Fable Foreman turns your session into a job site.
A free, open-source skill for Claude Code: your strongest model becomes the foreman — it plans, routes, and reviews — while cheaper workers swing the hammers. And nothing gets called done until an independent verifier, who never saw how the work was built, reproduces the evidence.
Version 0.4.0 · Updated August 18, 2026 · MIT licensed
You bought access to the smartest AI model on the planet — and then you asked it to rename variables for two hours.
That's like hiring a world-class architect and handing them a hammer. The expensive part of a frontier model is its judgment: what to build, what to check, what's actually wrong. The typing? A model a tenth the price does that just fine.
Fable Foreman flips that around. It's the management playbook nobody handed the boss — so your best model spends its time on the decisions only it can make, and the grunt work goes to workers who cost a fraction as much.
Free and MIT licensed. The repo is the home of the skill — grab it on GitHub, read the seventeen small files in about fifteen minutes, and install it.
View the skill on GitHub: https://github.com/olsenbrands/fable-foreman Release + Desktop ZIP: https://github.com/olsenbrands/fable-foreman/releases/tag/v0.1.0
Copy this and paste it into Claude Code. That's the whole install — Claude fetches the skill from GitHub and puts it where it belongs, along with its five agents. Works the same on Mac and Windows.
Get Fable Foreman
On Claude Desktop / claude.ai: Download the ZIP from the release page (https://github.com/olsenbrands/fable-foreman/releases/tag/v0.1.0) and upload it under Settings → Customize → Skills. Desktop runs a lighter "discipline mode" — the skill tells you plainly what it can and can't do there.
You've felt both of these
If you've run Claude Code hard, you know the two failure modes by heart: "You hit your usage limit at 2pm because your frontier model spent the morning doing find-and-replace." "You told a swarm of agents to 'just go' and watched an unsupervised budget evaporate on work nobody checked." The problem was never agents. It was that nobody gave the boss a management playbook.
What it actually does
Fable Foreman is instructions, not software — a skill your Claude reads and follows. No daemons, no config files, no hooks. Seventeen plain-text files. Here's the system it installs in your AI's head:
1. It probes the job site first. What model is running? Can it spawn agents? Is OpenAI's Codex CLI installed and logged in — and xAI's Grok CLI? Foreman detects what your machine actually offers and adapts, from full orchestration down to an honest discipline mode on Claude Desktop that tells you exactly what it can't do there.
2. It routes by judgment, then by evidence. Tasks are classed FRONTIER, WORKHORSE, or FAST by their judgment content — a 500-line mechanical rename is FAST work; a 10-line concurrency fix is FRONTIER work. Only then does price enter: the class resolves to a seat using a dated table of prices, context ceilings, and capability scores, not a hunch. "Economics chooses among the models that clear the quality bar. It never lowers the bar." — the First Law, verbatim from the skill
3. It speaks all three providers — and routes to what's actually there. Claude only, Claude plus Codex, Claude plus Grok, or all three: any combination works. Codex and Grok are discovered from your machine at runtime and chosen per task, and Codex's first dispatch asks for your OK because it's your money. A missing provider is a routing input, never a stall — Foreman never holds work waiting for one, and never tells you to go install one mid-run.
4. Zero hardcoded models. New Claude next month? New Codex tier? The rules never name dated models — they resolve live. It was built the week GPT-5.6 launched and survived that churn during its own review. The foreman seat is a capability class rather than a specific model, so any frontier-class Claude runs it identically — and if your session switches models mid-run, it re-checks its own seat instead of routing off a stale one.
5. It trusts no one — including itself. Every meaningful change gets a blind verifier: fresh context, read-only tools, handed your original request (never the worker's retelling of it), and assuming the work is broken until it reproduces the evidence. Off-family reviewers sharpen the read, but the accepting verdict is always the Claude verifier.
6. It respects your budget in both directions. Bounded retries (never a third identical attempt), announced fan-outs, one fix-worker per findings list instead of a swarm — and a degradation rule with a spine: if budget pressure would drop quality, it stops and tells you instead of quietly shipping worse work.
What's new in v0.4.0 — know what a seat costs
v0.3.0 made the crew visible. v0.4.0 makes the choice of crew checkable. Foreman now routes off a dated evidence table instead of a hunch — price per seat, the cache discount, the context ceiling, and a capability score, in one file you can read and argue with.
The seats, priced and scored
Prices and context windows from the vendors' own docs, verified 2026-08-17. Index = Artificial Analysis Intelligence Index v4.1.1, one consistent source so the column compares like with like.
The honest caveat: these are list prices — on a subscription login they are not money spent, only the best available proxy for how fast a dispatch depletes that provider's allowance — and the index is a composite, not a coding benchmark, so a two-point gap is noise rather than a ranking.
Grok 4.6 is the near-frontier value seat. Same index band as Opus 5 and sol (61 against 63 and 61) at roughly 44% of Opus 5's depletion. Same tier, far cheaper — that, not "Grok is cheap," is the reason to field it. gpt-5.6-luna dominates Haiku 4.5 outright — cheaper and far more capable (52 against 30). Where Codex is present, luna is the FAST seat and Haiku is the fallback. Sol and Fable are priced above their measured capability. Field them for a specific strength, for cross-family independence, or because their pool is the unloaded one — never by default.
Which seat gets which job
The honest caveat: the judgment content of a task still picks the class first, and the First Law is not suspended here — if no listed seat clears the bar, Foreman goes up a tier or stops rather than taking the cheap row because it is cheap.
