Three iteration modes
Build mode (default)
You prompt; the agent reads, plans, writes, hot-reloads. Use it for anything that touches code: features, refactors, bug fixes, design changes, schema updates. This is what 95% of your turns are.Plan mode
Plan mode is read-only — the agent can search, read files, browse, and reason, but it can’t write. Use it before any change you want a second opinion on.“We’re about to add multi-tenancy. Read the auth and DB code and tell me what would have to change. Don’t write anything yet.”The agent researches and responds with a plan. When the plan is right, prompt “go ahead and implement that” and the full toolset unlocks. Plan mode is the cheapest way to avoid five-turn rabbit holes on big changes.
Visual edits
Click any element in the preview, type a change (“make this button bigger”, “use the brand purple here”). The agent patches just that element — no re-think of the design system, no unrelated changes. Best for: design tweaks, copy edits, spacing adjustments, brand alignment.Shape your prompts to the change
| What you want | Prompt shape |
|---|---|
| A small, surgical change | ”In pages/Settings.tsx, change X to Y. Don’t change anything else.” |
| A feature touching many files | ”Add a dark-mode toggle. List the files you’ll touch first.” |
| A refactor | ”Refactor useDeals to use TanStack Query. Don’t change behaviour.” |
| A bug fix | ”When I click ‘Save’, nothing happens. Investigate, then fix.” |
| A design change | ”Make the dashboard feel more like Linear — denser, mono-spaced numbers, less colour.” |
| Anything ambiguous or risky | ”Plan how we’d do X first. Don’t write code yet.” |
When the agent gets stuck
- Re-read the brief. Half the time a stuck agent is acting on an outdated assumption from earlier in the chat. A two-line restatement of what you actually want resets it cheaply.
- Switch to Plan Mode. If the agent keeps making the same wrong edit, “Switch to Plan Mode and figure out what’s actually causing X” gets you a research turn instead of another doomed attempt.
- Ask the agent to clarify. If it’s guessing instead of asking, prompt: “If you’re uncertain about X, ask me — don’t guess.” Pin this in custom instructions so you never have to type it again.
- Restart the runtime. If the preview is wedged, hit Restart in the project menu — files are preserved, the runtime is rebuilt fresh.
Memory across turns
Long chats get compacted — older messages summarised to make room. Memory is what survives. Memory is a small file (.vibely/memory.md) that lives with your project and is read back at the top of every compacted block. Use it for facts you don’t want the agent to forget mid-session:
“Save to memory: this is a B2B app, never use any consumer-style copy. Always require login.”See Tips → Memory for what belongs there and what doesn’t.
When to start over
Sometimes the cheapest path forward is a new chat. Heuristics:- The chat is 100+ turns and the agent is making contradictory decisions
- You’ve changed the brief substantially since turn 1 (e.g. CRM → CMS)
- You want to try a fundamentally different stack (Next.js instead of Vite)
- The architecture you ended up with isn’t the one you’d have asked for
“Summarise the lessons from this session in 5–10 bullet points I can paste into a new chat.”
Next
Tips → Iterating
Deeper coverage of multi-turn habits.
Tips → Prompting
Prompt shapes that hold up across long sessions.
Project settings
Restart sandbox, version rollback, GitHub sync.
Tips → Troubleshooting
What to do when things break.