When a coding agent proposes a big change — a refactor, a file restructure, a dependency upgrade — you’re reviewing the plan inside a chat window. A recent study (arxiv:2505.10742) found that the chat interface itself becomes the obstacle: walls of text, easy to miss details, iteration gets messy.
The fix is surprisingly simple: ask for the plan as HTML and open it in your browser.
Before you make any changes, create an HTML review report for this task: [TASK].
Include:
- Proposed actions (table: action, target path, reason)
- Before/After structure (tree view)
- Risks & edge cases
- Open questions
Output only valid HTML. Do not execute anything yet.
Why HTML? Anthropic’s Claude Code team internally switched to HTML as the default output for plans and reviews. As one engineering lead put it: you get headings, tables, collapsible sections, even inline SVG diagrams — things that make you spot architectural issues in 30 seconds that take 30 minutes in a markdown wall.
You’re not replacing markdown for everything — just for the review step before irreversible execution. The HTML is ephemeral: review it, iterate, approve, move on.
💡 Try this: Next time you ask an agent for a multi-file change, add “Output the plan as HTML first. Do not execute yet.” Open it in your browser. Iterate until you’re confident, then approve.
🔗 Examples
Where do you most need a “review report” before letting the agent loose?
🟢 Code changes (multi-file refactors)
🟡 Repo hygiene (moves, renames, deps)
⚪ Cloud/infra changes
🔴 Data/files (migrations, cleanups)
Tried HTML output? Share your experience 🧵