Distill not transcribe

distill-not-transcribe

Build skills from judgment, not from steps. A skill earns its place in the context window by encoding decisions a bare model would not make on its own. (Chinese twin: SKILL.zh.md — same map, zh-specific copy checks live there.)

Stance — load-bearing biases

People will disagree with these. That is the filter, not a problem.

1. Judgment density before output. Material with steps but no decisions becomes a script, not a skill. Why: agents already execute steps well; what they lack is the author's "when X, choose A over B". Observed (2026-06): of 20,000+ skills in public directories, most are step lists; the #1 item on a paid curated marketplace had 116 installs. Skills without judgment die unnoticed.
2. Precision before recall. Write negative triggers. A skill that stays silent beats one that over-fires. Why: a mis-triggered skill pollutes the context window — decoration in another form. Official guidance favors "pushy" descriptions; this skill deliberately takes the opposite side.
3. Honest lineage. Derived work declares its source. Why: provenance is reverse-verifiable; concealed derivation, once detected, down-weights the entire repository, not just one file.

Route first (always)

Before anything else, check whether a comparable skill already exists (search directories / awesome lists / the marketplace).

• A comparable original exists → fork line → load reference/fork.md
• Nothing comparable → distill line → load reference/distill.md

Gate: judgment-density screen

Run before producing anything. Two or more failures → refuse (see Failure behavior).

1. Does the material contain at least three decisions of the form "when X, choose A over B, because ___"?
2. Has the author actually run this — can they point to one place where it failed?
3. Strip every opinion out. Does what remains still justify a skill? If yes, it is a generic script: do not build. The value must live in the judgments.
4. For each strong opinion: who does it help, and when would following it hurt? No answer → decorative bias → drop the opinion, or drop the skill.
5. Would a bare frontier model already do this at ~80% unprompted? Name the exact missing 15–20 points. Cannot name them → there is no skill here.

Workflow gate

Report the planned structure (sections, reference files, what loads when) and the gate results to the user BEFORE writing any file. Wait for confirmation.

Build

• Distill line: seven transformations → reference/distill.md
• Fork line: four questions + test-trace page → reference/fork.md
• Every skill ships the six mandatory fields → reference/fields.md
• Evidence rules for thickening (where observations may come from, where they may not) → reference/evidence.md

Acceptance — all four, in order

1. Before/after. Same input, bare model vs with-skill, side by side. No visible gap → do not ship. Visible gap → it goes to the top of the README.
2. Mis-trigger set. Five or more queries that must NOT trigger; verify silence.
3. Self-referential demo. The skill demonstrates itself on itself or on one real case.
4. Author redline grep. Scan all output for the author's declared redline terms (identity, organization, location granularity). The list is author-supplied; never assume it is empty.

Failure behavior

If the gate fails: stop and report "this material is a script, not a skill — it contains steps but no decisions", listing which gate questions failed. Do not pad with generic best practices to force a result. Silent fabrication — producing something plausible instead of reporting failure — is the worst failure mode this skill knows.

Lifecycle — this skill's own mandatory fields

• Version: 0.1 (2026-06-12). Pre-release: before/after evidence pending first real run (see examples/recursive-acceptance.md).
• Deprecation condition: when a bare frontier model, asked to "turn this document into a skill", reliably produces output with negative triggers, boundary conditions, and declared lineage — retire this skill.
• Composition: composes-with the official skill-creator: use its trigger-rate eval loop and packaging tooling; this skill governs judgment screening and trust structure, not tooling. Yields to explicit user instructions.
• Boundary: validated on structured, single-author protocol documents. Raw dialogue transcripts and multi-author material: see distill.md "Raw material" — that path is NOT yet validated.
• Time: judgments dated 2026-06. Ecosystem observations expected to decay within 12–18 months; the gate questions should outlive them.