How to Actually Prompt Fable 5, Straight From the Team That Built It

CyrilXBT · tweet · 2026-07-01 Reader 🖥️

Agent Harness Agent Memory Context Engineering Frontier-Lab Hype Long-Running & Managed Agents Multi-Agent Orchestration Prompting as Communication Riding Model Improvements

This thread is essentially Anthropic's official Fable 5 prompting guide, compressed and repackaged for practitioners as a six-step migration checklist — it restates the guide's core techniques (latency/timeouts, effort tiers, progress-auditing, proactivity constraints, memory, subagents, final-response format, refusal fallback) nearly verbatim while dressing a public-doc summary in 'straight from the team that built it' authority framing and layering in a few uncited specifics (the ultrathink/ultracode keywords, a sub-5% refusal rate, an 'audit beat softer phrasings' test) that the official extraction does not contain.

practical · expository thread (practitioner distillation of a vendor how-to) with a light persuasive/authority frame — A social-media thread that repackages Anthropic's official Fable 5 prompting guide into an actionable migration checklist. It asserts rather than argues; rationale is thin and inherited from the guide. Its one genuinely argumentative move is rhetorical, not evidential: the title's appeal to authority ('straight from the team that built it').

Structure

Part	Role
Hook and re-orientation ('Fable 5 is a different category of tool')	Warns against running Opus 4.8-era prompts on Fable 5 and frames the whole thread as a breakdown of Anthropic's published guide; asserts most teams undersell the model by testing it on the wrong things.
What Fable 5 is for + the latency change	Restates that Fable 5's advantage is sustained, autonomous, multi-stage work, and that its most surprising behavioral change is long response latency (correct deliberation, not a bug), forcing timeout updates.
Effort control (tiers + keyword invocation)	Presents effort as a cost/latency trade-off (high default, xhigh for first-shot correctness, medium/low routine) and adds thread-only keyword controls: 'ultrathink' for single-turn xhigh, '/effort ultracode' for session-level xhigh + workflow orchestration; notes summarized-only thinking and thinking.display values.
The copy-paste system-prompt blocks (progress audit, proactivity, memory, final response, subagents)	Reproduces the guide's verbatim instruction blocks for anti-fabrication auditing, constraining unrequested actions, a Markdown memory system, outcome-first final responses, and bounded subagent delegation.
Safety classifiers / refusal handling and vision prompting	Restates the refusal stop_reason + Opus 4.8 fallback (adding a sub-5% figure and Mythos/cyber-routing detail), and expands the guide's terse vision note into concrete crop/zoom and design-vs-screenshot comparison guidance.
The migration checklist and closing positioning	Synthesizes the preceding into a six-area checklist (timeouts, refusal, proactivity, progress verification, memory, final-response), sets 'first-attempt completion of hard work' as the benchmark, and positions Fable 5 as a specialist with Opus 4.8 cheaper/sufficient for everything else.

Key terms

ultrathink
A prompt keyword the thread says signals xhigh reasoning effort for a single response, without changing session settings or triggering workflow orchestration.
ultracode (/effort ultracode)
A Claude Code session setting the thread says combines xhigh reasoning with automatic dynamic workflow orchestration for every substantive task, and that requires an xhigh-capable model (Fable 5, Opus 4.8, Opus 4.7; not Sonnet 4.6 or earlier).
effort
Reframed by the thread as 'not a quality dial you crank up' but a cost-and-latency trade-off calibrated to the task. This drops the official guide's framing of effort as the intelligence/latency/cost control and omits its over-deliberation warning.
thinking.display
An API setting the thread says controls thinking-block content: 'summarized' returns a readable summary, 'omitted' (default) returns empty thinking fields; raw chain of thought is never returned for Fable 5 / Mythos 5.
audit (audit-language)
The load-bearing verb in the progress-verification block; the thread claims Anthropic tested multiple phrasings and found audit-language produced more reliable self-checking than softer equivalents.

