

Versatile — works after any phase that produces a large artifact:

/diverge synthesis -> /distill -> priority hierarchy
/converge report  -> /distill -> decision essence
Research notes    -> /distill -> core findings
Design doc        -> /distill -> essential requirements
Meeting notes     -> /distill -> action items and decisions

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sjarmak
• Source: sjarmak/agent-workflows
• License: MIT

Install and usage instructions live in the source repository linked above.