Ask Fable

Ask Fable 5

Turn the user's ask into a Fable-shaped task prompt and hand it to the fable subagent. Behavioral tuning (scope discipline, honest progress reporting, no early stopping, delegation, report style) already lives in the agent definition. Write the task, not the personality.

Task: $ARGUMENTS

1. Check the fit

Fable 5 is for the hard end of the range. Route elsewhere when:

• The task is a routine edit or a quick lookup — just do it in-session, or use a normal subagent. Say so rather than dispatching.
• The task touches offensive cybersecurity (exploits, malware, attack tooling) or biology and life sciences (lab methods, molecular mechanisms). Fable runs classifiers on those domains and returns stop_reason: "refusal" even for benign work. Handle it in-session or with an Opus subagent.

If the ask is vague, ask one or two clarifying questions before dispatching. Fable navigates ambiguity well, but a run this expensive should not start pointed at the wrong target.

2. Write the prompt

Include, in this order:

The why, not only the what. Highest-leverage addition. Context lets Fable connect the task to relevant information instead of inferring intent.

I'm working on [larger task] for [who it's for]. They need [what the output enables].
With that in mind: [request].

The definition of done. Concrete enough that a verifier subagent could check it against a specification.

Starting context. Paths, repos, commands to run, constraints, decisions already made — everything established in this session that Fable would otherwise re-derive from scratch.

Verification cadence, for anything long-running:

Establish a method for checking your own work at an interval of [X] as you build.
Run this every [X], verifying against the specification with subagents.

A notes path, when the work spans sessions or should accumulate lessons. Fable performs notably better with somewhere to record them.

Then dispatch, with a stable topic-scoped name so the agent stays addressable for follow-ups:

Agent(
  name: "fable-<short-topic>",     // scoped, so two invocations don't collide
  subagent_type: "fable",          // any type EXCEPT fork
  description: "<3-5 words>",
  prompt: "<the above>"
)

Never use subagent_type: "fork". Fork inherits the parent's model, so the spawn silently runs on your model instead of Fable's — you would be talking to yourself.

Front-load the prompt. A non-fork subagent inherits none of your context and there is no second chance to pick it up later. Brief it like a colleague who just walked in: background, what has already been tried, exact paths, the facts it needs. Under-briefing is the most common way to get a weak answer out of Fable.

3. Follow-ups

To continue the SAME conversation with its context intact, use SendMessage — do not call Agent again, which starts a fresh agent with no memory of the exchange:

SendMessage(to: "fable-<short-topic>", summary: "<5-10 words>", message: "<your reply>")

The name keeps working after the agent finishes; a send resumes it from its transcript. That is also why the name must be topic-scoped — reusing a bare "fable" across two different asks can resume the wrong prior conversation.

This flow needs top-level invocation. Named, addressable agents only exist when you are the main thread. If this skill is itself running inside a subagent, name is silently unavailable ("Only synchronous subagents are supported") — the spawn still works, but SendMessage cannot reach it and the conversation is lost. When nested, use single-shot Agent calls that re-paste the full context each time.

4. Autonomous runs

When the user will not be watching (overnight, background, scheduled), append:

You are operating autonomously. The user is not watching in real time and cannot answer
questions mid-task, so asking "Want me to…?" or "Shall I…?" will block the work. For
reversible actions that follow from the original request, proceed without asking.

5. Relay the result

The agent's report is not shown to the user. Lead with the outcome, then whatever it needs from them. Do not pad it with commentary.

Caveats

• Fable runs long: many minutes per turn, hours for autonomous work. Tell the user up front so a silent gap does not read as a hang.
• Never instruct Fable to echo, transcribe, or explain its internal reasoning as response text. That trips the reasoning_extraction refusal category and falls back to Opus.
• Resist over-specifying. Prompts and skills tuned for earlier models are usually too prescriptive for Fable 5 and degrade its output. State the goal and the constraints; let it pick the method.

Full prompting guide: references/prompting-fable-5.md.