video youtube.com Watch ↗

How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch

TL;DR

Nate Herk (AI Automation) reads Anthropic's own Prompting Claude Fable 5 doc and distills it into six paste-able habits under one through-line the source labels "the shift — smarter model, lighter touch": because Fable 5 follows short, clear direction and reasons better than older models, piling on rules and telling it how to think now backfires, so the leverage moves to (1) give it the why — supply intent/context, not step-by-step detail; (2) tell it what NOT to do — negative-prompt the boundaries ("do not fix/send/edit until I say go"); (3) let it act once it has enough — stop over-planning, and match reasoning effort (low → x-high) to the task's value rather than defaulting Fable 5 to everything; (4) make it prove it — the source's claimed "most important" habit — require evidence for any "done" instead of trusting the claim; (5) — the one Fable-specific rule — stop asking it to "show its reasoning," because a standing "explain your reasoning" line (especially in a system prompt/skill) can trip a safety classifier and silently reroute the task to a weaker backup model; and (6) say less, not more — a short instruction now steers as well as an over-detailed rulebook, which can actively make the answer worse. A closing explainer covers the safety-classifier handoff: requests that look like hacking, dangerous biology, or that ask the model to reveal its private reasoning silently route to Opus 4.8 — cheaper, capable, losing only Fable's edge on long/hard tasks. The source states many groundable specifics (pricing $10/$50 per M input/output, a promo→usage-credit access window, a FrontierCode accuracy-vs-cost chart, "Mythos 5" as a restricted higher tier) — these are attributed to the source and flagged for later verification, not adjudicated here.

Concepts introduced

Intent Context — Rule 1: give the model the why (intent, audience, the larger task) so it connects the request to the right information instead of guessing; "context over detail."
Negative Prompting — Rule 2: explicitly state what not to do / where to stop, because a capable model acts readily and will "get creative" past your ask.
Reasoning Effort Control — Rule 3 (effort half): treat reasoning effort (low/medium/high/x-high) as a dial matched to task value and cost, not a fixed high default.
Evidence-Gated Completion — Rule 4: require the model to point at evidence for any "done" and to say plainly when something is unverified; a claim of completion is not a fact.
Safety Routing Fallback — Rule 5 + the closing explainer: a safety classifier can silently reroute a "sensitive-looking" request (incl. asking the model to narrate its private reasoning) to a lower-tier backup model.
Concise Prompting — Rule 6: for a capable model wrapped in good context/tools, one clear instruction steers as well as an exhaustive rulebook — and an over-detailed prompt can degrade output.

Held, not dropped

Themes the capture touches that do not warrant their own concept page yet (spin out on demand):

Fable 5 economics — pricing ($10/$50 per M in/out), the promo→usage-credit access window, and the internal June 22/23 vs "July 7th" date inconsistency. Groundable, time-bound; held as a flagged claim, not a durable concept.
"Mythos 5" as a restricted higher tier — named model gated to Glasswing/biology partners with safeguards lifted. Held (attributed, for grounding).
FrontierCode benchmark — the specific accuracy-vs-cost eval and its numbers. Held as a source artifact, not a concept.
The "intern" mental model for prompting — "tell the intern what not to do; they don't know the process yet." A framing already echoed under Agent Supervision ("AI is like an intern"); held as framing.
"Act when you have enough" / stop over-planning (Rule 3, planning half) — not a new node; it corroborates Agentic Simplicity and the loop's own-plan-mode idea in Agent Loop. Logged there, not duplicated.
Second brain / AIOS as the setup that supplies context files — passing mention; see existing AI Second Brain. Held.
Sponsor segment (Hyper Agent "council of agents") — advertisement; not distilled.

