Lovable AI Playbook: Modes, Prompts and Credits That Last

Lovable AI is the agent behind every Lovable project: you describe a change in the project chat, and it writes, runs and checks the code. You do not pick its model. Lovable chooses, and says it uses Claude Opus 5 among its models and routes stuck or complex sessions to Fable 5.1 (September 2026). What you do control is how many credits each idea burns. Plan mode costs 1 credit per message plus any subagent research, while Build mode bills the work done, failed attempts included. The working habit is simple: diagnose in Chat, plan bigger changes when a redo would cost more than the plan, and fix copy by hand in the preview.

Chat, Plan and Build do different jobs

The project chat runs in one of three modes, and choosing the right one is most of the craft. Chat mode reads your code, files, app data and connected tools such as Slack or Notion, yet it never edits files, runs database migrations or publishes. It is the place for questions like why does this page load slowly? If a request would act through a connected tool, Lovable asks for approval first unless you saved Always allow. Your project itself stays untouched.

Plan mode inspects files, the database and logs, asks clarifying questions and writes an editable plan into .lovable/plan.md. It never touches code until you approve. Each message, including a revision of one highlighted part of the plan, costs 1 credit plus whatever subagent research it starts ^[1]. Approval switches to Build mode, billed separately on the work done. There is no upfront estimate. The pricing page lists sample Build costs: 0.50 credits to turn a button gray, 0.90 to remove a footer, 1.20 to add sign-up and login, and 1.70 for a landing page with images ^[2].

Free, Pro and Business workspaces also get 5 build credits a day: up to 30 a month on Free, and on top of plan credits on Pro and Business, where some regional plans cap them at 30 a month. Enterprise gets none. Daily credits reset at 00:00 UTC and never roll over, and Lovable spends such grants before general credits. Rollover windows and plan prices sit in our Lovable pricing guide.

Plan first when a redo would cost more than the plan

Plan mode has a clear break-even. A plan message costs 1 credit plus its subagent research, so planning a 0.50-credit color change costs more than the change itself. For work that touches several files, the database or sign-in, the math can flip. Lovable charges every AI action regardless of outcome, and reverting a version refunds nothing. A plan pays for itself only when the redo it prevents costs more than the plan message. A light plan with little research can beat one redo of sign-up and login at 1.20 credits. A research-heavy plan may not. Button tweaks never do.

Three controls keep long builds honest. Stopping a Build request charges only the work completed. Credit check-ins (beta) pause a single message when it crosses 20 credits by default, a level you can set up to 100,000; it is a prompt to continue, adjust or wrap up, not a hard cap, and a run can slip slightly past it. When the workspace runs dry mid-message, the build pauses with no deadline and offers Add credits or Wrap up.

Errors get their own allowance. Each account has 10 free Try to fix uses, shared across all workspaces and including fixes from the security view; each returns 24 hours after use, and past ten a fix is ordinary build usage. Security scans stay free. When the same error keeps returning, repeated Try to fix clicks become the expensive habit, because every attempt is charged. Stop and diagnose first: Chat mode can read the code and logs and explain what is failing without changing anything, then plan the repair. Reviewers on Trustpilot most often complain about credits running out, and fix loops show up among those complaints.

Do the small edits yourself

The preview toolbar, on every plan, replaced the older Visual edits. Inline text edits there skip the agent: each account gets 100 free ones, each regained 24 hours after use, plus a shared workspace pool of 2,000 per 24 hours ^[3]. Beyond that, they cost credits. Select elements and Draw annotation count as build usage. Comments are free, but sending a thread to Lovable is not. On Pro and above, editing code yourself in the code editor uses no credits; Free shows code read-only. Typos belong in the toolbar. Layout changes belong in a prompt.

Drafts are the safer way to try a risky idea. A draft is a copy of the project with its own chat, much like a branch; creating, accepting or deleting one is free, and only messages sent inside it spend credits. Accepting merges it back without publishing. The catch: a draft shares the project's database, files and settings, so data changed while testing is real. Git sync does not carry drafts yet. Version history keeps every change, with bookmarks for stable points, and a revert restores code only, never database records.

