grounding-before-action

To prevent AI confabulation, enforce a solo tool call and wait for the real observed results before composing any follow-on actions or statements about tool outputs. This ensures the AI relies on actual data.

The best way to emit AskUserQuestion is in its own dedicated turn. By separating the question emission from other tool calls, you prevent context bundling and ensure clear user interaction before proceeding.

To verify targets before irreversible mutations, enforce a post-result echo-back and validation step. The AI must echo back the observed results and verify target references before executing any sequential writes or edits.

Yes, you must wait for tool outputs before composing follow-on actions. The workflow requires sequential reads and writes, meaning all statements about tool outputs must come after actually observing those real results.

Guardrails for sequential reads and writes require verifying target references and titles before any edits. This prevents premature mutations by enforcing post-result echo-back and validation after observing real tool outputs.

AI confabulation occurs when tool calls are bundled because the AI attempts to compose follow-on actions without waiting for real observed results. Enforcing solo tool calls and sequential processing prevents this issue.