10 vibe coding tips from YC founders
What the best founders are doing to actually get results from coding agents
“This is the best f*cking time ever, in the history of technology, ever, period, to start a company.” - Sam Altman
Scrolling Twitter, you’d think AI had already automated away all devs. But if you’ve tried vibe-coding yourself, you know it’s not quite that simple.
Over the past few weeks, I’ve chatted with dozens of YC founders actually building with tools like Cursor and Claude Code. Most aren’t coding 100x faster, but many are moving 20–50% faster using specific tactics that make AI effective.
Here’s a distilled set of their best advice, so you can do the same. Happy vibe-coding 😄
1. Starting from scratch works best
These tools perform better when writing new scripts/components rather than modifying existing codebases.
Use AI for standalone utilities, UI scaffolding, or fresh feature prototypes rather than editing large, existing systems.
2. Create a plan before generating code
Ask the model to write an outline/pre-execution plan before making changes - this catches misalignment early.
o3 tends to be the model of choice for planning.
3. Document feedback along the way
Ask the AI to maintain a lightweight file updated any time you give feedback (e.g. CLAUDE.md).
This helps capture conventions, preferences, and patterns that emerge over time.
4. Break tasks into small, clear steps
You need to give Cursor/Claude focused, specific instructions.
It can feel like a pain, but take the time to break things down as small as possible.
Avoid vague prompts or requests that touch too many parts of the system at once. (Can’t stress this one enough!)
5. Build in safety nets
Always ensure changes can be undone easily using version control or checkpoints (git and branching are your friends).
If the model starts making crazy changes, just drop that branch rather than try to course-correct.
6. Don’t expect consistency across all tasks
Performance is often unpredictable—great on one task, unreliable on the next.
This is expected (at least for now)… over time you’ll build intuition for where it works and where it doesn’t.
7. Treat it as a collaborator, not a replacement
The best results come from pairing closely with the assistant—thinking through problems, breaking them down, and reviewing its work.
Expect to do diff reviews, bug fixes, and handholding—it’s not fully hands-off yet.
Think of it as a productivity multiplier, not a magic button.
8. Rework your workflow to take full advantage
While certain parts of coding can actually be slower with AI, the end-to-end workflow of building a feature is usually faster.
When product people can talk to customers, then immediately build an MVP (with AI) you get feedback quicker.
This shifts the effort from planning and delegation toward testing and iteration.
9. Use multiple models and compare results
Different agents have different strengths—some are better at small edits, others at broad refactors.
This post from Besimple.ai (YC X25) breaks down a popular strategy:
4o for cleaning up prompts
o3 for planning
Sonnet 4 for actually writing the code
10. Expect productivity gains, not magic
While not 10x faster (yet), a 20–50% speedup is common across many tasks.
The cumulative time savings add up—especially for boilerplate, UI scaffolding, and small migrations.
Even when it’s not faster, offloading routine work can free up cognitive space for higher-level thinking.
In summary, build workflows that lean into what AI coding tools are good at today, and steer clear of what they’re not.
Resources