Key claims

Because Fable 5 follows short, clear direction and reasons better, over-prescriptive prompting ("piling on rules and telling it how to think") can degrade output — the winning move is a lighter touch. observation — the source's central thesis, attributed to Anthropic's doc; corroborates Agentic Simplicity at the prompt level. → Concise Prompting
Give the model the why (intent/context) rather than step-by-step detail; "Anthropic says Fable does better when it understands your intent." (best practice) — context: capable instruction-following models where intent lets the model route to the right information. → Intent Context
Negative-prompt the boundaries — state what not to do and where to stop ("report and stop; don't fix/send/edit/delete until I say go"). (best practice) — context: a more-capable model that will otherwise act/expand past the ask; the source notes negative prompting works better on newer models than it used to. → Negative Prompting
Let it act once it has enough; stop forcing exhaustive up-front planning. (best practice) — context: models that can run for minutes at higher effort, where option-surveying burns budget on unused choices. → Reasoning Effort Control, Agentic Simplicity
Match reasoning effort to the task: the source relays Anthropic's recommendation to use high as the default, x-high for the most capability-sensitive workloads, and medium/low for routine work. (best practice) — context: cost-sensitive use of an expensive model; "using Fable 5 for everything is ~100% overkill." → Reasoning Effort Control
"Make it prove it": require evidence for completion and honest "not verified" over guessing — the source claims Anthropic's own testing found this one line "nearly eliminated made-up status reports." (best practice) — context: agentic/long-running work where models over-claim done; bake into skills/agents/CLAUDE.md rather than every prompt. → Evidence-Gated Completion
(Fable-specific) A standing "explain/show your reasoning" instruction — especially in a system prompt or skill — can trip a safety classifier and silently reroute the task to a lower-tier backup (Opus 4.8) with no on-screen notice; ask for the answer and read the thinking output instead of making it narrate. (best practice) — context: Fable 5's summarized-thinking-extraction safeguard; Fable-specific, may not generalize to other models. → Safety Routing Fallback
For a capable model in a good environment (context, tools, skills), a short instruction steers as well as a spelled-out rulebook — and reusing an over-detailed prompt can make the answer worse. observation — reconciled with Rule 1: keep the why, cut the redundant procedure. → Concise Prompting
The safety handoff routes hacking / dangerous-biology / reveal-your-reasoning requests to Opus 4.8; it is silent in the app/Claude Code but marked on the raw API, and costs less (Opus ≈ half of Fable), losing only Fable's edge on long/hard tasks. observation — a mechanism claim about routing behavior and pricing. → Safety Routing Fallback
Fable 5 is priced at $10 / M input and $50 / M output ("double Opus"); a promo window includes it on subscription plans before usage-credit billing. observation — groundable pricing/availability claim; the source's own on-screen doc ("June 22/23") and narration ("ends July 7th") disagree on dates — flag for verification. → held
On the source's "FrontierCode" accuracy-vs-cost chart, Fable 5 scales from ~11–12% (low) to ~31% (max), Opus 4.8 ~8–13%, GPT-5.5 ~5%, on the hardest 50 of 150 tasks; the source claims "Fable 5 on low ≈ Opus 4.8 on x-high, but cheaper." observation — benchmark numbers reproduced as the source presents them; not verified here. → held

Why this is novel (and corroborates Agentic Simplicity)

The dominant stance is novel: five of the six habits attach to no existing concept page and are spun out as new nodes (Intent Context, Negative Prompting, Reasoning Effort Control, Evidence-Gated Completion, Safety Routing Fallback, Concise Prompting). The secondary stance is corroborates: the through-line — "smarter model, lighter touch; piling on rules backfires; act when you have enough" — is an independent, doc-derived convergence on Agentic Simplicity applied at the prompt level (that concept already records this source's "say less, not more" and "act when enough" claims). Rule 4 ("make it prove it") is adjacent to Agent Supervision (grounding an agent's own progress claims) and Eval-Driven Development (don't trust vibes; show evidence), but is distinct enough — it is the model self-gating its completion claims, not the human reviewing output — to earn its own node, cross-linked rather than merged.

Faithfulness note: this source presents "Claude Fable 5" (and "Mythos 5", "Opus 4.8", "GPT-5.5", the prices, dates, and benchmark figures) as established fact. Per the distiller's lane, these are recorded as the source's claims and flagged for a later grounding pass — not adjudicated true or false here (this call has no external ground truth). The one in-vault tension surfaced is the source's internal date inconsistency (on-screen "June 22/23" vs spoken "July 7th"), written as a tension to resolve, not a verdict.

Illustrated walkthrough

Video is 10:44; visual coverage is "ok" (89/98 frames kept, 9% deduped). The largest un-illustrated stretch is ~72 s spanning the back half of Rule 5 into Rule 6 (≈462 s → 534 s) — the Rule 5 card (f0063) anchors the start of that window and the Rule 6 card (f0064) its end, but any mid-explanation slide changes between them were not sampled; absence of a frame there is not evidence of a static screen. Most of the deck is a purpose-built recreation of the Anthropic doc rather than the live doc, except two frames (f0007, f0037) that show doc-styled pages directly.

