# AI Quality Gaps Are Policy Problems Not Capability Problems Most of what reads as "the model isn't good enough" is actually the model applying a skill it already has at the wrong moment — a policy gap you can close today with a better harness, not a capability gap you wait out for a better model. The diagnosis comes from design, where an LLM "can write CSS fluently but has essentially zero design taste." The capability (fluent CSS) is present; what's missing is *when* to deploy it. Slop isn't random — it's the model "predicting the most likely next token," and "the average landing page is mediocre by definition," so the default output is the corpus's mode. You escape the default by engineering the policy: force purpose before layout (the real bug is "the model chooses layout before it chooses the purpose"), make quality countable instead of subjective (state coverage — "the median of AI-generated component is like 1.5 states" vs. a human's six), and choose representations where parameter-space distance tracks output-space distance (OKLCH over HSL) so the model can optimize the thing you actually care about. None of these waits for a smarter model; all of them are harness design you do now. ## Scope & conditions - Holds when the model demonstrably *can* do the sub-skill in isolation but produces mediocre defaults in context — design, writing, code style, accessibility markup, schema choices. - Doesn't hold for genuine capability ceilings (a task the model fails even when scoped tightly and prompted perfectly) — there, harness work won't save you and a better model is the real lever. - The test for which you're facing: scope the task to one sub-skill and force the non-default choice. If quality jumps, it was policy; if it doesn't, it was capability. ## Links - [[AI's Failure Mode Is Careless Adoption Not the Technology]] — extends: organizational supervision and labeling are adoption-policy choices that can create failure without any change in model capability. - [[3. Resources/Permanent Notes/Bounded Action Spaces Not Better Models Make Agents Safe to Run]] — the safety sibling of the same "not better models" family: safety comes from constraining the envelope, quality comes from constraining the policy — both engineer the harness instead of waiting on the model - [[3. Resources/Permanent Notes/Vague Prompts Are Expensive Specificity Cuts AI Error and Cost Together]] — the prompt-level version of the same lever: specificity (a policy choice) closes the gap cheaper than scale - [[3. Resources/Permanent Notes/Encoding What Good Looks Like Is What Makes Agent Verification Trustworthy]] — the verification counterpart: you have to encode the quality bar (state coverage, purpose-first) for the model to hit it or be checked against it <!-- atomicity: 1 idea ✓ --> - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the MOC lists this note under Amplifier thesis & the harness.