| name | judgment-compilation |
| description | Find every judgment call in a draft prompt and compile it into a form a Haiku-class model can execute: a decision table over countable inputs, an ensemble vote, an evidence artifact, or a mechanical escalation. Use when / Trigger: writing or reviewing ANY prompt text for this marketplace — especially before shipping a SKILL/command/agent containing words like assess, decide, right-size, when it matters, if needed, appropriately, thoroughly, sensibly. |
Judgment compilation
A frontier model reads "verify thoroughly when it matters" and supplies the
judgment; a small model reads it and resolves — fluently, every time — to "not
now." You cannot instruct that gap away, because following the instruction is
the missing judgment. So judgment is compiled once, at design time, by you
into structure the model only has to execute. The theory:
gap-model.
The move
-
Detect. Grep the draft:
assess|appropriate|when it matters|right-size|if needed|if unsure|thoroughly|carefully|use judgment|as necessary|sensibly|significant|complex|important|clean up|keep in mind
Every hit is a compilation site. So is every adverb of degree.
-
Compile each hit into exactly one of the four target forms:
| Target form | Use when the judgment is… | Compiled shape |
|---|
| decision table (M1) | a sizing/classification call | scored rubric over countable inputs → tier → fixed prescription |
| ensemble vote (M2) | a verdict that can be wrong | N independent runs (N from tier: 1/3/5), mechanical per-criterion majority |
| evidence artifact (M3/M8) | an honesty/"did it happen" claim | a file a script can stat/schema-check/grep; pre-written probe with attached output |
| escalation (M7) | "knows when it's out of depth" | countable trigger (retry ×2, split vote, flag present) → more votes → pinned model → human |
Full per-pattern recipes with before/after examples:
references/compilation-recipes.md.
-
Delete what compiles to nothing. "Clean up nearby code if appropriate"
compiles to removal — unrequested work is a scope defect, not a judgment
to preserve.
-
Pin the residue. Genuine invention (design synthesis, novel probe
authoring) refuses all four forms. Do not paper over it with a longer
instruction: pin the agent to a bigger model via model: frontmatter and
record why in an ADR, with a fixture-based downgrade trigger
(ADR-0001).
One example (full catalog in the recipes file)
Draft: "Verify the fix works when the change is risky."
Compiled:
risk flags (each +2): migration | auth/payments | public API | data deletion
score 0–1 → T1: run the named test command, attach transcript
score ≥2 → T2: + dispatch verifier with probe list P1–P3, 3 votes,
verdict.json into the evidence bundle
The model counts flags and follows a row. Nothing is left to "risky."
Rationalizations
| Thought | Reality |
|---|
| "Haiku will infer 'when it matters' from context" | It resolves to "not now" — every time, fluently |
| "I'll add emphasis: IMPORTANT, ALWAYS" | Emphasis doesn't survive the mid-prompt attention dip; structure does |
| "A table can't capture this nuance" | Then it's residue — pin it and say so in an ADR; don't ship the nuance as prose |
| "The verb is harmless, it's obvious" | Obvious to you — you have the judgment being deleted; grep says it ships |
Done when
Zero detector hits on the execution path; every compiled site names its target
form; residue (if any) is pinned + ADR'd; each new threshold has a boundary
fixture per the plugin-authoring done-bar.
