Bricks

Describe what you see, not just what you want

Read the AI's response before reacting — it often answers part of your question already.
When something looks wrong, describe what you see, where (page or route), and what you expected — not just what you want fixed.
Tell the AI which role you're connected as. “I don't see X” is ambiguous when you hold several roles — role visibility shapes what's on screen.
When the issue is visual, attach a screenshot of the actual screen with the page or route it lives on. Skip the screenshot when a one-line description says the same thing.
Refer to platform-mapped concepts (workflows, roles, tables, apps) by name. “The gallery on the members app” is much cheaper than “the screen with photos”.
The AI can read from your other apps in this organisation. If you already built something similar, mention it — it's cheaper than re-explaining.
Answer the AI's questions literally. When it asks which error, which page, or which role, give that information — saying “just fix it” sends the next turn back looking for the same missing piece.
Frustration costs the same as diagnosis. “It's broken!” or insults burn a full turn for zero signal. The same turn spent on the error message, the page, or what you already tried gets you to a fix.
Don't paste prompts or answers from another AI as-is. The AI here has your full project state; the other one didn't — its instructions for your case may be plausibly wrong. Summarize the idea in your own words instead of pasting raw.

Don't ask for too much at once. Bricks gives focused responses for small iterations — it's not an autonomous agent that builds for hours unattended.
If you can't convey a concept in 5 messages, you're heading the wrong direction. Stop, reset, try a different framing.
Let turns finish before starting a new one. Abandoning a turn mid-stream invalidates the prompt cache and makes the next turn more expensive too.
Don't ask the AI to “start over” on an existing app. Ask for surgical fixes instead — recreating from zero loses every iteration of context.
When the AI mentions “app_antipattern_warnings”, take it seriously and ask for the fixes on the next turn — those are SDK misuses the AI just shipped that compound silently.
Don't revert + republish an old app version lightly. The current boilerplate gets re-merged on republish, and old AI patterns can silently break on deprecated SDK fields.

During first build, don't fear running manual workflows to create, edit or delete test data. You can clear it later when real users arrive.
Don't fear reading data and workflows yourself. Having the AI show you is much cheaper than guessing in the dark.
Workflow expressions in {{...}} are intentionally simple — only ||, comparisons and basic arithmetic. No &&, no ternary, no parens. Each branch needs its own complete chain.
Bigger and more complex projects cost more per turn — growth is neither linear nor exponential. That's a context-size constraint, not a commercial choice. Future building modes will accommodate larger projects.

Prefer text-based PDFs over scanned ones — scans force the AI to read every page as an image, which costs more and reads less reliably.
For photos, use JPEG or WebP instead of PNG — same content, much smaller file.
Avoid 4K screenshots when 1080p captures the same information.
Strip embedded fonts, animations, and effects from documents — they bloat the file without helping the AI.

Decorative images, logos, separators, and graphic headers add tokens without adding signal — remove them before attaching.
Strip out repeated headers and footers from long PDFs.
If a section isn't relevant to your question, cut it out — the AI doesn't need to skim past it.

For data exports (CSV, JSON, member lists), 30 anonymized rows usually beat 30,000 raw ones — the patterns are the same, the cost isn't.
Long PDFs (more than ~25 pages) confuse the AI as much as they cost; split them into focused sub-PDFs and ask one question at a time.
If you don't know what the AI will need, paste a small sample first and let it ask for more.

Strip names, emails, phone numbers, and addresses before sending personal-data files to the AI.
Don't share photos of identifiable people without their consent — same rule as posting them publicly.
If your data has identifiers (member IDs, customer numbers), the AI rarely needs them — replace with sequential placeholders.