Why Fable 5 and GPT-5.6 Need Cleaner AI Setups, Not Better Prompts

After months of piling on rules and instructions to coax better AI results from Fable 5 and GPT-5.6, the real breakthrough came not from prompts, but from cleaning up the entire system around them. The key to smarter AI isn’t more commands—it’s a leaner, clearer harness that lets the model shine.

Why Adding More AI Rules Backfires

Over time, AI users tend to pile on more and more rules to fix every hitch. This video creator did the same with Fable 5 and ChatGPT 5.6, adding countless instructions, project files, and custom prompts—collectively called the AI’s “harness.” While initially helpful, this approach ended up cluttering the system, causing worse performance rather than better.

He explains that the harness includes everything wrapped around the AI model: from custom instructions and saved prompts to permissions and memory. Each of these elements influences how the AI understands and answers queries, even before you type a single word. Yet most users never thoughtfully design this harness—it accumulates organically like barnacles on a ship, weighing the AI down.

Making the Invisible Visible: Mapping the AI Harness

To tackle this, a specialized skill was built to map the entire harness in one place. This inventory revealed a staggering 66 reusable skills and 172 instruction files, some reaching 18,000 words. While some rules genuinely protect quality and guide research, many were outdated, overlapping, or unnecessary with the new AI models’ improved capabilities.

Importantly, the map distinguished between simple text instructions and actual “locks”—hard permissions or schemas that outright block invalid actions. Recognizing this difference radically improved insight into what truly mattered for AI decision-making.

Blame the Right Layer: Model or Harness?

Testing revealed a surprising truth. A slimmer harness—giving just clear goals, facts, and permissions—helped Fable 5 complete tasks correctly 100% of the time. In contrast, a heavier harness loaded with detailed scoring and classification often failed, breaking word count or JSON formats.

This shows that the AI and its surroundings are a team. When problems arise, it’s rarely the model’s fault alone. More often, it’s the cluttered, overly complex setup causing confusion.

Six Rules for a Smarter AI Harness

1. Map before you clean: Understand your whole setup before trimming rules.
2. Blame the right layer: Identify if the model or harness causes failures.
3. One rule, one home: Avoid duplicating similar instructions across multiple files.
4. Load specialist knowledge only when needed: Timing context matters more than sheer volume.
5. Hard requirements need hard checks: Use machine-verifiable schemas for critical rules.
6. Build for the model and product: Tailor the harness to how the specific AI version and platform operate.

For example, some rules about avoiding false opinions were scattered in 15 different places, allowing contradictions to creep in. Consolidating to a single source of truth avoided this conflict. Similarly, specialist editorial guides loaded only when their context arose kept the system focused.

How Models React Differently to Harnesses

Fable 5 struggles with overly heavy instructions upfront, failing to meet output requirements when burdened with thick method explanations. More concise setups let it inspect problems and develop its own approach, loading depth gradually. GPT-5.6, on the other hand, chokes earlier under a huge harness when trying to find its method at all.

The harness acts much like the chassis and drive shaft of a car: essential support but prone to being weighed down if built carelessly. Keeping it clean and dynamic allows the AI engine to propel tasks smoothly.

Cleaning Tools and Real-World Impact

The cleaner skill not only maps but proposes logical trimming based on job relevance, ownership, and utility, with clear before-and-after summaries and diagnostic receipts. This makes harness maintenance manageable, avoiding surprises.

For product managers, this means a streamlined note system that loads project requirements precisely. Developers get a single version of the truth instead of conflicting instructions. Casual AI users avoid unintended influence from outdated corrections or logs lingering from past model versions.

Why It Matters Moving Forward

AI models will continue evolving, but the surrounding harness must keep pace with deliberate design, not random accretions. Removing redundant instructions isn’t about stripping context but about delivering the right depth at the right time. The ultimate goal is a system that makes AI easier, safer, and more predictable for everyone.

For those eager to try, the cleaner skill is available publicly to audit and refine any AI setup. This sheds light on what’s been quietly accumulating and opens the door to genuinely efficient AI collaboration.

Whether you rely on Claude, ChatGPT, Fable 5, or another model, understanding and managing your AI’s harness will be key to unlocking its true potential.