Claim
Operators who use AI tools as upgraded versions of old tools (fancier autocomplete, slightly faster search) under-perform operators who treat them as new tools — pushing for ambitious, end-to-end outcomes and retrying from a clean prompt when stuck rather than iterating against broken context.
Mechanism
LLMs degrade fast when context is polluted by failed attempts. Repeated retries against the same broken state stack errors. Starting fresh with an ambitious end-state goal ("build the whole thing") gives the model room to plan its own decomposition, which is often better than the operator's. The discipline is two-part: ambitious framing, plus the willingness to throw away and re-prompt.
Conditions
Holds when:
- The operator can hold the ambitious end-state in mind without micromanaging.
- The cost of retry is low (a few minutes, a few tokens).
- The model is capable enough to handle the ambitious framing.
Fails when:
- Compliance or audit requires the original context preserved (legal review, regulated workflows).
- The operator does not yet have judgment to recognise pollution and reset.
- Retry costs are high (long-running jobs, expensive multi-tool chains).
Evidence
"People who use the new tools as if they were old tools tend to not succeed."
Specific pattern Mann names: Claude Code users who ask for ambitious changes and retry the prompt 3+ times outperform those who try once and bang on the same failure.
— Benjamin Mann on Lenny's Podcast, 2026-04-28
Signals
- Team velocity per AI-fluent operator climbs sharply versus AI-skeptical or AI-incremental peers.
- Stuck operators ship a "throw away and start over" reset rather than nineteen iterative tweaks.
- Output quality improves with retries-from-scratch, not with retries-on-context.
Counter-evidence
For some classes of debugging, preserving context is essential and resetting destroys hard-won state. The discipline is a heuristic, not an absolute.
Cross-references
- The Economic Turing Test — would you hire the agent if you didn't know it was a machine? — same operator, the destination this discipline reaches
- Give the model tools and a goal; do not hard-code the workflow — Boris Cherny's architectural complement