Plan Summary
What is Plan Summary?
While Intent Preview shows WHAT the agent will do, Plan Summary explains WHY and HOW. When an agent breaks a complex goal into subtasks, users need to understand the agent's reasoning - not just its intended actions. This is especially critical for knowledge work where there are multiple valid approaches. 'Research competitor pricing' could mean scraping websites, reading analyst reports, or checking public databases - the strategy matters as much as the outcome. The Plan Summary provides goal interpretation, strategy explanation, a subtask checklist that updates in real-time, explicit assumptions the user can correct, and resource and time estimates. This pattern extends explainability from retrospective ('here's why I gave this answer') to prospective ('here's why I'm taking this approach').
Problem
While Intent Preview shows what the agent will do, users also need to understand why and how. When an agent breaks a complex goal into subtasks, users can't evaluate whether the approach is sound without seeing the reasoning and assumptions behind the plan.
Solution
Provide a structured plan summary with goal interpretation, strategy explanation, a subtask checklist with real-time progress, explicit editable assumptions, and resource/time estimates. Keep it concise by default with full reasoning available on expansion.
Real-World Plan Summary Examples
Implementation
Practice in Courses
When to use Plan Summary, and when it backfires
Use it when
- The goal is genuinely ambiguous and several valid approaches exist, so the strategy matters as much as the result.
- The run is long or expensive enough that discovering a wrong interpretation at the end is costly.
- The user holds context the agent cannot have, and a stated assumption is the only place that context can enter.
Don't, or minimize, when
- The task has one obvious approach. Narrating it is ceremony, and users learn to approve without reading.
- The run is cheap and reversible. Let it run and show the result. Reviewing a plan costs more than redoing the work.
- You cannot let the user change anything. A plan presented for approval with no edit path is a loading screen with extra reading.
The trap
The plan with no edit button: a well-structured summary whose only controls are Approve and Cancel. Cancel means starting over, so everyone approves. The assumptions were the valuable part and the interface gave the user no way to touch them, so the plan reads as accountability while functioning as a delay.
Take it into your own product
- 1
Lead with how the goal was read, not what will be done.
'Research competitor pricing' can mean three different jobs. One sentence naming which one the agent chose catches more bad runs than a twelve-item checklist, because that is the step where the misunderstanding actually happens.
- 2
If the user cannot edit it, it is not a plan.
Approve and Cancel is a two-button loading screen. Cancel means losing the work and starting again, so people approve. Make each assumption a field they can change in place, and regenerate the affected steps when they do.
- 3
Assumptions are the payload. Everything else is packaging.
Users cannot judge whether a strategy is optimal, but they can instantly spot 'assumed you meant the UK market' being wrong. Put the assumptions where the eye lands first, not in an expandable section under the checklist.
- 4
Say which step is the last free exit.
Mark the subtasks that cannot be undone and name the point after which stopping costs something. A plan that reads as uniformly safe gets approved with the same attention whether step four sends email to customers or writes to a scratch file.
- 5
Keep the plan alive while it runs.
Replacing it with a progress bar throws away the thing that made it useful. When a step disproves an assumption, show that against the plan, so the user sees the moment the approach stopped matching reality rather than reading it in the summary afterwards.
Save Plan Summary as a Claude skill
Saved skills collect on your dashboard, ready to download one at a time or as a pack for your repo. Once a skill is in place, Claude Code applies it whenever you work on a surface this pattern covers.
Check if your product already has this pattern
Upload a screenshot. We'll tell you which of the 38 patterns your AI interface uses and where the gaps are.
Audit My DesignTake all 38 patterns as Claude skills
Get the whole library as skill files, so Claude applies these patterns while it builds instead of after you catch them in review.
- One skill file per pattern, all 38
- Drop them in .claude/skills and Claude applies them as it builds
- Daily AI UX newsletter (unsubscribe anytime)
More in Trustworthy & Reliable AI
Explainable AI (XAI)
Make AI decisions understandable via visualizations, explanations, and transparent reasoning.
Responsible AI Design
Prioritize fairness, transparency, and accountability throughout AI lifecycle.
Error Recovery & Graceful Degradation
Fail gracefully with clear recovery paths when things go wrong.
