Claude Fable 5 is BACK (How we're USING IT BETTER) — Superbash Learn

Superbash Learn

Click to load the YouTube player

Reviewed companion

Useful notes, receipts, and next steps

Format: tutorial
Reviewed: Jul 18, 2026

TL;DR

Ron used Fable 5 on extra high for the first one to three prompts, then stopped once the product requirements document (PRD)—a brief defining what to build and why—specifications, and foundation were in place. (source video jw9nhsWB4Ck, 01:29; source video jw9nhsWB4Ck, 01:40)
He then handed execution to GLM 5.2 or a model transcribed in the source as “Kimmy 2.7 code.” This is a dated July 2026 workflow, not a claim about which models are currently available or best. (source video jw9nhsWB4Ck, 01:52)
His example was an unfinished indicator rebuild in Pine Script V6, the language used for Ron’s indicator, with context from 68 videos. Ron said Fable 5 was the only model he had played with that could one-shot a working indicator, but he did not present a controlled comparison in the transcript. (source video jw9nhsWB4Ck, 00:44; source video jw9nhsWB4Ck, 02:28; source video jw9nhsWB4Ck, 02:37)
Prompt for the outcome and the reason behind it, not a numbered implementation recipe. Ron said this reversed the detailed, step-driven style he had used with Opus 4.8. (source video jw9nhsWB4Ck, 03:46; source video jw9nhsWB4Ck, 04:02; source video jw9nhsWB4Ck, 04:13)
On July 2, Ron said Fable 5 counted toward up to 50% of weekly usage limits through July 7 and would move to usage credits afterward. July 7 had passed by this companion’s July 18 update; this article does not confirm current access or pricing. (source video jw9nhsWB4Ck, 00:13)

Ron’s verdict

Use Fable 5 where expensive reasoning can change the whole project: the initial plan, a product requirements document, a specification, a difficult refactor, or a new foundation. Keep that burst to one to three prompts, then move repetitive work to another capable model. Ron’s test gives a useful routing pattern, not a blank cheque for autonomy: his Pine Script example was still in progress, and his earlier indicator had conditions he was unhappy with. Verify the artifact before calling the hand-off complete. (source video jw9nhsWB4Ck, 01:29; source video jw9nhsWB4Ck, 01:52; source video jw9nhsWB4Ck, 02:55; source video jw9nhsWB4Ck, 03:11)

Key moments

00:44 — The 68-video Pine Script job: Ron rebuilds a CBR entry indicator with much more video context than his first attempt.
01:29 — The one-to-three-prompt rule: use extra high for planning and foundation work, then stop.
02:28 — Why Pine Script gets the premium model: Ron describes a language where his other model attempts struggled to produce a working indicator.
03:42 — Goal-driven prompting: state the outcome and why it matters, then let the model form the plan.
04:30 — Overnight autonomous runs: Ron discusses reports of 12-hour Claude Code sessions returning 80–90% complete.
04:52 — Intelligence versus tighter safeguards: Ron gives his interpretation of the model’s changed behavioral experience.

Useful quotes

“I think when it comes to a different type of programming language where it’s not Python, extra high might be worth it.” — Ron, source video jw9nhsWB4Ck, 02:22

“stop writing numbered prompts.” — Ron, source video jw9nhsWB4Ck, 04:00

“the intelligence is the same, but the behavioral experience is tighter.” — Ron, source video jw9nhsWB4Ck, 05:17

Put Fable 5 at the foundation, not everywhere

The core idea is a tiered model workflow: reserve the premium model for decisions that shape everything downstream, then route bulk execution elsewhere. Ron named PRDs, specifications, code refactors, regression finding, skill.md improvements, and memory-system reconstruction as foundation work his group was compressing into one to three Fable 5 prompts. (source video jw9nhsWB4Ck, 01:40; source video jw9nhsWB4Ck, 02:03)

