# AI-Native Tiers

Six-tier ladder for how someone uses LLMs day-to-day, due to Elena Verna ([[elena-verna-job-2027]]). The framing is non-engineer-shaped — the climb goes from text help to shipping production code — but the rungs survive recasting for engineering use.

The tiers are ordered by *what you let the LLM do*, not by sophistication of the prompt:

1. **Writing buddy.** LLM as editor. Catches unclear sentences, suggests stronger phrasings. Verna's discipline: write your own version first, then have the LLM critique it. Skip that and you're already in [[cognitive-debt]] territory.
2. **Meeting partner.** LLM as note-taker and after-action coach. Granola-style transcription plus prompts like "If you were my manager, what are my strengths?"
3. **Thinking partner.** LLM as adversarial collaborator. The whole point is to defeat [[ai-sycophancy-loop]]: prompts that explicitly request disagreement ("What's the strongest counter-argument?"), holes-in-thinking checks, idea-divergence-then-pruning ("Give me 20 ideas, cut to 3").
4. **Builder.** LLM builds working artifacts — internal tools, landing pages, prototypes — under user direction. The "vibe coding" tier in engineering vocabulary.
5. **Shipping to prod.** LLM-written code goes into the production system. For non-engineers this is the unlock Verna pitches; for engineers this is the [[simonw-vibe-coding-agentic]] tier where Willison admitted his own line collapsed.
6. **Agents.** Self-running recurring tasks ("every Monday, summarize sales and post to #revenue"). Verna's own caveat: "often require more setup, oversight, and cost than just doing the thing yourself." Consistent with [[agentic-coding-fatigue]], [[agentic-coding-is-a-trap]], and the supervisor-paradox of [[skill-atrophy-supervision-paradox]] from the engineering side.

## Why the ladder is useful

Most discourse about "AI use" treats it as binary — you adopt or you refuse. The ladder separates two things people conflate: *what tier you operate at* and *whether you operate at all*. Someone at Tier 2 every day and Tier 6 never is using AI fundamentally differently from someone at Tier 4-5 daily.

It also makes failure modes locatable. Tier 1 done lazily is cognitive debt. Tier 3 done without the disagreement-prompt is just sycophancy. Tier 5 done without verification is the principal-agent gap from [[agent-principal-agent-problem]]. Tier 6 done eagerly is the [[ceo-ai-psychosis]] / [[tokenmaxxing]] failure pattern.

## Where engineering changes the ladder

For engineers, Tier 4 (prototype building) and Tier 5 (shipping to prod) are the same skill at different levels of review discipline. The interesting line for them is between Tier 5 (you review and ship) and Tier 6 (the agent reviews and ships), where the [[code-review-principal-agent]] effort-signal collapse hits.

The ladder also flattens at the bottom for engineers: Tiers 1-3 are continuous with how a developer has used Google + Stack Overflow + a rubber duck for two decades, just with better latency. The original *novelty* in the ladder is at Tiers 4-6.

## Related framings

- [[jxnl-codex-maxxing]] — Jason Liu's operating loop reads as a Tier 5-6 production setup (durable threads, goals, heartbeats, side panel).
- [[average-is-all-you-need]] — Tier 5 democratizes "average" output, which is the real economic effect.
- [[dont-outsource-learning]] — Osmani's "form a hypothesis first, ask for explanation before code" is Tier 3 done as a learning discipline rather than a shortcut.
- [[contributor-poker]] — what you bet on a Tier 6 agent's PR vs a human contributor's first PR.

## See also

- [[elena-verna-job-2027]] — the source piece
- [[ai-sycophancy-loop]], [[cognitive-debt]] — Tier 1/3 failure modes
- [[skill-atrophy-supervision-paradox]] — Tier 6 erosion mechanism
- [[simonw-vibe-coding-agentic]] — Tier 4/5 line collapse from inside engineering