Fall-On-Your-Face Taste Calibration
Deliberately push AI past its limits in a safe environment to build a gut feel for what it can do.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 85%
To use AI tools well you need 'taste' — a gut feeling for what complexity of task a model can handle and how much you need to specify. Truell argues this can't be taught verbally; you build it by intentionally being over-ambitious in a low-stakes setting to discover where the model breaks, and you must rebuild it whenever a materially new model ships.
Origin
Michael Truell's advice for new Cursor users, framed as tacit knowledge that must be acquired experientially.
Core principles
- 01Much of the skill of using AI is tacit and can only be learned by doing, not by instruction.
- 02You learn a model's real boundary by crossing it, not by staying safely inside it.
- 03People more often underestimate than overestimate what current models can do.
- 04Each new model has different quirks and personalities, so the gut feel must be periodically recalibrated.
How to run it
- 1
Pick a safe environment
Choose a side project or low-stakes context, not your critical professional work, so failures cost nothing.
Pro tip Especially useful for developers wedded to existing workflows who are anchored on old assumptions.
- 2
Deliberately go for broke
Be maximally ambitious — explicitly try to make the model fail so you can see where the limits actually are.
Pro tip You may be surprised at the places where the model doesn't break.
Watch out Don't run this experiment on production professional work where a failure has real consequences.
- 3
Map what worked and what didn't
From the failures and surprises, build an internal map of task complexity the model handles and how much specification it needs.
- 4
Recalibrate on every major model release
Treat a significant new model as a reason to re-run the exercise, since quirks and capabilities shift.
Watch out Don't assume last model's limits still hold; each has different quirks and personalities.
In the wild
Truell says they 'run into people who haven't given the AI yet a fair shake and are kind of underestimating its abilities,' and advises those developers to explicitly try to fall on their face on a side project to discover the true limits.
→ A grounded, current gut feel for model capability that replaces stale assumptions.
Common mistakes
Giving the AI a timid shake and concluding it's weak
Staying within assumed limits means you never see the model's real ceiling and you under-use it, especially common among senior engineers anchored to old workflows.
Never recalibrating after a new model ships
Capabilities and quirks change materially between models; relying on an outdated gut feel leads to both over- and under-trusting the new model.
Is it for you?
Best for
New AI-tool users and experienced engineers who are anchored on prior assumptions about model capability.
Not ideal for
High-stakes production work where you can't afford exploratory failures.
From the transcript
“I would encourage people to explicitly try to fall on their face and try to discover the limits of uh what these models can do…”
“You might be surprised in some of the places where the model doesn't break”
“haven't given the AI yet a fair shake and are kind of underestimating its abilities”
“each of these things have slightly different quirks and different personalities”
From the episode
The rise of Cursor: The $300M ARR AI tool that engineers can’t stop using
Michael Truell (co-founder and CEO)