Agents

Best practices for using agents

Buildprint agents are powerful, but they’re still AI — the quality of what you get out depends directly on what you put in. Follow these guidelines to get the best results and avoid frustration.

Garbage in, garbage out

Agents respond to what you ask for. A vague prompt like “fix the checkout” gives the agent a wide range of possible interpretations — and it might pick the wrong one. A specific prompt like “the Stripe webhook isn’t updating the order status when payment succeeds” gives the agent a clear target.

Before prompting, take a moment to think about exactly what you want changed. The more precise your instructions, the more predictable the result. This applies to both questions and editing requests.

Stay in the loop

Agents work best as collaborators, not autopilots. Review what the agent proposes before confirming changes. Ask clarifying questions if the plan doesn’t look right. Check the agent’s work in the Bubble editor after it makes changes.

The agent explains its plan before executing. Iterate on the plan until it matches what you actually want — don’t just approve and hope for the best.

When things go wrong, restart with a narrower prompt

If an agent conversation goes off track, it’s almost always easier to start a new conversation with a better prompt than to try to get the agent to fix its own mistakes. Agents don’t always recover well from compounding errors.

The pattern is:

  1. Restore to your last savepoint if the agent made unwanted changes.

  2. Identify what went wrong — was the prompt too broad? Did you skip reviewing the plan?

  3. Start a fresh conversation with a more specific, narrower prompt.

Vague prompts lead to a wider set of possible outcomes. Narrowing the scope reduces the chance of unexpected changes.

Design in code

When working on design, it can be helpful to ask Buildprint to produce designs in HTML first before implementing them. This lets you iterate more quickly, without editing the Bubble app. It also improves the quality of output as the agent will not stop until it has completely implemented the updated design.

Always test in development first

Never deploy agent-generated changes directly to production. Use a development or staging version of your Bubble app to test changes before merging them to your live app.

Start in Plan mode, switch to Build only when needed

The safest way to use agents is in Plan mode. This lets you ask questions and understand your app without risk of accidental changes.

Only switch to Build when you’re actively working on a specific change. Switch back to Plan when you’re done — this prevents accidental edits from vague or misunderstood prompts.

Review agent suggestions before confirming

When an agent proposes a change, read through what it plans to do. Ask clarifying questions if anything seems wrong or incomplete. Common issues to watch for:

  • Privacy rule changes that might expose sensitive data

  • Workflow modifications that skip validation or error handling

  • Database field changes that break existing workflows

  • Overly broad changes that affect more than intended

Verify AI outputs against your Bubble editor

Agents can hallucinate or misinterpret your app’s logic. Cross-check their answers for critical configurations like privacy rules, conditional logic, database field types, and API connector settings.

If an answer seems wrong, ask follow-up questions or rephrase your prompt. Agents improve with more context.

Use reviews for critical branches

Run reviews on every branch that touches high-risk areas: payments, authentication, user data, or privacy rules. Reviews help catch regressions that manual testing might miss.

Don’t rely solely on the AI-generated test plan — add your own manual test cases based on your product knowledge.

Understand the limitations

Buildprint is in free beta and doesn’t guarantee agent behavior, uptime, or accuracy. Always maintain backups and version control for your Bubble app.

AI-generated changes are not guaranteed to be correct. You are responsible for testing and verifying all outputs before shipping to production.

Forking chats

Buildprint allows you to fork (branch) chats from any message. This will compact the conversation and allow you to continue it in a new chat with clean context. This is useful if the agent is stuck in an endless cycle, or the context is too long. You can keep costs down by keeping conversations short. It will summarise the previous conversation and the new conversation will have that context, as well as any attachments in the previous conversation.

Was this helpful?