agentsclimarketplace
Sign in

Consult

Skill mckissinger/keel/skills/consult

Consult the read-only oracle (claude-fable-5) on ONE judgment question — a choice the spec or situation underdetermines, where several options are viable and a reasonable reviewer could pick differently. Distills the question into a brief (question + viable options + relevant paths), dispatches agents/oracle.md, and reports its recommendation, rationale, and disposition. Decorrelated judgment, not authority — it recommends; it decides nothing and changes nothing.From its SKILL.md

Install
npx -y skills add mckissinger/keel --skill consult

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.4 KB, as published. Nobody here has run it

Consult

Ask a decorrelated frontier model one well-formed judgment question and get back a recommendation you can act on. This skill owns the consult mechanism; it is used attended, by you or by a session that has hit a real fork in the road.

Why it exists: keel's two-model policy rests on decorrelation (references/model-routing.md) — a check that shares the builder's blind spots is weak exactly where it matters. The verifier applies that after work exists. A consult applies it before a choice is made, on the narrow class of question where better judgment, not more authority, is what's missing.

What is consultable

A judgment question: the spec or situation underdetermines a choice, several options are genuinely viable, and a reasonable reviewer could pick differently. If the spec already answers it, or one path is obviously right, or it is a routine style pick — don't consult, just decide.

An oracle consult is never how an authorization question gets answered — the oracle recommends on judgment; authorization is the human's, mode or no mode.

The mechanism (owned here)

  1. Distill exactly one question. One dispatch, one question. If you have three, consult on the one that actually blocks you, or run three dispatches — never bundle them into one brief.
  2. Write the brief. Three parts, always: the question; the viable options you considered (with what each costs and buys); the relevant paths — the spec sentence, the files, the code the question turns on, so the oracle can ground its answer instead of guessing.
  3. Dispatch agents/oracle.md with that brief. It is read-only (claude-fable-5, effort high) — it reads the repo to ground the answer and mutates nothing.
  4. Report back: the oracle's recommendation, its rationale, and its disposition — one of answered, uncertain, reframed-as-authorization. Report the disposition as given; never launder an uncertain into a confident answer.

A worked brief

Question. The spec says "store the item's history" without saying whether history is a separate table keyed by item id, or an append-only JSON column on the item row. Which shape?

Viable options. (a) Separate item_history table — clean queries over time ranges, real foreign keys, costs a join on every detail read. (b) Append-only column on items — single-read detail page, no join, but history is unqueryable in aggregate and the row grows without bound.

Relevant paths. specs/features/inventory.md §3 (the ambiguous sentence), db/schema.sql:40-70 (the existing items table), app/items/detail (the only current reader).

Attended semantics

This skill runs attended — you are present — so it carries none of the halt choreography an unattended run needs:

  • No cap. The per-feature consult cap exists to bound unattended ambiguity, where nobody is watching the volume. Attended, you are the cap: consult as often as the questions are real.
  • uncertain halts nothing. It means the oracle could not confidently pick either — useful information. The conversation with you simply continues, usually toward getting whatever would resolve it.
  • reframed-as-authorization halts nothing. It means the question was a "may I", so it is simply yours to answer directly — no consult needed, and none was ever going to substitute.
  • The decision tail. When you adopt a recommendation and the choice is durable — architectural, doctrinal, hard to reverse — this skill offers to record it as a decisions/ entry so the reasoning outlives the transcript. It asks; it never writes one on its own, and declining writes nothing.

What this skill does not own

The mechanism is owned here; the autonomy policy is not. Mode-gating, the judgment-only boundary as a run rule, the per-feature cap, the mode's disposition-halt rules, and the ledger records all belong to the consult contract in skills/auto/SKILL.md — cited here by reference, never restated. Under an active mode, that contract governs; this skill's attended semantics above do not apply there.

Boundaries

  • A consult recommends; it never decides and never authorizes. Adopting a recommendation is your call (or, under a mode, the contract's) — the oracle's report is an input.
  • Read-only, always. The oracle mutates nothing. A consult costs tokens and nothing else.
  • One question per dispatch. A bundled brief gets a worse answer to every part of it.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.