Stage	Fable 5’s job in Ron’s workflow	Hand-off rule
Plan	Work in plan mode at `extra high`; establish the PRD and specifications. (source video jw9nhsWB4Ck, 01:29; source video jw9nhsWB4Ck, 01:40)	Stop when you believe the hard foundation is mostly laid. Ron described this as about 90%, not a measured completion score. (source video jw9nhsWB4Ck, 01:45)
Foundation	Refactor code, find regressions, improve skills, or reconstruct memory systems. (source video jw9nhsWB4Ck, 02:03)	Keep the premium burst to roughly one to three prompts. (source video jw9nhsWB4Ck, 02:13)
Execution	Preserve the plan as the artifact another model continues from. This is a practical inference from Ron’s hand-off pattern, not an instruction he stated.	Continue with GLM 5.2 or the “Kimmy 2.7 code” model named in the transcript. (source video jw9nhsWB4Ck, 01:52)
Acceptance	Run or inspect the finished work before trusting it. This check is an inference from Ron’s example, not an instruction he stated.	Ron’s earlier indicator had incorrect conditions, so “one shot” still needs a real output check. (source video jw9nhsWB4Ck, 02:52)

Three details that change how to use the workflow

First, Ron chose extra high for Pine Script V6, the indicator language in his example, because he found it unusually hard for AI models. He supplied all 68 videos after deciding a one-video attempt may have missed important conditions. (source video jw9nhsWB4Ck, 02:20; source video jw9nhsWB4Ck, 02:28; source video jw9nhsWB4Ck, 02:52)

Second, the graphics-card advice belongs to Ron’s media pipeline. He used faster Whisper on a GPU, or graphics processing unit, to transcribe the videos and said this job would be impossible without one. That does not establish that every Fable 5 task needs a GPU. (source video jw9nhsWB4Ck, 03:13; source video jw9nhsWB4Ck, 03:19; source video jw9nhsWB4Ck, 03:24)

Third, state the outcome, explain why it matters, and let Fable 5 build the plan. Ron said his mistake was importing the numbered prompt style he used with Opus 4.8. (source video jw9nhsWB4Ck, 03:46; source video jw9nhsWB4Ck, 04:13) Ron also relayed reports from users on X of 12-hour Claude Code runs returning an estimated 80–90% complete; the transcript does not provide a controlled result. (source video jw9nhsWB4Ck, 04:34; source video jw9nhsWB4Ck, 04:46)

Keep the safeguard story inside the evidence boundary

Ron believed Fable 5’s intelligence had not been diluted. He said a retrained classifier, a system that decides how a request is handled, blocked a known jailbreak more aggressively and could silently route some coding, debugging, cybersecurity, or biology tasks to Opus 4.8. He repeated an “over 99%” block figure. (source video jw9nhsWB4Ck, 05:08; source video jw9nhsWB4Ck, 05:22; source video jw9nhsWB4Ck, 05:32; source video jw9nhsWB4Ck, 05:37)

The transcript does not provide model telemetry, classifier documentation, or a controlled before-and-after test. Ron says his group had used the returned model for about a day and did not sense weaker intelligence. Treat both the routing explanation and the no-dilution conclusion as Ron’s early July 2026 assessment, not independently established product behavior. (source video jw9nhsWB4Ck, 06:04; source video jw9nhsWB4Ck, 06:06)

A five-question routing check

Will this prompt define the architecture, specification, or major refactor? If yes, it fits Ron’s foundation tier. (source video jw9nhsWB4Ck, 01:40; source video jw9nhsWB4Ck, 02:03)
Have you stated the outcome and the reason? If not, rewrite before adding numbered steps. (source video jw9nhsWB4Ck, 03:46; source video jw9nhsWB4Ck, 04:02)
Is a non-Python language the hard part? Consider Ron’s extra high pattern, but treat it as his judgment from this test rather than a universal setting. (source video jw9nhsWB4Ck, 02:20)
Does the workflow need local video or graphics processing? Check the hardware pipeline separately from the model choice; Ron’s 68-video transcription depended on GPU work. (source video jw9nhsWB4Ck, 03:13; source video jw9nhsWB4Ck, 03:24)
Can another model continue from a written artifact, and can you verify the result? If yes, hand off after the foundation. Requiring verification is a practical inference from Ron’s flawed first indicator and unfinished second attempt, not an instruction he stated. (source video jw9nhsWB4Ck, 01:52; source video jw9nhsWB4Ck, 02:44; source video jw9nhsWB4Ck, 02:52)

What changed since this video

This video was published July 2, 2026. Its “through July 7th” usage-limit window had passed by this companion’s July 18, 2026 update. This article relies only on the July 2 source for its access terms, model names, extra high behavior, Sonnet 5 comparison, classifier explanation, and reported autonomous-run results. Treat them as a dated record, not confirmation of current product state. (source video jw9nhsWB4Ck, 00:13; source video jw9nhsWB4Ck, 06:24)

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=jw9nhsWB4Ck