video youtu.be Watch ↗

Do THIS Before You Lose Access to Fable 5 — war-game the missions, keep the blueprints

TL;DR

Mark Kashef's "third move" for the last days of subscription access to Claude Fable 5 (the source states most users lose it on July 8th and then pay eye-watering API prices): don't spend the remaining tokens building every idea, and don't ask Fable for ordinary plans — ask it to war-game your hardest missions instead. A plan assumes linearity and shows only the success path; a wargame has the model "fight the mission on paper move by move," and for every move write the expected observation, the most-likely failure and the signal that reveals it, the counter-move, a fork trigger ("if you observe X, take route B"), RECON NEEDED flags for assumptions it can't settle, and terminal abort conditions — the action → reaction → counteraction loop pre-simulated before anything runs. The payoff is portability: the premium model's contingency reasoning, captured as markdown, becomes a blueprint a cheaper executor (Opus 4.8, GPT-5.5, Sonnet 5, GLM) can run "end to end without asking a single question" long after Fable is gone. The conceptual frame is borrowed from an Anthropic field guide the source attributes to Thariq (@trq212), "Finding Your Unknowns": with a model this capable you're no longer bottlenecked by raw intelligence but by the unknowns you hold as orchestrator, sorted into four boxes (known knowns / known unknowns / unknown knowns / unknown unknowns), and the wargame drags the last three into the light. The hands-on half shows a reusable prompt template, a fable-last-week/ folder (tasks/, wargames/, SUCCESS.md, LEDGER.md, ASSUMPTIONS.md), and running all ten missions in bulk via /goal and a 20-minute /loop that fans out parallel agents. All model names, prices, dates, and the field-guide attribution are the source's claims, flagged for a later grounding pass, not adjudicated here.

Concepts introduced

Wargaming (Adversarial Contingency Planning) — new, the core. Adversarial contingency planning: fight the mission on paper move by move — expected observation, most-likely failure + signal + counter-move, fork triggers, RECON NEEDED markers, abort conditions — the action→reaction→counteraction loop pre-simulated into a portable blueprint a cheaper executor can run.
Orchestrator Unknowns (Finding Your Unknowns) — new. The "Finding Your Unknowns" frame: past a capability threshold the bottleneck is the operator's unclarified unknowns, sorted into a four-box matrix; the wargame is how you surface the three hidden boxes.

Held, not dropped

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

Fable 5 access economics — the "couple days left," the July 8th loss-of-access date, "eye-watering" API prices. Time-bound, groundable; held as flagged claims (overlaps the Fable-economics theme already held under Make Fable 5 80% Cheaper (& Other Usage Cheat Codes) — five levers for spending a premium model less and How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch).
The /goal + /loop bulk-run harness — running ten missions with a 20-minute loop that fans out parallel agents ("a million tokens later"). A concrete workflow instance; folded into Wargaming (Adversarial Contingency Planning) and adjacent to Loop Engineering / Agentic Workflow Patterns rather than spun out.
The fable-last-week/ scaffold (tasks/ / wargames/ / SUCCESS.md / LEDGER.md / ASSUMPTIONS.md) — a specific folder convention; held as an artifact of the technique, not a concept.
"Tailor the wargame to a named executor" via a Claude Code Guide agent — reading a target model's docs/system card to gear the plan to it. A nice extension; held (touches Harness / Model Fit — a plan tuned to one executor's contract).
The prior "mine your old Fable conversations" video — referenced in passing as a different move; not distilled here.
Named executor models as portability targets (Opus 4.8, GPT-5.5, Sonnet 5, GLM, "open-source") — attributed observations / routing targets, held under Model-Tier Routing, not re-spun.

