Intent Preview
What is Intent Preview?
The Intent Preview pattern addresses a core anxiety in agentic AI: users need to understand what will happen BEFORE it happens. Unlike traditional AI where the user explicitly types a prompt and evaluates the response, agentic actions may be initiated proactively or involve consequences that are difficult to reverse - sending emails, booking flights, modifying files. This pattern shows a clear, scannable summary of planned actions using plain language (not technical jargon), with each step marked for reversibility and editable by the user. The preview must be sequential for multi-step operations, highlight irreversible actions visually, and never auto-dismiss. This transforms the approval moment from a binary yes/no into a structured review that builds trust and catches misunderstandings before they cause harm.
Problem
When an agent is about to take a multi-step action, users need to understand what will happen before it happens. Without an intent preview, users experience anxiety leading to constant monitoring or blind trust that erodes at the first mistake.
Solution
Before any significant action, present a clear, scannable preview showing planned steps in plain language, with reversibility indicators, edit controls for individual steps, and explicit approve/reject buttons. Never auto-dismiss the preview.
Real-World Intent Preview Examples
Implementation
When to use Intent Preview, and when it backfires
Use it when
- The action is consequential or hard to reverse: sending mail, moving money, deleting files, booking, posting. The cost of a wrong action is higher than the cost of a glance.
- The agent inferred intent rather than being told step by step. The preview is where you check that the agent understood what you actually meant.
- The plan is multi-step and the user can usefully edit it: drop a step, change a recipient, fix an attachment before anything runs.
Don't, or minimize, when
- The action is trivial and reversible. A preview for every autocomplete or draft just trains users to approve on reflex, and a reflex-approval is no approval at all.
- You can only show a preview that summarizes the plan but isn't the literal contract the executor runs. A pretty summary stapled to a different execution path is worse than no preview.
- The preview would lag the real state. If the agent re-plans after approval, the thing the user signed off on no longer exists, and you've collected consent for a phantom.
The trap
The cosmetic preview: a tidy, reassuring summary that doesn't match what actually executes. The diff quietly omits the step that overwrites the file. The summary says 'archive 3 threads' while the job deletes them. 'Email the team' renders as one calm line and fans out to 400 external addresses. The user reads it, it looks reasonable, they approve, and the agent does something they never agreed to. This is worse than no preview, because a preview manufactures consent: now there's a record showing the user said yes to an action they were never actually shown. Real approval requires that the words on the card are the exact instructions that run. If the summary and the execution can drift apart, you didn't build a preview, you built an alibi.
Take it into your own product
- 1
The preview must be the thing that runs, not a description of it.
Render the card from the same plan the executor consumes. The moment the summary is built separately from the instructions, they can drift, and a preview that can lie is worse than none: it collects consent for an action the user never saw.
- 2
Preview consequence, not just intent.
'Send email' and 'send to 400 external addresses' are the same intent and wildly different actions. Show reversibility and blast radius per step so the irreversible, wide-scope move can't hide inside a calm one-liner.
- 3
Approval is for the plan as shown, and only that plan.
If the agent re-plans after the user approves, the thing they signed off on no longer exists. A stale yes carried forward is a phantom consent. Re-prompt when the plan changes; never auto-run on a timeout.
- 4
Make it editable, or it's just a confirmation dialog.
The value of a preview is catching the misunderstanding before it executes. If the only options are approve and reject, you've reproduced the yes/no you were trying to replace. Let users drop a step or fix a recipient in place.
- 5
Record what was shown next to what ran.
Persist the previewed plan alongside the executed actions. If they ever diverge, that's the incident that proves your preview was honest, or proves it wasn't. Without that record, you can't tell a real approval from an alibi.
Add Intent Preview to your product
Copy the prompt below into Claude Code or Cursor in your repo. It encodes the four moves on the left and asks Claude to find your AI decision surfaces and update them. Claude reports what it changed and asks before adding dependencies.
Check if your product already has this pattern
Upload a screenshot. We'll tell you which of the 36 patterns your AI interface uses and where the gaps are.
Audit My DesignMore in Human-AI Collaboration
Contextual Assistance
Offer timely, proactive help and suggestions based on user context, history, and needs.
Human-in-the-Loop
Balance automation with human oversight for critical decisions, ensuring AI augments human judgment.
Augmented Creation
Empower users to create content with AI as a collaborative partner.