The effort rule: Raising a model's reasoning effort never changes its per-token price — it changes how many tokens get generated. The payoff depends on the task shape: on analysis, knowledge work, and review the accuracy curve is nearly flat (low gives up only 1–3 points for 33–50% less cost), so Foreman defaults to medium. On long-horizon coding the curve is steep (low gives up about 8 points), so it defaults to high and does not economize there. The honest caveat: that curve is Anthropic's published measurement on Anthropic models — a strong prior for the other families, not a proven transfer.
Cross 200,000 tokens and Grok reprices the entire request at 2x, so cost jumps rather than climbs. At 210K in / 15K out, Sonnet 5 is 44% cheaper on list price ($0.570 against $1.020) for the same job. So the rule is a routing boundary, not a price bump: above roughly 200K, change seats. Above 500K Grok is ineligible outright — that is a hard context ceiling. The honest caveat: the within-Grok 2x jump is xAI's published policy; the cross-provider comparison is list-price math to confirm against measured costs.
Every number above is dated, and the skill's own currency rule says the live provider docs win when they disagree — re-derive the table rather than route off a stale one. You can read the whole thing, sources and all, in references/model-matrix.md: https://github.com/olsenbrands/fable-foreman/blob/main/skills/fable-foreman/references/model-matrix.md
What's new in v0.4.0
Still in there from v0.3.0
That release's theme was "trust the log, see the crew." It's history now, so here it is in short:
1. You can watch Codex work. Codex jobs used to be invisible shell commands. Each one now rides its own visible subagent in the Claude Code harness, so you see the crew working and completion arrives as a notification instead of a status bar to babysit.
2. It stopped taking any model's word for it. Which model actually served a dispatch is graded by evidence in three tiers — served, routed, requested — never by the worker's self-report. With no hard evidence, the ledger says "seat: unverified" instead of pretending.
3. Three conventions became scripts. The environment probe, the ledger bootstrap, and the Codex launcher stopped being English a model could misread under pressure and became small, injection-hardened shell scripts.
4. Setup that proves itself. An agent-executable runbook verifies your environment step by step and shows the evidence, rather than asserting you're all set.
About the cost — the honest version
Anthropic has published that multi-agent systems can use roughly 15x the tokens of a single chat (https://www.anthropic.com/engineering/multi-agent-research-system), and its own cost guidance (https://code.claude.com/docs/en/costs) recommends putting cheaper-tier workers under a stronger lead. Fable Foreman is built around exactly that.
So this is about smarter spend, not a discount. It won't promise to cut your bill by some percentage — subscription users all share one quota — but it does stop your most expensive model from doing your cheapest work, and it refuses to quietly burn budget on a runaway loop.
Built in public — and stress-tested by a rival AI
So the skill that teaches Claude to distrust AI self-reports was itself stress-tested by a rival AI until it stopped breaking. That's a stress test, not an endorsement — and the commit history on GitHub shows every round.
This skill wasn't vibe-coded in an afternoon. The research phase alone ran about 66 AI agents across three passes — reading every serious orchestration project on GitHub, the official Anthropic and OpenAI docs, and the failure post-mortems — before a line was written. Around 5.4M tokens, with every repo claim verified against the GitHub API.
Then the first draft went in front of the harshest reviewer available: GPT-5.6 "Sol," OpenAI's flagship model, released that same week, running in a read-only sandbox with instructions to tear the draft apart. Across five adversarial review rounds the findings fell 21 → 12 → 10 → 7 → 2 → fixed. Sol caught things like a one-word bug where "commit or stash" would have made the verifier validate the wrong version of the code.
v0.3.0 got the same treatment. The work started by studying the claudemix project, then went through three more rounds of adversarial review, this time from OpenAI's frontier Codex model, plus live head-to-head testing against v0.2.0. Sixteen findings came back; every one was fixed or explicitly disclosed. The skill's own verification discipline was used to build the version that sharpened that discipline — reviewed by its own rival again.
v0.4.0 went further. It was hardened by three adversarial reviews from Grok itself — the very model the release was adding. Grok returned four blockers on the first pass; three were conceded and the plan was cut back. Then came a blind Claude verification, and then an independent Fable review that found what all four passes had missed: a billing policy that had been swept in without being asked for, a percentage printed backwards in the cost-cliff table, a metric bug in the rule that reacts to that cliff, and a transport wrapper that fabricated a relay. Every one of those was fixed and re-verified before this version was published.
It's seventeen files. Read them yourself.
No daemons, no telemetry, no magic. Fable Foreman is SKILL.md, seven reference docs, four short shell scripts, and five role agents — scout, worker, verifier, and one transport wrapper each for Codex and Grok — plus the manifests that register it. All plain text, MIT licensed, and it passes Anthropic's claude plugin validate. Read the whole system in about fifteen minutes, then decide — which is rather the point of a skill built on verifying instead of trusting. Read the source on GitHub: https://github.com/olsenbrands/fable-foreman
Who it's for
→ Claude Code users on Pro or Max who watch their usage bar like a gas gauge.
→ Anyone running big multi-file tasks who's been burned by unverified agent output.
→ People with both a Claude and a ChatGPT/Codex subscription who want them working as one team.
→ People with a Grok subscription — Grok 4.6 gets used as a cheap adversarial reviewer and as a workhorse implementer under 200K.
→ Claude Desktop users — you get a lighter but honest version that's upfront about its limits.
Frequently asked questions
Make your best model the boss
Grab Fable Foreman on GitHub — free, MIT licensed, and small enough to read in about fifteen minutes.
Don't sleep on what's next. Free tools, prompts, and skills for Claude Code and OpenClaw — sent when they drop. No spam, unsubscribe anytime.
Your frontier model is worth more than grunt work. Put it in charge.
Don't sleep on AI.
