> ## Documentation Index > Fetch the complete documentation index at: https://docs.vibely.sh/docs/llms.txt > Use this file to discover all available pages before exploring further. # Iterating > Stay in flow across many turns without breaking what already works. A Vibely session can run for hundreds of turns. Most of what makes a long session productive — not exhausting — is in how you iterate, not in how fast the model is. **TL;DR** * One change per turn. Revert, don't patch. Pin invariants you keep re-saying. * Use [Plan Mode](/docs/tips/plan-mode) before anything that touches 3+ files. * "Green state" = preview works + latest turn made sense + types clean. That's when you push to GitHub. ## Keep each turn small A turn that does one thing is easy to roll back, easy to verify, and unlikely to break anything else. A turn that does five things is the opposite. When you catch yourself writing a prompt with multiple "and"s, split it. Five small turns take less *total* time than one bundled prompt that fails twice. ## Trust the live preview The preview reflects every turn within seconds. Click around between turns — don't wait until you have ten changes stacked up. When something works, say so explicitly: > *"Yes — that's the layout I want. Keep that pattern going forward."* Confirmations get picked up by future turns. They prevent the agent from second-guessing a good answer. ## Roll back instead of patching If a turn produces something wrong, the right next prompt is almost always a revert. > *"Undo the last change to the orders page and try again — this time use a stacked card layout instead of a table."* Reverts are clean. Bug-fix-on-top-of-bug-fix prompts accumulate cruft and make future iterations harder. ## Pin invariants early Once you're happy with a foundational decision — component library, layout primitive, brand colour, copy tone — write it down in [project memory](/docs/tips/memory). Ask the agent: *"Save to memory: …"* and it lands in `.mana/memory.md` for every future turn. The earlier you pin, the less drift you fight later. ## Sync to GitHub when you're in a green state Green state checklist — push when **all** of these are true: * The preview loads and renders without errors. * The latest turn did what you asked. * The code your eyes scanned looked right. * `tsc` / lint is clean (or no worse than before the turn). * You'd be okay with the project ending here. If a future turn breaks something, branch off the green commit and replay. Don't treat the Vibely project as your only history. GitHub is your safety net. ## Use [Plan Mode](/docs/tips/plan-mode) for anything risky When the next thing you're about to ask for touches more than one or two files, ask for a plan first. > *"Plan how we'd switch from a single-tenant data model to multi-tenancy. List the files you'd touch."* You get a written approach back. Edit, push back, ask for alternatives — until the plan is right. Then *"implement it"* unlocks the agent's full toolset. [Plan Mode](/docs/tips/plan-mode) is the cheapest way to avoid expensive mistakes — flat 1 credit per turn regardless of how much it reads. ## Know when to start over Sometimes the cheapest path forward is a new chat. Heuristics: * **The agent has looped on the same error 3+ turns in a row.** Context is poisoned. * **The chat has crossed 100 turns and the agent is contradicting itself.** Compaction has eaten too much. * **You've changed the brief substantially mid-session.** The agent is still trying to satisfy the original brief. * **The architecture you ended up with isn't the one you'd have asked for.** A fresh chat with the lessons from this one usually lands somewhere better. Starting over *doesn't lose your code* — files are stored, the new session restores them. What you lose is the chat context, which at that point is more confusing than helpful. Before starting over, ask the agent: > *"Summarise the lessons from this session in 5–10 bullet points I can paste into a new chat."* ## Use comments and visual edits for design tweaks For visual changes, two faster paths than prompting from scratch: * **Visual edits** — click any element in the preview, type a change ("more padding", "use the brand purple"). The agent patches just that element. No re-think of the design system. * **Pin comments** on the preview ([Project settings → People → Collaborators](/docs/workspace/project-settings#comments-on-the-preview)). Prompt *"address the open comments"* and the agent takes them as a batch. ## Next Prompt shapes that hold up across long sessions. What survives compaction, what doesn't. When the loop breaks down. GitHub sync, version history, and other safety nets.