Key claims

A wargame differs from a plan by pre-simulating failure: every move carries its expected observation, most-likely failure + signal, counter-move, and fork triggers — a plan assumes linearity and shows only the success path. principle → Wargaming (Adversarial Contingency Planning)
Instructing the model "you are war-gaming, not executing" makes a capable model front-load contingency reasoning rather than an execution path — "a language model will know the difference." observation — a behavioral claim about current models, attributed to the source. → Wargaming (Adversarial Contingency Planning)
Bound the simulation to a chosen order of consequence (second/third/fourth), because a simulation without a defined end never terminates; the operator sets the depth. (best practice) — context: contingency planning that can recurse indefinitely; depth is a stakes judgment. → Wargaming (Adversarial Contingency Planning)
End every wargame with abort conditions — the error classes at which an unattended executor should stop (e.g. no access to a required system). (best practice) — context: handing a brief to an executor that will run without you. → Wargaming (Adversarial Contingency Planning)
The value is portability: once the premium model's contingency reasoning is written to markdown, a cheaper/open executor can run the brief confidently without live access to the premium model. principle → Wargaming (Adversarial Contingency Planning), builds on Execution Commoditization
Past a capability threshold the bottleneck on output quality is the orchestrator's unclarified unknowns, not the model's raw intelligence — "Fable is the first model bottlenecked by my ability to clarify its unknowns" (attributed to the Anthropic field guide). principle → Orchestrator Unknowns (Finding Your Unknowns)
Every unknown lives in one of four boxes (known knowns / known unknowns / unknown knowns / unknown unknowns); the prompt only fills the first, and the capable model can be used to surface the other three. (best practice) — context: working with a model strong enough that specification, not ability, is the limiter. → Orchestrator Unknowns (Finding Your Unknowns)
The source states most users lose subscription access to Fable 5 on July 8th and then face "eye-watering" API prices. observation — groundable, time-bound access/pricing claim; flagged for verification, not adjudicated here. → held
The source attributes an Anthropic field guide, "A Field Guide to Fable: Finding Your Unknowns," to Thariq (@trq212), and paraphrases Anthropic release notes that "Fable is an executor; give it a plan from Opus; don't use Fable for planning." observation — named source and released-doc claims; flagged for grounding. → Orchestrator Unknowns (Finding Your Unknowns)
Named executor models for running the blueprints — Opus 4.8, GPT-5.5, Sonnet 5, "a very sophisticated open-source model like GLM." observation — model names/versions reproduced as the source presents them; not verified here. → held

Why this is novel (and builds on / corroborates the Fable cluster)

The dominant stance is novel: the video's net-new contribution is the war-gaming pattern (Wargaming (Adversarial Contingency Planning)) — adversarial contingency planning as a distinct artifact from a plan — together with the four-box unknowns frame (Orchestrator Unknowns (Finding Your Unknowns)); neither had a page, and both spin out here.

The secondary stance is builds_on / corroborates across the existing Fable cluster, recorded in prose rather than by duplicating nodes:

builds_on Execution Commoditization — that page already states "the method you feed the model becomes a distribution channel for differentiation." A wargame is exactly that method, made portable and shippable: the differentiation moves off execution and onto a captured, transferable planning artifact.
adjacent to Model-Tier Routing — this is the offline, artifact-first cousin of the architect/ executor split: instead of consulting tiers live (plan-then-dispatch, advisor mode), you distill the expensive model's reasoning once into a brief any executor can run. Logged as a relation, not merged.
corroborates Imagination Constraint / Tacit Capability Awareness — the "your unknowns are the bottleneck, not raw intelligence" field guide is an independent, Anthropic-attributed convergence on "the human, not the model, is now the limiter," seen from the execution-quality side rather than the task-selection side. Orchestrator Unknowns (Finding Your Unknowns) records the distinct altitude.

No in-vault concept contradicts the video, so there is no contradicts tension to surface — only new nodes plus corroboration of the existing cluster.

