1 min lesson
Explore the codebase to write the spec - no code required
Describe what "Explore the codebase to write the spec" changes in practice.
Step 1 of 3
Explore the codebase to write the spec - no code requiredWhere the spec comes from
If the small project hands you an unfamiliar repo, you don't need to read it line by line to scope a real change. Drive Ask modeA read-only mode for asking questions about a codebase without changing files; the safe way to explore unfamiliar or legacy code. Press Enter for the full definition. against the actual code: pull the target folder into context, interrogate it, find the gap, then have the agent write the spec for you. The whole explore-to-spec sequence happens without you writing a line.
Learn more
Full explanation
Explore → Plan → Spec, Without Writing Code
Interactive diagram. Step through it with the Next and Previous controls below, or Tab to a region to read its detail.
Run this against a real repo in Ask mode to scope a change, then let the agent draft the spec. A PM did exactly this live and never touched code.
A PM ran this exact sequence on a vendor-waterfall app in a workshop: orient, diagram, probe geography, find the unused data, outline the fix, then have the agent draft a 1300-line spec - all without writing code. The interview lesson is the same as the demo: the artifact is cheap once you've used Ask modeA read-only mode for asking questions about a codebase without changing files; the safe way to explore unfamiliar or legacy code. Press Enter for the full definition. to find the real gap. Spend your exploration finding the unused-data kind of insight, not reading the repo top to bottom.
Learn more
Full explanation
One-page spec skeleton
# <Feature> - onsite spec
Problem: The paper cut I feel daily, in one sentence.
User: The single persona this is for (e.g. solo pro on a big repo).
Slice: The thin end-to-end thing I'll ship today.
Cut: What I'm deliberately NOT building and why.
Metric: The one number I'd watch in week 1 (e.g. % agent diffs
reviewed per-hunk vs blanket-accepted).
Risk: The thing most likely to make this wrong (latency? trust?).Narrate your prompting as design decisions, not magic. “I asked the agent to scaffold the diff view, then I rewrote the accept handler because it auto-applied without a confirm - that's a trust bug for this feature” shows you drive AI and exercise judgment over its output. That's the exact builder-PM signal Cursor is screening for and it's invisible if you type in silence.
The trap is rabbit-holing on a polished detail and arriving with nothing to show. Protect the demoable slice above everything. A thin working vertical with honest gaps and a clear metric beats a beautiful fragment you can't run or defend.
Learn more
Full explanation
Iterate once on the harshest feedback
Iterate once on the harshest feedbackThe second rep
After the demo, take the single sharpest piece of feedback and redo that one piece. The willingness to revise live, without getting defensive, is a culture signal at a place that prizes truth-seeking over ego. Don't rebuild everything - fix the one thing that was most wrong.
QYou're 70 minutes into a 90-minute prototype project and one layer is polished but the end-to-end path doesn't run. What do you do?