KM4BI KM4BI

Every prompt drags 18,384 words of junk: how I cut it down

Nate B. Jones explains how to map and slim down the AI harness around a model so the right context loads at the right time with real checks.

Nate B. Jones starts from a practical discovery: his AI setup had become too heavy. Every time a model made a mistake, he added another rule. Over time, those custom instructions, project files, memories, skills, tools, permissions, and checks could pull in thousands of words before a writing task even began.

He calls this surrounding layer the “harness”: everything around the model that shapes the answer before the user writes the prompt. The point is not that all context is bad. Some rules protect important work, such as keeping research separate from the user’s own opinion. The problem is losing track of which rules still protect the work, which ones duplicate each other, which load too early, and which confuse the model.

His cleaner begins by mapping the harness. Each control needs a row that explains where it lives, when it loads, what job it performs, who owns it, what evidence shows it still helps, and what can go wrong if it is misused.

The cleanup follows six principles: map before cleaning, blame the right layer, give each rule one home and one owner, load specialist knowledge only when the work needs it, turn hard requirements into hard checks, and design for the actual model and product doing the work.

The tests in the video show why this matters. A thicker setup sometimes produced richer analysis, but it also failed delivery constraints such as JSON validity and word limits. A more compact setup respected the requirements more reliably. For both Fable 5 and ChatGPT 5.6 in Codex, the lesson is not to starve the model of context; it is to deliver depth at the right moment and let schemas, file checks, tool restrictions, and run receipts enforce what machines can verify.

The takeaway is simple: a clean harness is not an empty harness. It is a readable system with clear sources of truth, specialist skills loaded at the right phase, automated validation, and enough diagnostic evidence to understand what actually happened.

Source