"How to Get the Most Out of Fable 5" — @Laura Entis

Why this is in the vault

Practical, worked-example guidance on operating Fable 5 — the exact model this agent was upgraded to on 2026-06-09. Every's team spent a week using it in production across four real workflows, and the piece doubles as a delegation playbook: when to spend Fable 5 tokens, how to frame tasks, and how to verify output. The issue is a hybrid: highlights from Dan Shipper's AI & I episode with Mike Krieger (head of Anthropic Labs), then the team's own usage patterns with reusable prompts. Two house-promo CTAs (Fable 5 Camp on June 12, Codex Camp on June 26) open the email, but they are a brief intro, not the substance — no third-party sponsor present.

The core argument

The fastest way to be disappointed by Fable 5 is to drive it like GPT-5.5 or Opus 4.8, where you iterate back and forth with careful prompting. Fable 5 behaves like a capable coworker: front-load the context, the goal, and the definition of done, then step aside and review the finished work. Every frames this as the manager mindset — "think like a manager" — and notes it only pays off when the task deserves it.

Task-selection filter. Good Fable 5 candidates have four qualities: you can supply organized, deep context; the goal is well-defined; there is a clear definition of done; and the task's importance justifies the cost. Because the model carries tasks all the way to completion, stale data or conflicting goals send it confidently to the wrong place — there are fewer mid-task checkpoints where a human would catch the drift.

Krieger's highlights (AI & I episode):

• Overnight delegation is real: he ends his day briefing the model on what to finish while he sleeps, and it routes around obstacles (downed services, broken tools) instead of stalling.
• The head-to-world gap is closing for non-engineers — an Anthropic recruiter could finally build the exact internal tools she needed.
• Engineering shifts from writing code to setting direction, reviewing agent output, and making judgment calls; the PM/engineer divide blurs.
• Verification is where attention moves: regression tests on known workflows, visual checks (including feeding the model video captures of its own work to catch animation glitches screenshots miss), and mock backends for what can't be tested live.

Four worked examples from Every's team:

1. Fix a broken workflow (Nityesh Agarwal): pointed Fable 5 at the session log of a failing PowerPoint-deck skill. It diagnosed the root cause (agents hand-editing raw slide XML) and built a CLI tool giving agents targeted edit operations. Takeaway: use Fable 5 to diagnose failures and build the tooling, then let cheaper models run that infrastructure.
2. Go-to-market strategy (Austin Tedesco): a vague "make a plan" ask produced an expensive consensus summary. Re-framed with full sources of truth (surveys, PostHog, positioning docs), a concrete business objective, and a specific deliverable (10 insights + a stack-ranked list of 10 moves with evidence), the output jumped to hire-of-the-quarter quality. Takeaway: ask it to test assumptions against data, not summarize agreement.
3. Feedback into batched changes (Kieran Klaassen): the "AI sandwich" (human–machine–human) scaled up — Fable 5 pulled two days of colleague feedback from Slack, derived a fix list, and shipped 30 fixes in one batch with a cross-conflict check, instead of 10 serial reviews. Next layer: scheduled feedback pulls evaluated against the product vision doc and personas, surfaced for approval. Takeaway: the model is strongest wired into a feedback loop, and output quality tracks input quality.
4. Build from an original spec (Willie Williams): code-only inspection produced a confident wrong fix for a memory-leak-style bug; telling the model to run the app locally and watch it led to the real fix. Given the original product spec, Fable 5's from-scratch build beat Opus 4.8 and GPT-5.5. Takeaway: hand it what you'd hand a senior engineer, and make it verify in the environment where the thing actually runs.

Cost and speed. Fable 5 rides on Claude paid plans until June 22, then moves to token-based pricing: $10/M input, $50/M output — roughly 2x Opus 4.8 and 3x Sonnet 4.6. It is also slow at higher effort levels. Reserve it for large, delegable jobs (workflow repair, feature builds, heavy synthesis, codebase review); keep quick edits and brainstorming on faster, cheaper models.

Mapping against Ray Data Co

Strong mapping — this is operating doctrine for the agent reading it, one day after the Fable 5 upgrade.

• The four-quality task filter is a dispatch gate. "Organized context + defined goal + definition of done + cost-justified importance" is nearly isomorphic to the pre-registered ACCEPTANCE CONTRACT that /verify-dispatch already checks for. The addition worth adopting: explicitly scoring whether a task justifies Fable 5-tier effort at all, or whether it belongs to a cheaper subagent seat. With 1M context available, the temptation is to throw everything at the top model; this argues for routing by task shape.
• Diagnose-then-tool is the /improve and /skillify loop, named. Nityesh's pattern (top model reads the failure log, builds the tool, cheaper models run it forever after) is exactly the thin-harness/fat-skills economics RDCO already bets on. Concrete extension: when a pipeline seat or skill repeatedly fails, feed Fable 5 the session log and ask for root cause + infrastructure, rather than patching prompts.
• Batch-with-conflict-check beats serial review. Kieran's 30-fixes-in-one-batch pattern maps to todo-file + /loop builds: batch related fixes, have the agent verify non-interference, review once at the end. Cheaper in founder attention, which is the binding constraint here.
• Verification patterns worth stealing. Video captures of the agent's own work (vs screenshots) for catching motion glitches matches the existing /video-critic instinct; "run it where it actually runs before trusting a fix" is the verify-skill philosophy and a caution against code-read-only diagnoses. Krieger's regression-on-known-workflows suggests keeping a small suite of known-good RDCO workflows to re-run after harness changes.
• Austin's GTM reframe applies to strategic outputs. "Test assumptions against data, produce ranked decisions with evidence and flagged single-source conclusions" is a better dispatch shape for investing briefs and /verify-strategic-output targets than "synthesize a plan" — the failure mode he hit (expensive consensus summary) is the exact overconfident-synthesis pattern that skill exists to catch.
• Pricing cliff confirmed. The issue corroborates the known June 23 subscription-to-usage cliff: paid-plan access through June 22, then $10/$50 per M tokens (~2x Opus 4.8). At ~3x Sonnet 4.6, the task-routing discipline above is also a cost-control measure for this always-on agent.

The camps (Fable 5 Camp June 12, Codex Camp June 26) are paid-member live events — relevant only if the founder wants a live walkthrough; the substance of this issue is the playbook itself.

Related

• [[2026-06-09-every-vibe-check-fable-5-best-coding-model]] — Every's week-of-testing review this issue builds on
• [[2026-06-10-stratechery-fable-5-anthropic-alignment-ai-tiers]] — same-day strategic read on the Fable 5 launch and pricing tiers
• [[2026-06-03-every-opus-4-8-gets-in-your-way]] — the prior-generation contrast this issue argues against driving Fable 5 like
• [[2026-06-02-every-eight-levels-ai-adoption]] — the adoption curve referenced for who already delegates comfortably (L7-L8)
• [[2026-04-22-every-bread-in-ai-sandwich]] — Kieran's human-machine-human pattern that example 3 scales up