Correction & Refinement
Vague corrections trigger blind regeneration. No product in the audit prompts for clarification before rewriting — producing a new output that may preserve the exact problem the user was trying to fix.
Problem
AI output is rarely perfect on the first attempt. When users express general dissatisfaction, every product audited responds with blind regeneration — producing a new output without asking what specifically was wrong. The result is regeneration that may preserve the exact structural or tonal problem the user was trying to fix.
Prescription
Three requirements: preserved history (all prior versions accessible and navigable, back arrow returns to original), clarification on vague correction (structured options before regeneration, including a tone dropdown with specific options), and inline text selection (selecting specific text surfaces a contextual toolbar for targeted refinement).
Design decisions
Tone change as dropdown vs. flat button — dropdown communicates that specific options exist rather than leaving users to describe tone freeform. "Be more specific" rather than "Fix fact or accuracy" — more accurate description of the output. Inline selection toolbar appears on text selection, not on hover.
Tradeoffs
Clarification prompts add a round-trip before the user gets a new response — skippable for power users who prefer to rephrase immediately. Version history increases state complexity. Inline selection requires careful hit-testing so the toolbar does not interfere with normal reading.
Pattern definition
Problem
AI output is rarely perfect on the first attempt. When users express general dissatisfaction, every product audited responds with blind regeneration — producing a new output without asking what specifically was wrong. The result is regeneration that may preserve the exact structural or tonal problem the user was trying to fix.
Prescription
Three requirements: preserved history (all prior versions accessible and navigable, back arrow returns to original), clarification on vague correction (structured options before regeneration, including a tone dropdown with specific options), and inline text selection (selecting specific text surfaces a contextual toolbar for targeted refinement).
Design decisions
Tone change as dropdown vs. flat button — dropdown communicates that specific options exist rather than leaving users to describe tone freeform. "Be more specific" rather than "Fix fact or accuracy" — more accurate description of the output. Inline selection toolbar appears on text selection, not on hover.
Tradeoffs
Clarification prompts add a round-trip before the user gets a new response — skippable for power users who prefer to rephrase immediately. Version history increases state complexity. Inline selection requires careful hit-testing so the toolbar does not interfere with normal reading.
Interactive demo
Audit finding
Every product in the audit responds to vague corrections with blind regeneration — producing a new output without asking what specifically was wrong. GitHub Copilot is the only product that explicitly marks an interrupted response as incomplete. No product prompts for clarification before rewriting, meaning the regenerated output may preserve the exact structural or tonal problem the user was trying to fix.
All states
Clarification interceptor
Triggered by vague correction input — short phrases or known vague terms. Presents structured options before generating. Prevents blind rewrites.
Version navigation
Arrow controls with v1/3 format. Back arrow returns to the original response. Full history preserved for the session.
Inline text selection
Selecting any text in the response surfaces a contextual toolbar: Rephrase selected or Flag error. Submits the specific string to the next generation turn.
Tone dropdown
Change tone opens a dropdown with four options — More professional, More academic, More conversational, Simpler. Communicates that specific options exist rather than leaving users to describe tone freeform.