t=00:03 (f0007) — the "Availability" doc page. On screen: "Claude Fable 5 is available everywhere today. Claude Mythos 5 is restricted to Glasswing partners (with cyber safeguards lifted) and soon to select biology researchers… Pricing for both models is $10 per million input tokens and $50 per million output tokens. Developers can use claude-fable-5 via the Claude API." Plan windows read "from today through June 22 … included at no extra cost. On June 23 we'll remove Fable 5 … using it after that will require usage credits." Spoken narration instead says the promo "ends July 7th" — a source-internal date inconsistency worth grounding.
t=02:25 (f0034) — the through-line slide, "How to prompt Claude Fable 5 · from Anthropic's own docs". On screen: "Fable 5 follows short, clear direction better than older models. … The model got smart enough that piling on rules and telling it how to think backfires. Give it the why, tell it what not to touch, and make it show the receipt." Below: "THE SHIFT — Smarter model, lighter touch. Older models needed long, detailed rulebooks. Fable 5 follows one clear instruction on its own, so piling on rules now gets in its way. The habits that used to help can now hurt." This is the spine every rule hangs off.
t=03:57 (f0037) — the "Capability improvements" doc page. Lists Fable 5 gains compared with Claude Opus 4.8: long-horizon autonomy, first-shot correctness, vision, enterprise workflows, code review/debugging ("bug-finding recall … noticeably higher than Claude Opus 4.8"), navigating ambiguity, delegation. A callout box: "Claude Fable 5 runs safety classifiers that target offensive cybersecurity techniques …, biology and life sciences content …, and extraction of the model's summarized thinking. Benign … work may also trigger these safeguards. To re-route declined requests automatically, configure server-side or client-side fallback to Claude Opus 4.8." and "requests in those domains can return stop_reason: "refusal"." This page is the doc basis for Rules 5 and the handoff.
t=04:17 (f0041) — the six-card summary, "Six habits, six sentences". Sub-line: "Fable 5 is powerful and not cheap, so a sharper prompt is also a smaller bill." Cards are tagged any model vs Fable-specific ("five are solid prompting anywhere, only one is Fable-specific"). Card 1 Give it the why (CONTEXT OVER DETAIL) — instead of "Write me an email to a client about the delay," paste "I'm working on [the bigger task] for [who it's for]. They need [what the output enables]. With that in mind: [your request]." Card 2 Tell it what NOT to do (NEGATIVE PROMPTING) — instead of "Take a look at this problem and handle it," paste "When I'm describing a problem or asking a question, the deliverable is your assessment. Report what you find and stop. Don't fix, send, edit, or delete anything until I say go. Do the simplest thing that works, and skip cleanup I didn't ask for."
t=05:43 (f0059) — Cards 3 & 4. Card 3 Let it act once it has enough (STOP IT OVER-PLANNING) — "More deliberation is not better. On a model that can run for minutes, endless option-surveying just burns time and money on choices it will never use." Paste: "When you have enough information to act, act. Don't re-derive what we've already settled or narrate options you won't pursue. If you're weighing a choice, give a recommendation, not an exhaustive survey." Card 4 Make it prove it (CATCH A FAKE "DONE") — "'Done and working' is a claim, not a fact. Anthropic's own testing found this one line nearly eliminated made-up status reports, even on tasks built to provoke them." Paste: "Before you tell me something is done, point to the result that proves it. Only report work you can show evidence for. If something isn't verified, say so plainly instead of guessing."
t=06:03 (f0061) — the "FrontierCode: Accuracy vs Cost" chart. Three series: Claude Fable 5 (red), Claude Opus 4.8 (green), GPT-5.5 (black dashed), plotting score % against mean cost/task (log scale). Fable 5 climbs from ~11–12% at low (~$5) to ~31% at max (~$19); Opus 4.8 ranges ~8% (low) to ~13% (x-high); GPT-5.5 sits flat near ~5%. Caption: "Diamond subset comprises the hardest 50 of 150 tasks." Narration's claim off this chart: "Fable 5 on low … is similar to Opus 4.8 on x-high and max, [but] is cheaper" — motivating effort-matching and "realistically only reach for Fable ~5–15% of the time." (All figures attributed to the source's chart, flagged for grounding.)
t=07:42 (f0063 / f0064) — Cards 5 & 6. Card 5 Stop asking it to "show its reasoning" (THE ONE THAT FLIPS A RULE · Fable-specific) — "On Fable 5, a standing 'explain your reasoning' line, especially saved in a system prompt or skill, can trigger a refusal and hand your task to a less capable backup model (Opus 4.8) with no error on screen. You get a weaker answer without knowing it switched." Delete "Explain your reasoning step by step, then give me the answer"; do instead "Just ask for the answer and let it think privately. If you need to see its thinking, read the thinking output rather than making it narrate." Card 6 Say less, not more (ONE LINE BEATS TWELVE) — "A short instruction now steers as well as spelling out most of the rules by name. Reusing last year's over-detailed prompt can actively make the answer worse." Explicit reconciliation: "Not a contradiction with tip 1. Add the why (real signal the model can't guess), and cut the rulebook (redundant procedure it already follows)." Paste: "Lead with the outcome, keep it simple, and pause only when the work truly needs me."
t=09:22 (f0081) — "What happens when Fable hands off to Opus 4.8". A 3-step flow — You ask Fable 5 → It trips a safety check (the request looks sensitive) → Opus 4.8 answers (a capable, cheaper backup) — with four Q&A tiles: When? hacking, dangerous biology, or asking the model to reveal its private reasoning (ordinary coding/security can occasionally trip too). Do you see it? silent in the Claude app and Claude Code; on the raw API the response is clearly marked. What do you pay? whichever model does the work; Opus 4.8 is ~half of Fable, so a handoff costs less, and a request declined before it writes anything is free. How to avoid? don't tell it to show/explain its thinking, keep security questions plainly worded, and use Opus 4.8 by default for genuine cyber/biology work. Banner: "A handoff isn't a surprise bill or a failure … the only thing you give up is Fable's extra muscle on the long, hard tasks." Footer: "Source: Anthropic docs · Prompting Claude Fable 5 and Refusals and fallback."