| name | fable-advisor |
|---|---|
| description | Consult a stronger Fable 5 advisor model for strategic guidance at a decision point — before committing to a non-obvious approach, when stuck on a recurring problem, when weighing a change of approach, or before declaring a substantial task done. Spawns a Fable subagent (a focused brief, or a read-only investigation) and returns weighted advice. FOR OPUS-MAIN SESSIONS ONLY (Opus orchestrates, Sonnet does workers, Fable advises). Do NOT invoke when the main model is anything other than Opus — Fable advising Fable is pure waste, and this advisor is configured to sit above an Opus orchestrator. |
Fable Advisor
Approximates the Claude API "advisor tool" (which is API-only and NOT available inside Claude Code) by spawning a Fable 5 subagent as a second opinion at genuine decision points. The working split: Opus orchestrates, Sonnet does the volume worker subagents, Fable advises on the hard calls.
Precondition — Opus-main only
Only invoke this when the current session's main model is Opus. If the main model is Fable, there is nothing stronger to consult and Fable→Fable is pure token waste — do NOT use it. If the main model is Sonnet or Haiku, this skill is not the configured path (the intended setup runs Opus as the orchestrator, Sonnet as workers, Fable as advisor). Check the actual current model before consulting; if you are not on Opus, skip.
When to consult (the decision points)
- Before substantive or irreversible work — before writing/editing files, committing to an interpretation, or building on an assumption. Orientation (reading files, searching, fetching a source) is NOT substantive — do that first, then consult so Fable sees real context.
- When stuck — errors recurring, an approach not converging, results that don't fit.
- When considering a change of approach.
- Before declaring a substantial task done — and make the deliverable durable FIRST (write the file, save, commit): the consult takes time; a durable result survives an interruption, an unwritten one doesn't.
- For design / architecture / risk judgment calls where your answer would be analysis or a recommendation with no other tool call — that's where a second opinion is highest-value.
Restraint — this is a rationing tool
Fable is the priciest tier and drains the shared usage pool fastest; reserving it is the whole point. So:
- Reserve consults for genuine decision points on non-trivial tasks. One consult before committing to an approach and one before declaring done is the norm for a multi-step task.
- Skip on short reactive turns where the next action is dictated by tool output you just read, on trivial/mechanical edits, and on simple factual or arithmetic questions.
- Do NOT consult every turn. If you're unsure a decision is "high-value," it probably isn't.
How to consult — pick the mode per call
Spawn a subagent with model: "fable" (this is the explicit exception to the
default-Sonnet subagent rule). The subagent does NOT see this conversation — it starts fresh
with only the prompt you give it. Two modes:
1. Focused brief (default — design, architecture, risk, "which approach", judgment calls). Package the context into the prompt yourself:
- the task/goal and any hard constraints,
- the approach(es) under consideration,
- what you've already tried and the key results/evidence,
- the specific question you want answered.
End with: "Keep your guidance focused — a sharp starting point, not a comprehensive plan
(aim ~150 words)." Use a read-only agent type — general-purpose with an explicit
"advise only, do not modify anything" instruction, or Plan (read-only architect).
2. Read-tools investigation (code review, "check my work", decisions that hinge on details
across many files). Spawn Fable with a read-only agent (Explore, or general-purpose
instructed not to edit) plus pointers to the relevant files/dirs and the question. Let it read
the code itself and advise. More Fable tokens — use when hand-packaging context would lose the
detail the decision depends on.
Either way, the subagent's final message returns as the tool result — that's the advice.
How to weight the advice (from the advisor tool's own best-practice guidance)
- Give it serious weight — it's a stronger model with a fresh perspective.
- Adapt only on empirical failure (you followed a step and it failed) or primary-source contradiction (the file says X, the paper says Y against a specific claim). A passing self-test is NOT evidence the advice is wrong — it may mean your test doesn't check what the advice checks.
- Don't silently switch away from evidence you already have. If you'd gathered data pointing one way and Fable points another, reconcile in ONE more focused Fable call: "I found X, you suggest Y — which constraint breaks the tie?" A reconcile call is cheaper than committing to the wrong branch.
After consulting
Incorporate the advice, then briefly tell the user what Fable advised and how it changed or confirmed the plan — don't bury a course-correction.