Faithfulness note: this source presents "Fable 5," "Opus 4.8," "GPT-5.5," "Sonnet 5," "GLM," the July 8th access-loss date, the API-pricing framing, the Anthropic release-note paraphrase, and the "Finding Your Unknowns" field guide (and its author, Thariq / @trq212) as established fact. Per the distiller's lane these are recorded as the source's claims and flagged for a later grounding pass — not judged true/false, current/stale, or real/fictional here (this headless call has no external ground truth; a stale prior on a post-cutoff fact is exactly how a confident error would enter the vault). The technique itself (war-gaming, the four-box matrix) is method, attributed to the source and the field guide it cites.

Illustrated walkthrough

Video is 13:58; visual coverage is "low" (47/49 frames kept, 4% deduped). The largest un-illustrated stretches are ~60 s gaps — one across the middle of the war-game prompt walkthrough (≈390 s → 450 s) and one across the local-AI/tax use cases (≈540 s → 600 s) — and 21 of the kept frames are coverage-floor grabs on a static-looking background. The frame sampler misses text-on-solid-background changes, so absence of a sampled frame is not evidence the slide didn't change; treat illustration as incomplete even though the load-bearing slides (the unknowns matrix, the prompt template, the folder) were captured cleanly. Most of the deck is a purpose-built HTML slide deck with a tab bar (WHY WARGAME · THE UNKNOWNS · 01 Website … 10 Automation · RUN THE LIST · COMPACT VIEW).

