Detailed Analysis
The article outlines a practitioner's account of "harness bloat" — the accumulation of system prompts, custom instructions, skills, memory settings, and permission rules that build up around AI models like Claude and ChatGPT (referred to speculatively here as "Fable 5" and "GPT-5.6") over months of iterative use. The author describes an almost universal pattern among heavy AI users: every time a model made a mistake, they added a new instruction to prevent recurrence, never removing old rules even after they became redundant or contradictory. Running an inventory on their own Codex and Claude setups revealed 66 reusable skills and 172 instruction-related files, with a single writing task capable of pulling in an 18,000-word prompt before any actual work began. The piece frames this not as a prompting problem but as a systems-design problem, introducing "harness" as a term for everything wrapped around a model — the equivalent of a car's chassis and drive shaft relative to the engine.
The core argument matters because it reframes a common failure mode in AI usage: when a model's behavior degrades or seems inconsistent, users tend to blame the model itself rather than the accumulated instruction layer sitting between them and the model. This is especially relevant as providers like Anthropic and OpenAI increasingly route requests dynamically — silently switching underlying models mid-conversation, retiring older versions, or escalating difficult queries to more capable variants without explicit user awareness. When that routing happens, a harness built around the quirks and weaknesses of a previous model doesn't automatically adapt, and the mismatch can produce worse outputs even though the new model is nominally more capable. The author's proposed fix — building a "skill" that maps the entire harness into a visible table, noting where each instruction lives, when it loads, who owns it, and what evidence exists for its continued usefulness — treats prompt engineering as an auditable system rather than an ever-growing pile of ad hoc patches.
A key empirical finding in the piece is the tradeoff between richness and reliability: a "thicker" harness with detailed methodology, scoring rubrics, and classification schemes produced more thorough analysis but failed delivery constraints (broken JSON, exceeded word limits) in two of three trials, while a leaner harness with just the goal, facts, permission boundaries, and completion criteria succeeded in all three attempts. This suggests that as frontier models grow more capable at reasoning and following implicit intent, the marginal value of exhaustive scaffolding diminishes and can even become counterproductive, introducing failure points that stricter, more literal-minded instructions would have avoided. It's a finding with implications beyond a single hobbyist's workflow: it echoes a broader industry conversation about "context engineering" superseding traditional prompt engineering, where the discipline shifts from writing better instructions to curating leaner, more intentional systems around increasingly agentic models.
More broadly, the article reflects a maturing phase in how professional and semi-professional users relate to AI systems like Claude and ChatGPT. As these tools evolve from single-turn assistants into agent-like systems with memory, tool access, skills, and autonomous permissions, the "harness" surrounding them becomes as consequential as the model itself — arguably more so for day-to-day reliability. The piece's implicit critique of default AI behavior (users adding rules reactively, with no visibility into the cumulative system) points to a gap in tooling and user education that both Anthropic and OpenAI will likely need to address as their models become embedded in increasingly complex, rule-laden workflows. Whether through built-in harness auditing tools, better default memory management, or clearer signaling when models are swapped behind the scenes, the underlying tension the article surfaces — that better models don't automatically fix bad configurations — is likely to become a more prominent theme as AI systems grow more autonomous and stateful.
Read original article →