How to Combine Fable 5 and Soul 5.6 in One Claude Code Skill
How to Combine Fable 5 and Soul 5.6 in One Claude Code Skill
Stop asking which model is better. The real move is running Fable 5 and Soul 5.6 (GPT 5.6) together in a single workflow — Fable plans, Codex builds, Fable reviews. I built a Claude Code skill that does exactly this, and in a live build it shipped a working web app while burning only about 130,000 tokens on the Fable side. Here's how the four-stage pipeline works and how to set it up yourself.
Everyone wants to know if Soul 5.6 beats Claude Fable. That's the wrong question. The models are good at different things, and the price gap between them means you're leaving efficiency on the table if you force one model to do everything. The smarter play is a division of labor: use the best planner to plan, the cheapest capable builder to build, and the best reviewer to review.
Why Combine Fable 5 and Soul 5.6 Instead of Picking One?
There are two reasons to run these models together instead of choosing a side: raw capability and token efficiency.
First, capability. Soul 5.6 — OpenAI's GPT 5.6, shipping in Codex — posts strong numbers on Terminal Bench 2.1. Both the standard 5.6 and 5.6 Ultra variants land ahead of Claude Mythos, let alone Fable 5, on that benchmark. Take it with a grain of salt since the numbers come from OpenAI, but the trend is real.
Second, and more importantly, token efficiency. There's a reason so much content right now is about reducing Claude usage — things like advisor mode, where you have Fable plan and Opus execute. But if Opus is doing the execution at a given price, why not have a cheaper, more efficient model do it instead?
The data backs this up. GPT 5.5 on extra-high effort scored 23% pass rate at $1.24 on the benchmark. GPT 5.6 scored 25% — a higher score at lower cost. Each generation has gotten leaner: 5.6 is more token efficient than 5.5, which was more token efficient than Opus 4.6. In direct comparisons of 5.5 versus Opus 4.8, it's not close — higher pass rates, lower cost.
So the construct is the same one people already use with advisor mode. You're just swapping in a better and arguably cheaper builder than Opus.
What Are the Two Skills This Workflow Uses?
There are two skills here, and they serve different needs.
- codex-build — the standalone version. You've already created a plan with Fable, and Codex just goes and builds that feature or product. Use this when the planning is already done.
- grill-me-codex — the comprehensive four-stage skill. This is the big one. It runs the whole loop from interview to review, and it's an updated version of a skill I've covered before, now with GPT 5.6 building things for you.
If you just need something built from an existing plan, reach for codex-build. If you're starting a feature from scratch and want the full adversarial planning treatment, use grill-me-codex.
How Does the grill-me-codex Skill Work?
The skill runs in four stages: a deep interview, adversarial planning between Fable and Codex, the Codex build, and a final Fable review. Here's each stage.
1. The Interview (Grill Me)
You kick it off with a prompt describing what you want to build. The first thing that happens is an interview — this is literally the Grill Me skill from Matt Pocock. It's plan mode on steroids. It goes far deeper than Claude Code normally would, asking eight to ten pointed questions about your project.
In the demo, I built "Trip Atlas," a stylized cinematic trip-planner web app. The skill asked things like what the tool is for — I said a real personal tool, not just a video demo — and about geocoding choices. For every question, it gives its own recommendation, so if you don't know what to pick or why, that's spelled out for you. Fable drives this whole stage.
2. Adversarial Planning
Once Fable has a plan, the skill pushes that plan over to Codex. In the demo that's GPT 5.5; going forward it's 5.6. Fable and Codex go back and forth for up to five iterations. Fable presents the plan, Codex responds with what to accept and what to change, Fable agrees or disagrees, and they keep going until they reach consensus.
The skill logs the entire back-and-forth in a markdown file so you can read exactly how the two models argued. In my run it only took two rounds to get approval. The second round surfaced 12 findings around hardening the data core — the kind of thing a single model working alone tends to miss.
3. The Codex Build
Once the two models agree, the actual build gets pushed to Codex — 5.5 today, 5.6 going forward. This is where the upgrade pays off. It's better than passing the build to Opus or Sonnet or using advisor mode inside Claude Code, because these GPT models are stronger than the smaller Anthropic models and they're cheaper.
At this point you actually have options. Codex can build it. Claude can build it if you'd rather keep everything in-house. Or you can stop at the plan. In the demo I let Codex build.
4. The Fable Review
After Codex finishes, Fable reviews everything Codex built and goes back to Codex with corrections — "this is wrong, this was right." It runs up to two iterations of that. If the build still isn't right by the third pass, Fable takes the wheel and writes the code itself.
In my run, Fable found a couple of deviations from the plan, judged them all reasonable, walked the files, and then asked whether I wanted to commit or take a look first.
What Did the Skill Actually Build?
The Trip Atlas app worked. It had a world map with custom graphics that Codex generated using the GPT image generator. You could name a trip, add stops, and list what you'd do at each location. There was a cinematic replay with a plane hopping from spot to spot as an SVG, complete with passport stamps.
The interactions were real, not decorative. Delete a stop and it's gone. Reorder stops up and down. Add Tokyo and it calculates and shows how far that stop is once you add it to the route. For a first pass built through a fully automated model handoff, everything actually worked.
And the headline number: the whole thing cost about 130,000 tokens on the Fable side. That's the entire point — you get the planning and review quality of Fable with the build cost pushed onto a cheaper, more efficient model.
How Do You Install the Skill?
Installation is straightforward. The GitHub link is in the video description. Once installed, you invoke it with /grill-codex and hand it your prompt describing what you want to build. From there the four stages run on their own, pausing only for the interview answers and your final commit decision.
Frequently Asked Questions
Is Soul 5.6 actually better than Claude Fable 5?
On Terminal Bench 2.1, Soul 5.6 and 5.6 Ultra post higher numbers than Claude Mythos and Fable 5 — but those benchmarks come from OpenAI, so treat them with skepticism. The more defensible advantage is token efficiency: 5.6 scores higher than 5.5 at lower cost, and both beat Opus on price-per-pass. The point of the skill isn't to crown a winner; it's to use each model where it's strongest.
Why not just use advisor mode with Opus?
Advisor mode has Fable plan and Opus execute. The problem is Opus is expensive and less token efficient than GPT 5.6. You're paying more for a weaker builder. This skill keeps the same plan-then-execute construct but swaps Opus out for a cheaper, more capable executor.
What's the difference between codex-build and grill-me-codex?
codex-build is standalone — you already have a Fable plan and Codex just builds it. grill-me-codex is the full four-stage pipeline: interview, adversarial planning, Codex build, and Fable review. Use codex-build when planning is done; use grill-me-codex when you're starting from scratch.
How many tokens does this actually save?
In the live demo, the full build of a working trip-planner web app cost roughly 130,000 tokens on the Fable side, because the heavy execution work was offloaded to Codex. The exact savings depend on your project, but the structural win is that expensive Fable tokens go toward planning and review rather than line-by-line building.
Does this require GPT 5.6 specifically?
No. The workflow runs today on GPT 5.5 through Codex and gets better when 5.6 lands. The skill is model-agnostic on the Codex side — it uses whatever Codex is currently serving.
If you want to go deeper into combining Claude and Codex in your own builds, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.