t=00:00–00:24 (f0009, f0013) — the hook and "the third move." Talking-head open: "you have a couple days left to use Fable 5 as part of your existing subscription … after that the costs are eye-watering." Most people either panic-burn tokens or shrug; the offered third move is "not just building every single idea … not building simple plan files" but something "infinitely more potent that gives pretty much any model the ability to go through what Fable would have done."
t=01:04–01:33 (f0021) — the surgeon / executor framing. The source paraphrases Anthropic's release notes: "Fable is an amazing executor for long-range tasks; give it a plan created by something like Opus … don't use Fable for planning, it's a waste of tokens." His analogy: asking a top surgeon to draw how they'd operate instead of operating. His pivot: "the fix isn't asking for better plans, it's not asking for plans at all."
t=01:40 (f0024) — the "Finding Your Unknowns" field guide. On screen is an X/article card: "A Field Guide to Fable: Finding Your Unknowns," attributed to Thariq @trq212 ("Claude Code @anthropic, prev YC W20") with 205 comments / 1.1K reposts / 7.2K likes / 2.1M views, © 2026 X Corp. The visible prose: "Working with Claude Fable 5 keeps re-teaching me an old lesson: the map is not the territory. The map … is my prompts and skills and context … The territory is where the work needs to happen, the codebase, the real world … Fable is the first model where I find the quality of the work is bottlenecked by my ability to clarify its unknowns." A two-panel diagram labels "The map" (a clean dashed line) vs "The territory" (a jagged path with question marks), the gap between them tagged "your unknowns." (Author/existence of the field guide is attributed, flagged for grounding.)
t=01:46 (f0025) — the four-box unknowns matrix. Slide title: "Every unknown lives in one of four boxes." A 2×2 over WHAT YOU KNOW / WHAT YOU DON'T × YOU'RE AWARE / YOU'RE NOT: Known knowns ("what's already in your prompt," the only green box) · Known unknowns ("the gaps you can name") · Unknown knowns ("obvious to you, never written down") · Unknown unknowns ("what you never thought to ask"). Bottom line: "Your prompt only fills the first box. The wargame drags the other three into the light." Narration ties the unknown-knowns box to tacit knowledge you assume the model has but that lacks your life experience.
t=03:05–04:22 (coverage-floor frames; slides not sampled cleanly) — what a wargame is. Spoken: action, reaction, counteraction — "the AI makes a move, reality humbles it by throwing an error, and it takes a counteraction … this is the modern-day agentic loop." Worked example: building a new API endpoint on an existing platform — a normal plan just describes building it; a wargame says "assume at every juncture there are a couple different possible scenarios; based on your experience, how would you handle them?" so an Opus 4.8 / GPT-5.5 / GLM executor can "combine its harness with all the possible realities Fable already simulated and execute much more confidently." You must define an end — second/third/fourth-order consequences — and the operator decides how deep to war-game.
t=04:42–05:31 (coverage-floor; slides not sampled cleanly) — July 8th, two futures, and 10 missions. The left/right slide: most people on July 8th hold "a series of plans that executed ~80% of the way, then have to go back to Opus 4.8 to finish the last 20% — ironically the hardest part"; the alternative is "your hardest ideas fully simulated and war-gamed" to run at your leisure. Then a laundry list of 10 meaty missions to war-game (only #1 shown in detail; all provided via the video's link).
t=07:30 (f0036) — the war-game order prompt (the money slide, "01 Website" tab). Template body: "Then fight the mission on paper, move by move, and write it to wargames/01-website.md: — every move states its expected observation, exactly what you should see if it worked — every move carries its most likely failure, the cause it signals, and the counter-move — every fork gets a trigger, if you observe X, take route B — assumptions recon could not settle get marked RECON NEEDED with the exact check that settles it — end with abort conditions, and the verification runs the executor must perform with what pass looks like for each. Write it so the executor can run the brief end to end without asking a single question." Below, === THE MISSION BRIEF (the executor's orders, not yours) === with fill-in slots: "I'm rebuilding the marketing site for {{BUSINESS}} because the current one {{PROBLEM, e.g. looks dated and doesn't convert}}. Visitors are {{AUDIENCE}} and the one action I want is {{CTA}}. Build a complete static site in ./site. Plain HTML/CSS/JS, no frameworks … Mobile first, no horizontal scroll at 375px … When you believe you are done, verify before reporting. Open each page, exercise every link…" The prologue (spoken) is the shared header: "War-game order. You are not executing this mission, you are purely war-gaming it," plus an optional "a cheaper executor model (e.g. Opus) will run the brief — feel free to look at Anthropic's docs on it to tailor this."
t=10:30 (f0041) — a second brief, "04 Tax." Same template, different mission: "I run {{ENTITY TYPE, e.g. a Canadian corporation}} in {{JURISDICTION}} doing roughly {{REVENUE}} … Act as my accountant's analyst, not my accountant. Produce a tax strategy memo I will hand to a professional for review … Audit every number against the materials I gave you. Never estimate silently. Anything you cannot verify … gets marked unverified. This memo informs a conversation, it does not replace one." Illustrates that the same wargame scaffold spans web builds, copy, local-AI setup, tax, offers, chatbots, bug-fixing, financial models, competitor analysis.
t=12:00 (f0044) — the folder and the run. Claude Code with a fable-last-week/ tree: .claude/, tasks/, wargames/, and top-level ASSUMPTIONS.md, LEDGER.md, SUCCESS.md; the wargames/ finder window holds 01-website.md … 10-automation.md. Terminal shows /model set to Fable 5 and "run nothing that changes state," the status bar reading fable-last-week · Fable 5 · 13% · 5h: 0% · 7d: 9% · bypass permissions on, and a 135100 tokens counter. SUCCESS.md = your definition of a good wargame; LEDGER.md = where it records blockers and emits {{placeholder}} variables it needs input on. Run all ten with /goal ("every mission file in /tasks has a first-draft war game; loop through, do the recon mission for each, write a wargames file, draft all 10 before polishing any"), then a /loop every 20 minutes to refine — which "fans out a series of parallel agents … a million tokens later, a fully populated war games folder" of 10 blueprints any AI can execute.
t=12:44–13:34 (f0047, coverage-floor) — tailor to the executor, and the close. Optional level-up: tag a "Claude Code Guide agent" and say "tailor this war game to exactly how Sonnet 5 would execute it," spinning sub-agents through that model's docs / system card. Thesis restated: "taking the paradigm of planning to its natural extreme … pull out how Fable would do something and emulate it with cheaper models that hopefully won't disappear in the snap of a finger." Kit (prompts + folder structure) offered via the description.