Agent Taste¶
peter-steinberger's framing of the skill that survives the automation of code production. "If an AI can write the code, what is an engineer's job? Taste."
The lowest bar: AI-smell detection¶
The entry point of taste is the ability to instantly know something is AI-generated. Works across mediums:
- UI/UX — "the purple gradient" / "the color border on the left"
- Copy — too wordy, too many dots, inappropriate apologies, certain cadence
- Code — generic naming, over-abstracted, telltale boilerplate
- Architecture — wrong seams, awkward dependency directions
"Even if you can't pinpoint it, you will know. That's why I call it a smell."
This isn't a parlor trick. In an environment where 80% of output is agent-generated, AI-smell detection is the first line of quality defense.
Higher-order taste: the details agents won't produce¶
The next level: what the agent wouldn't produce unprompted, but makes the thing feel alive. Examples from OpenClaw:
- Occasional roast messages when you run the CLI
- Personality tuned for each chat platform (WhatsApp voice ≠ Slack voice ≠ Telegram voice)
- The soul-md file — hand-crafted persona, not prompt-templated
- Easter-eggs and lobster references
These emerge from prompting taste into the system design, not from asking the agent to "be fun."
Why it's the moat¶
- Not benchmarkable → not optimizable → not commoditized.
- Compounds with exposure: taste is pattern-recognition built from many examples across many domains.
- Can't be outsourced — agents can execute taste you specify, but they can't originate it (yet).
- See animals-vs-ghosts: ghosts inherit whatever's in the training data, which is the average, which is slop by construction.
Pairs with¶
- System design — Peter: "If you don't think about that, you will eventually swipe yourself into a corner." alex-krentsel independently reaches the same conclusion from the systems side: see design-over-implementation — when the code is AI-written and gross-but-working, architectural taste is what's left.
- Saying no — the ability to refuse the wild-idea-that's-a-prompt-away.
- Big-picture syncing — the agent is dropped in blind; taste provides the context the agent can't infer.
Cross-references¶
- peter-steinberger — originator of this framing at the 2026 talk
- vibe-coding — Karpathy's adjacent frame (taste as what raises the floor)
- agentic-engineering — taste as what differentiates professional from vibe
- driving-into-mud — failure mode when taste checks are missing
- emergent-cursor-rules — rules crystallize the taste corrections
- soul-md — one concrete artifact of taste