AI Agent field notes

Don’t ask for perfection

Status: searched 2026-07-31. The over-engineering effect is documented; the causal-versus-correlational framing and the convert-or-delete rule are the deltas.

Claim: asking a model for perfection makes the work worse. “Highest quality”, “impeccable”, “extreme performance” produce over-engineered solutions and ornate prose — because quality words are causal for a human and correlational for a model. A human routes “perfection” through taste. The model routes it through association, and in its training data the vocabulary of maximal quality co-occurs with maximal ceremony.

The asymmetry

Ask a senior human for their highest-quality work and you get their judgment applied harder: often simpler code, fewer moving parts, the right thing stripped of the unnecessary — because their model of quality is causal, a theory of what makes software good. Ask a model and the quality words shift its output toward what “highest quality” sounds like: more abstraction layers, more defensive machinery, more configuration, more patterns, more superlatives in the prose. It performs thoroughness. The request selects the aesthetic of quality, not the substance — intuition trained on humans transfers exactly wrong.

The deeper reason is that “perfection” is a symlink to a target that exists only in your head — the full specification of what good means for this project. You can dereference that pointer; the model cannot. Hand it the symlink and it resolves to the generic file its training says “perfection” points at: ceremony, layers, ornament. So you cannot just tell it what you want — you have to pour the requirements out. The thing the word points to never leaves your skull unless you write it down.

The incident

My harness’s memory profile rendered my engineering bar as superlatives — “perfection”, “impeccable”, “impenetrable”, “extreme” — and loaded them into every session for weeks (the model’s model of you). The measurable result was not higher quality. It was over-engineering and ornate prose: speculative hardening that became an attack surface, invented machinery, dense writing. The evidence is written into my instruction file as its entire corrective half: “quality is subtractive here — corrections remove code”, “more machinery is not more quality”, and a scar list of deleted model-built systems — a 64KB “graceful” drain buffer that was a CPU-abuse vector, a hand-rolled rate limiter, a whole subsystem nobody asked for. Every one of those was the model being maximal at me, as requested.

The rule

Every quality adjective converts to a constraint, a budget, or a test — or it gets deleted:

  • “Impeccable security” → validate inputs, fail closed, tear down on malformed. Decidable.
  • “Extreme performance” → zero allocations on the measured hot path; simplest correct thing everywhere else. Measurable, and bounded — the second half matters as much as the first.
  • “Perfect reliability” → the specific failure modes handled, and a deliberately-absent list for the ones accepted. Enumerable.
  • The test for any remaining quality word: can a session tell whether it complied? If not, the word is register, and register is what gets mirrored (the dirty house) — not obeyed.

The deliberately-absent list deserves special mention: writing down what the project intentionally does NOT do (“no clustering, no video, no speculative hardening”) turns out to restrain a model far better than any superlative motivates it — it is the one place where quality can be stated as a boundary instead of an aspiration.

Prior art

Verdict: PARTIAL. The effect is documented and the fix has vendor convergence. Max Woolf’s “write better code” experiment (2025) — this page’s existing neighbor, now verified — shows Claude 3.5 spontaneously going “enterprise,” adding cutting-edge optimizations and redundant machinery when iteratively asked to improve. OpenAI’s GPT-5 prompting guide (2025) independently prescribes the fix: convert a superlative (“world-class”) into a self-authored rubric before executing, and it cites Cursor’s production finding that unconstrained quality instructions produced worse output, fixed by an enumerated constraint rather than a better adjective. The persona-prompting line (arXiv:2512.05858) confirms expert personas don’t reliably help; verbosity/length bias in the LLM-as-judge literature is the mechanism-level reason quality and ceremony co-occur in training data. What no source states: the causal-for-humans / correlational-for-models framing of quality words, the generalization to standing-instruction superlatives baked into a memory or config file, and the convert-or-delete rule with the deliberately-absent list as the one place quality can be a boundary. Those are the deltas.