⏳ The window ends end of day July 7 (US Central).
The principle: author on Fable, execute on anything
Fable 5's edge is judgment over long horizons — the design, the review, the synthesis, the test of whether a thing is actually right. That edge is exactly what you want baked into artifacts. A skill file written and stress-tested by Fable runs fine on Opus 4.8 in August. A migration with a Fable-authored test suite stays correct no matter who edits it next. A diligence template distilled from your ten best past decisions doesn't care which model fills it in. Spend window tokens on authorship; save the cheap tokens for execution.
The six durable assets
Turn your recurring workflows into tested skills
Anything you do more than weekly — report formats, review checklists, content pipelines, deploy rituals — becomes a skill/command file. Fable is unusually good at writing these AND at testing them against real examples before you trust them.
I'm building a library of reusable skills so future (cheaper) models can run my workflows with your rigor. Watch me describe [workflow]. Write it as a skill file: trigger, steps, quality bar, failure modes. Then TEST it against these 3 real past examples [paste], score its own output against what I actually shipped, and revise the skill until it matches or beats them. Deliver the final skill file only.
After July 7: the skill runs on Opus/Sonnet daily; you re-author on Fable only when a workflow changes.
Have the best orchestrator design your orchestration
Fable manages subagent fleets better than any model — so have it design the scaffolding your smaller models will run inside: task decomposition templates, verification loops, memory-file conventions, checkpoint patterns. The scaffold outlives the engine.
Design an agent scaffold for [recurring job: e.g. weekly competitive research]. Deliverables: (1) an orchestrator prompt that decomposes the job for parallel subagents, (2) a fresh-context verifier prompt that checks their work against the spec, (3) a memory-file convention (one lesson per file, one-line summary on top), (4) a dry run on this real example [paste], with the scaffold revised from what the dry run exposed. Assume the agents running this next month are weaker than you — build the guardrails they'll need.
After July 7: Sonnet 5 staffs the fleet at $2/$10 — inside a structure a frontier model designed.
Distill your best judgment into rubrics
Investors, buyers, hiring managers: your edge is pattern recognition from past calls. Fable can read a stack of your past decisions — deals, vendors, hires, projects — and extract the rubric you were unconsciously using, then harden it with the cases where you were wrong.
Here are [N] past decisions with outcomes [attach memos/notes]. Reverse-engineer the evaluation rubric I was actually using: the criteria that predicted the good outcomes, the red flags I missed on the bad ones. Deliver a one-page diligence template with scoring anchors and a "walk away if" list — written so a colleague (or a cheaper model) applying it cold reaches my quality of judgment.
After July 7: every new deal gets the template on any model — the judgment is in the document now.
Clear the backlog nobody wanted to touch — with the tests that keep it cleared
The migration, the rewrite, the dependency upgrade: window-perfect work because Fable's first-shot correctness on well-specified problems is its signature. The durable part isn't the code — it's the test suite that guards it.
Scope this [migration/refactor] end to end. Before writing implementation code, write the test suite that defines "done" — including the regression cases that would catch the failures this codebase has actually had [link history]. Then implement until the suite passes, and verify with a fresh-context review subagent against the spec. The tests are the deliverable; the code just has to pass them.
After July 7: any model (or any human) can touch the code — the Fable-authored tests hold the line.
Write the runbook while the writer is brilliant
SOPs, runbooks, onboarding docs — the boring compounding assets. Fable can read a quarter of tickets, transcripts, or shell history and produce the manual your operation never wrote.
Read [tickets/transcripts/history]. Write the operating manual this reveals we need: the recurring situations, the correct responses, the escalation triggers, the mistakes that repeat. Organize by situation, not by system. Flag every place where the source material shows us contradicting ourselves — those are decisions we owe ourselves, listed separately as an appendix.
After July 7: the manual onboards people and models alike. Update quarterly on whatever's cheap.
Bootstrap the memory your future agents will read
Anthropic's own playbook says Fable performs best with a memory system — and recommends bootstrapping it by having the model mine your past sessions for lessons. Do that mining now, with the best miner.
Reflect on our previous sessions and project history [point at logs/repo]. Use subagents to identify the core themes, corrections, and confirmed approaches. Store them as memory files — one lesson per file, one-line summary at the top, including why each mattered. Don't save what the repo already records. This memory will be read by weaker models later: write for them.
After July 7: every future session, on every model, starts smarter. (This is Recipe 05 supercharged.)
The remaining days, scheduled
| When | Do | On which model |
|---|---|---|
| Each morning | Check usage, verify last night's output, file the artifacts (commit skills/templates/manuals), note lessons in memory | Opus / Sonnet — don't spend Fable on admin |
| Each afternoon | Prep the night's brief: pick ONE asset, gather its inputs (repos, docs, examples), write the goal-reason-constraints prompt | Opus — prep is cheap-model work |
| Each evening | Launch the Fable run with checkpoints + memory instructions (Recipe 07), let it work long | Fable 5 — this is the window spend |
| July 7, evening | Last run: the biggest single job left. July 8+: execution moves to the ladder — Opus 4.8 daily, Sonnet 5 for fleets, Fable only if credits genuinely pencil | — |
Moral: the window isn't a free trial of a chatbot — it's four days of frontier authorship. Leave July 8 holding artifacts, not screenshots.