Skip to main content

Safe prompting

The guardrails do the heavy lifting, but how you ask still matters. These habits get you better, safer results.

Say what, not how​

Describe the outcome you want, not the code you imagine. Let the AI use the Task Recipes and MCP queries to find the correct implementation.

  • ✅ "Show quarterly production for the last three years as a line chart."
  • ❌ "Write a useEffect that fetches /api/prod and parses the JSON."

Name real things​

Refer to data and features in business terms. The AI will map them to real endpoints via the API MCP. If you paste a made-up endpoint or component name, you invite it to follow you off the guardrails.

Ask it to verify​

When the AI reports something is done, ask it to run the verify step and show you the result:

Run the build and confirm it passes before we move on.

A passing verify gate is your proof, especially when you can't read the code.

Work in small steps​

Request one feature at a time. Small changes are easier for the guardrails to verify and easier for you to sanity-check.

Watch for these red flags​

If the AI does any of the following, stop and ask it to re-check against the MCP servers:

  • References an endpoint, token, or component without querying for it.
  • Says "done" but skips or can't show the verify result.
  • Suggests disabling a guardrail, instruction file, or check to "move faster."
  • Hardcodes secrets, tokens, or credentials into the code.

A good prompt, end to end​

Using the production endpoint, add a KPI card showing total production this quarter versus last quarter, styled with the design system. Then run the build and confirm it passes.

That tells the AI what you want, keeps it on real endpoints and tokens, and demands verification — all three guardrail layers, in one sentence.