Propositions

explicitFable 5 is a different category of tool, not a smarter version of Opus 4.8: its advantage is sustained, autonomous, multi-stage end-to-end work that takes a person hours, days, or weeks; testing it on simple, already-solved workloads undersells it, so evaluate it on your hardest previously-abandoned problems.conf: high
supportsPrompting Claude Fable 5 · p1 — Near-verbatim restatement of 'point Fable 5 at your hardest problems; simple-workload testing undersells it.' Secondary-source echo of the same official guide — adds no independent evidential weight.
explicitThe biggest behavioral change migrating from Opus 4.8 is response latency — minutes at high effort, hours on autonomous runs — which is the model planning, self-checking, and expanding context correctly, not a bug; so timeout settings almost certainly need extending.conf: high
supportsPrompting Claude Fable 5 · p10 — Restates the async/timeout migration point (runs last minutes to hours; extend timeouts). Secondary-source echo — adds emphasis ('most common source of confusion') but no independent evidential weight.
explicitEffort is a cost-and-latency trade-off you calibrate to the task (high default, xhigh when first-shot correctness matters, medium/low for routine subtasks) — 'not a quality dial you crank up for better answers.'conf: high
supportsPrompting Claude Fable 5 · p5 — Restates effort as a cost/latency trade-off with the same tiers, but FLATTENS the docs: it drops effort's intelligence axis and omits the over-deliberation warning. Secondary-source echo; the flattening is a selective distortion, not a contradiction.
explicitMaximum reasoning can be invoked by keyword: include 'ultrathink' in a prompt for single-turn xhigh; set '/effort ultracode' in Claude Code for session-level xhigh plus automatic dynamic workflow orchestration; ultracode requires an xhigh-capable model (Fable 5, Opus 4.8, Opus 4.7 — not Sonnet 4.6 or earlier).conf: high
explicitThe raw chain of thought is never returned for Fable 5 and Mythos 5; the thinking.display setting controls thinking-block content ('summarized' returns a readable summary, 'omitted' is the default and returns empty thinking fields).conf: high
explicitOn long autonomous runs, adding an instruction to audit each progress claim against an actual session tool result before reporting nearly eliminates fabricated 'I did it' status reports; this belongs in the system prompt, not just the task prompt.conf: high
supportsPrompting Claude Fable 5 · p8 — Reproduces the progress-audit block near-verbatim. Secondary-source echo — adds no independent evidential weight; the thread's extra 'audit beat softer phrasings' test is uncited and not in the primary.
explicitFable 5 is noticeably more proactive and will take unrequested actions (e.g. drafting an unrequested email, creating a defensive git-branch backup); constrain it in every unattended system prompt with an explicit block stating that an assessment is a complete deliverable — report findings and stop — and that state-changing commands must be evidence-gated.conf: high
supportsPrompting Claude Fable 5 · p9 — Same proactivity constraint block and the same two examples (unrequested email, defensive git-branch backup). Secondary-source echo — adds no independent evidential weight.
explicitcorroboratedEquip Fable 5 with a persistent Markdown memory system (one lesson per file with a one-line summary; record corrections and confirmed approaches and why they mattered; do not duplicate what the repo/chat already holds; merge duplicates; delete notes proven wrong) plus an end-of-session subagent reflection — this is the mechanism behind coherence across multi-day tasks.conf: high
supportsPrompting Claude Fable 5 · p12 — Reproduces the Markdown memory-note format and the session-close reflection bootstrap. Secondary-source echo — adds no independent evidential weight.
supports ←20 Loop Design Patterns Every AI Engineer Should Know · p9 — Success Pattern (store successful approaches and what made them work, not just failures) is instantiated by the Fable-5 persistent-memory practice of recording 'confirmed approaches and why they mattered' alongside corrections -- the same experience-memory mechanism as a concrete engineering instruction.
explicitAdd a final-response constraint to any long-running workflow (state the outcome first, then key supporting details; no working abbreviations, internal labels, or arrow chains in user-facing output) so the end user gets a clean answer rather than an internal technical dump.conf: high
supportsPrompting Claude Fable 5 · p16 — Restates the final-response readability addendum (outcome first; no arrow chains/internal labels). Secondary-source echo — adds no independent evidential weight.
