delegate-to-worker-models

Delegate implementation to cheap or small models safely by pre-deciding everything. Use when splitting work across agent lanes, picking a model tier, or writing a worker prompt: "can a small model do this", "delegate this", "fan this out", "spawn workers for these", "which model for this task". Encodes the lane contract template, the pre-verified guide requirement, and the honest track record of where cheap models succeed and where they quietly fail.

Aller à l'installation

Dépôt: ariaxhan/the-agent-library
Dernière activité de la source: 7 juillet 2026 à 00:03
Langue détectée de SKILL.md: anglais
Étoiles: 1
Forks: 0

Options d'installation

Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.

Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.

Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.

npx skills add https://github.com/ariaxhan/the-agent-library --skill delegate-to-worker-models

La commande reste sur une seule ligne. Faites défiler horizontalement pour la vérifier avant de la copier.

Vous préférez une copie locale ? Téléchargez les fichiers actuellement disponibles dans SkillsMP.

Vérifiez les fichiers source

Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.

Explorateur de fichiers

2 fichiers

SKILL.md

Instructions source · Aperçu en lecture seule

Retour à la recherche Retour à l’installation

extract-invariants

Turn a borrowed metaphor, analogy, or cross-domain inspiration into a testable mechanism, or prove it is decoration and discard it. Use when a design leans on language from another field ("immune system", "evolution", "market", "quorum", "crystallization"), when someone proposes copying how another domain solves a problem, or when reviewing a system whose names sound deeper than its code. Produces either a named domain-free invariant with evidence and falsifiers, or an explicit rejection with the reason it failed.

2026-07-20

design-loud-failures

Design systems so a failure can never present as success. Use when building or reviewing pipelines, extractors, reconciliations, checkers, or any computation whose output humans act on; when missing inputs, unreadable files, or crashes could pass silently; when a tool "found" an explanation that feels too convenient; or on "make this fail loudly", "could this lie to us", "harden the failure paths". Produces named invariants, explicit skip and no-answer states, honest cause classes, and adversarial regression fixtures.

2026-07-10

debate-decisions

A compact decision protocol for non-trivial choices, reviews, governance calls, and live disagreement. Use implicitly at decision points, even when nobody says "debate", whenever a choice is high-impact, hard to reverse, policy-bearing, or has multiple plausible paths that evidence alone cannot settle. Skip for trivial, cheap-to-reverse, or deterministically decidable choices.

2026-07-07

Tout voir dans ce dépôt

Delegate to Worker Models

Cheap models are not junior engineers; they are excellent typists. Observed across production agent runs, small models are reliable in exactly two roles: total-spec execution (every decision already made, written down, and verified) and mechanical, evidence-only verification (run these commands, report the raw output). Everywhere they were handed a decision, they drifted: inventing spec where the prompt was silent, or dying on a trap the coordinator knew about but did not restate. The wins are a decomposition achievement, not a model achievement: the coordinator pre-deciding everything is what makes cheap models safe.

The two safe jobs, and the tell

Total-spec execution: the lane implements a written guide with zero open questions.

Evidence-only verification: the lane runs named checks and returns raw output, never a judgment.

The tell: if the prompt needs the phrase "use your judgment" anywhere, it is a strong-model task. Do not lengthen the prompt to compensate; change the model or make the decision yourself first.

The pre-verified guide (a lane is only as safe as its guide)

Before delegating anything non-mechanical, write a guide the lane executes. Every API signature in it is grepped from the installed package source (the type declarations on disk) the day the guide is written, never recalled from memory: recall is the hallucination risk, in the guide author as much as in the lane. Name the repo conventions with a concrete file to copy from. A lane without a guide is a coin flip; a lane with one lands green.