Teach the agent once instead of in every prompt

Instructions, called Knowledge until October 2026, live at workspace and project level with up to 10,000 characters each; only owners and admins manage the workspace set. Lovable also reads an AGENTS.md or CLAUDE.md file at the project's top level with every message. Architecture decisions get written into AGENTS.md as rules that travel with the code to GitHub, and on paid plans Lovable may rewrite or remove rules that have gone stale. Put stack choices, naming conventions and files never to touch there, once.

Memory is on by default on every plan. Project memory is shared by everyone who can edit the project, personal memory is private and follows you across projects, and project memory wins when they disagree. On paid plans, Dreaming tidies project memory in the background at no credit cost and never changes code, though it skips projects with Git sync. One caution: Lovable strips values that look like API keys, tokens or passwords from memory but can miss other confidential text.

Skills are reusable markdown playbooks at workspace level, applied automatically when a task matches or called with a slash command. Owners and admins add them by writing them, building them with Lovable, importing from a public GitHub repository or uploading a ZIP. They share the SKILL.md shape of Anthropic's Claude skills, so a skill built in Claude uploads to Lovable as is. Using one adds nothing on top of the message. Skills do not work in Chats, the workspace-level agent outside any project.

Long runs with /goal

A single Build message can work for up to 10 hours. The /goal command goes further: Lovable keeps working until the goal is met, asks no questions and stops early only when it needs something only you can provide ^[4]. It keeps the final half hour to finish and summarize. A goal costs noticeably more than a regular Build message, because the same work uses more credits and a goal does more of it. It still pauses at your check-in level, and stopping it charges only for completed work.

Goals suit well-specified work you would otherwise babysit, such as finishing a feature list from an approved plan. Set the check-in level before you start, and raise it only after you have seen what a goal that size costs. For shorter sessions, follow-ups, which replaced the older message queue, let you line up the next requests while Lovable is still working. Line up small, independent changes. Keep dependent ones in a single plan, so the agent sees the whole job before writing code.

Models: who picks them, and AI inside your app

You cannot choose the model for Build, Plan or Chat. Subagents, the read-only helpers that research and inspect a project, sometimes several at once, consume model usage like other agent work, with routine lookups routed to lighter models. They cannot edit files. Benchmark claims stay Lovable's own: it says Fable 5.1 beats Fable 5 by up to 17% on its hardest iterative code-fixing tasks at 31% lower cost, on its internal test suite (September 2026).

Inside your apps, the choice is yours. The Lovable AI gateway adds chat, image, video, voice and search features with no provider account or keys, billed per model call in credits after a monthly grant of 4 AI credits. GPT-6 Astra is the default chat model and GPT-6 Luna the lower-cost option; Claude models run only in apps created from May 2026 (June 2026 on Enterprise) on TanStack Start. Showing the model's thinking makes answers cost more. When a model gets a retirement date, Lovable emails the owner, then redirects requests to a replacement or lets them fail. More in our Lovable Cloud guide.

Two facts matter once the app is live. Visitors' AI calls spend the workspace owner's credits, and at zero those features fail with an HTTP 402 error while pages keep serving. Free and Pro content may be used to train Lovable's models from September 2026 unless you opt out in account settings; Business and Enterprise are excluded by default. The AI app context setting also keeps your app's full AI requests and responses for 90 days, secrets removed, on by default on Free and Pro.

The bottom line

You tune Lovable's agent by picking the mode, not the model. Diagnose in Chat before you rebuild, plan changes whose redo would cost more than the plan, type copy fixes into the preview toolbar, and store your rules in instructions, AGENTS.md and skills so you stop repeating them. Set a check-in level before any /goal run. Treat every failed attempt as spent, because it is. Work that way and more of each month's credits land in changes you keep, not in fix loops you throw away.