Source: Every Prompt You Send Drags 18,384 Words Of Junk. Here’s How I Cut It.
Type: YouTube Video Transcript
Author: nate-b-jones
Video ID: PDJfciNhyHU
Source URL: https://www.youtube.com/watch?v=PDJfciNhyHU
Summary
In this video, Nate B Jones addresses the critical issue of harness bloat—the accumulation of redundant instructions, rules, and files wrapped around AI models over time. He defines an AI harness as everything wrapped around a model (custom instructions, project files, saved prompts, memories, skills, tools, permissions, and checks) that shapes the model’s output before a user even enters a prompt.
Nate shares his personal audit, which revealed 66 reusable skills, 172 instruction-related files, and over 27,000 characters of description, causing a single writing job to drag in an 18,000-word context file. To combat this, he introduces the Six Principles of Stable Harness Design and demonstrates how a cleaner skill can map and optimize these systems. He also analyzes how frontier models like fable-5 and chatgpt-5-6 fail differently when overloaded with heavy harnesses.
Key Takeaways
- What is an AI Harness?: The complete wrapper around an AI model, including custom instructions, project files, saved prompts, memory, skills, tools, permissions, and checks. It acts like a car’s chassis and drive shaft, transferring the model’s raw engine power to the actual work.
- The Problem of Harness Bloat: Users accidentally build bloated harnesses by adding “one-off” rules and instructions every time an AI makes a mistake. Over time, these rules overlap, drift, and degrade model performance.
- Six Principles of Stable Harness Design:
- Map the harness before you clean it: Inventory every control, its location, load trigger, purpose, owner, and efficacy.
- Blame the right layer: Differentiate between model failures and harness failures. Over-prompting to solve prompt-induced issues leads to a vicious cycle of bloat.
- One rule, one home, one owner: Consolidate duplicate rules (e.g., authorship or source guidelines) to prevent version drift.
- Load specialist knowledge when needed: Use lazy or selective loading to bring in deep context (like editorial guides) only during relevant workflow phases.
- Hard requirements need hard checks: Shift rigid constraints (like word counts or JSON schemas) from prompt prose to machine-verifiable schemas and system locks.
- Build for the model and product doing the work: Optimize harnesses for specific model lineages (fable-5 vs chatgpt-5-6) and the specific product interface (e.g., Claude.ai vs Claude Code vs API).
- Model-Specific Failure Modes:
- Fable 5 (Claude): Fails late in the delivery phase when overloaded with heavy methods, struggling to meet output constraints (like JSON or word limits) despite high-quality reasoning. It thrives on compact setups with clear goals, boundaries, and room to plan.
- ChatGPT 5.6 (Codex): Fails early in the routing phase, getting overwhelmed by a massive harness layer and failing to find the correct execution path.