LLM Delegation — The Cheapest Thing That Can Answer
SkillAI & modelsApply when a step in a script, a workflow or a CI job needs a judgement rather than a computation, a gate, a route, a triage, a verdict, a severity, a merge risk, and when deciding what answers it. Esposter LLM-delegation conventions: which tier a judgement is answered at, deterministic code, a Jev typed decision or a headless session, in automation as much as in a session.
Instructions available. Your AI can read the instructions. Execution depends on the setup they require.
Account requirements not reviewed. Check the skill instructions before use; ahel provides instructions and does not run this skill.
Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
Then ask your AI: use the LLM Delegation skill
What this skill tells your AI
The instructions your AI receives, as published by esposter/esposter in .agents/skills/llm-delegation/SKILL.md and read by ahel’s review.
Which intelligence answers a question. The role split — what the main session thinks about versus what a subagent executes — is the model-delegation skill's, and so is the ban on naming a model version where it chooses what runs; this skill owns the tier a single judgement is answered at, in automation as much as in a session.
The scarce resource is one shared Claude Code account limit: every headless session the collector spawns is taken out of the same window a person's session draws on (apps/web/content/docs/infra/review-collector/drain.md). A decision moved down a tier is not a saving on a bill, it is a window a session can still run in.
Settled — do not re-propose
- Jev for an authoring step. It returns typed decisions, never text, so a conflict resolved, a fix written, a commit split or a file edited is a session's work however mechanical it looks. Jev decides whether and which; it never produces the artefact.
- A frontier text model as the cheap classification tier — a small model prompted to "reply with one word". It samples text, so it can emit a value outside the set, and it then owes a parse, a validator and a retry path; a typed decision cannot leave its own domain and needs none of them.
The tiers
| Tier | Answers | Shape |
|---|---|---|
| Deterministic code | a fact the tree, git, the filesystem or an API already states | free, exact, testable |
| A Jev typed decision | a semantic judgement over state already in hand, whose answer is one of a fixed set | sub-second, fractions of a cent, cannot answer off-set |
| A headless session | work that must open files it cannot be handed, or must write | a slice of the one shared window |
A question escalates; it never descends
Ask the lowest tier first, and let it hand up what it cannot answer. A tier that could have answered and was skipped is the whole failure mode this skill exists for — its tell is a session prompt restating facts the caller already computed, or a session whose entire output is one word the caller then parses.
- A predicate over data already read is code. A count, a trailer, an sha ancestry, a file count, a label, a threshold: no model.
- A judgement over prose or over a diff — is this concern real, how severe is this finding, which label does this issue take — is Jev's, when its answer is one of a fixed set and the state fits in the request.
- Only what must open files it cannot be handed, or must write, is a session.
A deterministic tier does not have to know whether it applies
The usual reason a mechanical answer is skipped is that nothing can tell in advance whether it fits, so the question goes up to be classified and the tier above spends its whole turn deciding which command to run. Where the repository already holds a verifier for the result, invert it: produce the mechanical answer, then ask the verifier. The classification never happens, and the roster of cases it would have needed is never written and never drifts.
Two conditions. The verifier has to be cheap against the tier above it — a check suite against a slice of the window is the trade the collector's repair makes (apps/web/content/docs/infra/review-collector/repair.md). And a failed attempt has to restore exactly the state the tier above would have found, or the cheap path has made the expensive one harder instead.
Jev — the typed decision tier
choice, noul and score, every question about one state in one call, confidence gating the escalation rather than the answer, and every call through scripts/src/services/jev/readAnswers.ts (references/jev.md).
A spawn's model is a property of its role
A session's model is read off SessionRoleModelMap for its role, and every role runs one family on purpose (references/sessions.md).
Every spawn owes a gate that can decline it
No session launches unconditionally: a deterministic check, a typed decision or a recorded marker stands before it, and a gate that cannot decide hands the case up (references/sessions.md).
Reference pages
references/jev.md— when a judgement goes to Jev.references/sessions.md— when automation launches a headless session.
Signals
- GitHub stars
- 23
- Forks
- 3
- Last commit
- Sep 2026
Advanced
- Item type
- skill
- Key
llm-delegation- Source
- github.com/esposter/esposter
Related picks
Skill · mattpocock
The pick for TypeScripttypescript-pro
Skill · jeffallan
The pick for TypeScriptskill-creator
Skill · anthropics
More in AI & modelswayfinder
Skill · mattpocock
More in AI & modelsalgorithmic-art
Skill · anthropics
More in AI & modelscode-review-and-quality
Skill · addyosmani
More in AI & models