Making a $3 Model Work Like a $30 Model: The Fable Distillation

The Real Problem With Budget Models

Local models and cheap API models don't fail because they can't write code. They fail because they declare victory without checking, follow dead plans past the point of no return, and report "fixed" when they mean "tried."

These aren't capability gaps — they're discipline gaps. And discipline can be taught at inference time.

The Fable Distillation is a behavioral overlay — a set of explicit patterns injected at the start of a session — that transforms how a model works. No fine-tuning, no RLHF, no model merging. It's prompt engineering, but engineered from postmortem analysis of real agent failures rather than vibes.

Three Layers That Stack

Layer 1 — Operating Rules. Think before acting. Work from goals instead of step-lists. Gather your own context instead of asking permission to look. Verify everything before claiming it's done. Match effort to blast radius.

Layer 2 — Execution Process. The phase-by-phase machine for how real work moves: intake (understand before touching), orient (map before changing, baseline before modifying), plan (goal-path not ritual, riskiest unknown first), act (smallest change then fastest check), debug (reproduce, rank, bisect, three strikes), verify (adversarial pass), deliver (outcome first, honest state).

Layer 3 — Long-Horizon Survival. Named failure modes with countermeasures (hallucinated success, confidence laundering, zombie loops, fix stacking, goal mutation), orchestration protocols for delegated work, and opening moves specific to each task shape.

The Claims-Audit Gate: A Real Postmortem

Here's what happened when I asked a model running the overlay to merge two complex workflow files — 260 nodes and 300+ connections:

It did real reasoning: identified a design flaw in the handoff between stages, redesigned it using a data-dependency pattern, wrote good code. The intake and orient phases clearly worked.

Then at the end it reported: "306 links — all valid, no broken references."

That was true in a narrow sense — it verified every link's endpoints referenced existing nodes. But the serialisation format stores each connection in three places, and it only checked one. When I loaded the workflow, nearly every connection in the second stage was silently missing.

The honest report would have been: "306 links resolve per endpoint validation; full graph consistency is inferred, not verified."

So I added a harness-side gate — a stop-hook that fires when the model's closing message makes totality claims ("all valid", "no errors") and forces it to either name the exact check behind each claim or downgrade it to inferred. That's how the distillation evolves: real failures, root cause analysis, specific countermeasures. Not "be more careful."

"There Are Free Fable Packs on GitHub"

There are — and some are good. With Fable 5 leaving included plan usage on July 7, everyone with a Claude subscription spent the free window asking it to write down its own habits, and the results are floating around GitHub as skill packs. If Opus 4.8 is your daily driver, grab one of the evidence-backed ones. Opus is strong enough that six markdown files of discipline habits genuinely move the needle.

Local models are a different problem, and it's the problem this pack exists for. A 30B-class model — or DeepSeek, GLM, Qwen over API — will follow discipline prose beautifully for the first hour and then quietly stop. We've watched it happen in session logs: the instructions are still in context, the model still complies at the detail level, and the summary it writes at the end is fabricated anyway. Prose discipline decays over long sessions on distilled models. That's not a prompt-quality issue; it's how attention over a 200-message transcript works.

So this pack is built differently in two ways. It's iterated against the failure modes local models actually have — the enforcement mechanisms were built from logged failures of DeepSeek, GLM, and Qwen running real multi-hour tasks, and the pack ships those postmortems so you can see which failure produced which countermeasure. And it doesn't rely on the model remembering. The pack ships harness enforcement: a stop-hook that fires when the model tries to finish its turn, mechanically checking whether it modified files without ever executing anything, and whether its closing message makes totality claims ("all valid", "everything works") it can't back with a named check. Failed checks bounce the turn back. That check can't decay, because it isn't in the prompt.

Markdown skills can't do that second part. It's the difference between telling the model to be honest and making dishonesty fail to compile.

What It Doesn't Do

It doesn't make a small model smarter. A model that can't write valid Python won't start writing valid Python because you told it to verify. What changes is what happens around the code: whether it checks its work, whether it notices when its approach stopped working three attempts ago, whether it reports "done" or "done, per this specific check."

In practice, the gap between a capable-but-undisciplined model and a capable-and-disciplined one is bigger than the gap between model sizes. And on evidence: we're not going to claim blind-graded benchmarks we haven't run. What this pack ships is the postmortem trail — the real failures it was built from and the specific countermeasure each one produced. Published evals are on the roadmap; until then, the receipts are the postmortems.

Getting the Full Distillation

If you want the complete system — all three layers, the harness enforcement hooks, the failure-mode countermeasures, and the platform-specific integration guides — I've packaged it as a one-time download. Three formats: a Claude Code skill, a Continue/Cursor rules file, and a raw system prompt for any wrapper.

Get the Fable Distillation — $14 AUD

And if you're running local models or budget APIs for real work and the "it said it's done but nothing works" problem is eating your time, I want to hear what you're running and where it fails — [email protected].