explicitFable 5 can spawn and coordinate its own subagents but needs explicit delegation permission, a bounded scope with explicit success criteria per subagent, an instruction to keep working on independent parts while subagents run, and a rule to report a subagent's failure clearly in the synthesis rather than inferring what it would have found.conf: high
supportsPrompting Claude Fable 5 · p11 — Restates bounded subagent delegation with async continuation and failure-reporting. Secondary-source echo — adds no independent evidential weight.
explicitFable 5's safety classifiers (offensive cybersecurity, biology/life sciences, extraction of summarized thinking) return stop_reason 'refusal' as an HTTP 200 (not an error), so error handling must check stop_reason separately from HTTP status; configure fallback to Opus 4.8 (SDK middleware provided). Refusals affect under ~5% of typical developer queries; Mythos 5 removes the cyber classifiers (Glasswing partners only), and for everyone else cyber-adjacent work routes to Opus 4.8 at Opus pricing.conf: high
supportsPrompting Claude Fable 5 · p18 — Restates the refusal stop_reason + Opus 4.8 fallback. Secondary-source echo of the same mechanism — the added sub-5% figure and cyber-routing detail are uncited and not in the primary, so they add no evidential weight.
explicitFable 5's vision is meaningfully upgraded: it reads dense technical images and raw, noisy, or blurry screenshots accurately (often with fewer output tokens than Opus 4.8), is trained to use bash and crop tools on degraded inputs rather than interpret them directly, and can compare a design mockup against a live screenshot to identify and close a UI gap — so instructing it to 'crop and zoom' unlocks that preprocessing behavior.conf: high
explicitMigrating an Opus 4.8 workflow to Fable 5 is not just swapping model names; audit six areas before production — timeouts, refusal handling, proactivity constraints, progress verification, memory structure, final-response format — and validate on at least one task that was genuinely hard or impossible for Opus 4.8. First-attempt completion of that hardest workflow is the real benchmark of whether the migration was worth doing.conf: high
supportsPrompting Claude Fable 5 · p2 — Operationalizes the guide's 'a capability jump is a prompt to re-evaluate instructions, tools, and guardrails' into an ordered six-area migration checklist. Same content, repackaged as procedure; the packaging is the thread's additive value, not new evidence.
explicitFable 5 is not a replacement for Opus 4.8 across all use cases — it is a specialist for sustained autonomy, first-shot correctness on complex tasks, and long-context coherence; for everything else Opus 4.8 is faster, cheaper, and sufficient.conf: high
refinesPrompting Claude Fable 5 · p5 — Adds a cost-routing caveat the docs de-emphasize: Fable 5 is a specialist and Opus 4.8 is faster/cheaper/sufficient for routine work — more cost-cautious than the guide's 'lower effort on Fable 5 still performs well and often exceeds prior models' xhigh' framing.
implicitThe guidance can be trusted and applied to the reader's own workflows because it comes 'straight from the team that built' Fable 5.conf: high
implicitAnthropic's official guide is accurate and current, so its version-specific claims (latency, effort defaults, classifier domains, thinking.display values, model-support lists) can be relayed as settled fact.conf: high
implicitThe recommended prompt blocks reliably produce their described effects (the audit block eliminates fabrication; the proactivity block stops unrequested actions).conf: high

Arguments

Adopt this prompting guidance (and re-tool your Fable 5 workflows accordingly). strength: low

Explicit premises:

This is 'the full breakdown of what [Anthropic's official] guide actually says.'
It is 'straight from the team that built' Fable 5.
The teams seeing the best results follow exactly these practices.

Implicit premises:

Guidance originating from (or attributed to) the model's builders is authoritative and needs no independent verification.
An anonymous account's compression of a public guide inherits that guide's authority intact, with nothing added or distorted.
'Straight from the team that built it' denotes privileged insider access rather than a reading of the published documentation.

Inferences:

From builder-origin, infer trustworthiness; from trustworthiness, infer adopt-without-checking.

Do not shorten timeouts or treat long Fable 5 runtimes as inefficiency; leave them to run. strength: medium