PTN—21

Targeted regeneration

Regenerate the sentence that’s wrong, not the paragraph that’s right — redo scoped to a selection, like inpainting.

Whole-response regeneration is a destructive act sold as a convenience. The answer is ninety percent right, one sentence is off, and the regenerate button burns all of it — including the phrasing the person had already silently approved — to fix the ten percent. Targeted regeneration borrows the image world’s answer: select the region that is wrong, regenerate inside it, and leave everything outside untouched. Inpainting, for text and everything else.

The scope is the contract, and it must be visible before anything changes. A selection highlight that marks exactly what is up for revision — and, by omission, exactly what is locked — is what lets the person consent to the operation. Systems that “fix the second paragraph” and also quietly retouch the first have broken the one promise the pattern makes. Outside the scope, not a character moves.

Inside the scope, the replacement is a proposal until the person says otherwise. Show it against the original — side by side or as a marked diff — so accepting is a comparison, not a leap of faith. Then make both directions one click: accept commits it, undo restores the original exactly. When redo is scoped, previewed, and reversible, its cost collapses — and cheap, safe correction is what turns “almost right” from a frustration into a workflow.

The pattern also sharpens feedback. A whole-response regenerate tells the system “something was wrong somewhere”; a scoped regenerate says precisely which sentence failed and implicitly endorses the rest. That is dense, honest signal — the kind a product can actually learn from — produced as a side effect of the person just fixing their document.

Anatomy

Fig. 01 — Targeted regenerationLive specimen

Follow-up email · draft

Thanks for having us out Tuesday — Rivera finished the AC tune-up and both systems are running clean. Per our records you are contractually obligated to remit the outstanding balance forthwith. If anything sounds off before the season turns, call us and we’ll get someone out same-week.

Tap the sentence that reads wrong — the middle one is trying too hard.

A selection-scoped redo: the region is marked, the rest is locked, and the replacement previews against the original.

1Visible scope selection. The region up for regeneration is highlighted before anything changes — scope is consented, not inferred.
2Locked surroundings. Everything outside the selection is untouched; not a character outside the scope moves.
3Replacement preview. The new candidate is shown against the original before it commits.
4One-click accept and undo. Committing and reverting are each a single action, so trying a redo costs nothing.

When to use

Outputs are mostly right, and whole-response regeneration keeps destroying the good parts.
The output has selectable structure — sentences, sections, image regions, code ranges — to scope against.
People are hand-editing around the AI because a full redo costs more than it fixes.
Iteration is convergent: each round should narrow toward done, not reroll the whole draw.

Design considerations

01Make the selection the trigger — the affordance appears on the selected region, where the intent is.
02Guarantee immutability outside the scope, and treat any leakage as a correctness bug, not a nuance.
03Preview the replacement in context; a candidate shown in isolation hides the seams.
04Blend the boundaries — the regenerated span must agree with the grammar, tense, and tone around it.
05Accept an optional instruction with the redo (“more specific”, “cite the figure”), scoped to the selection.
06Keep undo alive after acceptance; scoped redo deserves scoped reversal.

Pitfalls

✕Scope leakage — the system “improves” text outside the selection, and trust in the pattern dies on the first occurrence.
✕Seam artifacts: a replacement that ignores its surroundings, leaving tense shifts and broken references at the boundaries.
✕Commit-first redo, where the replacement lands before the person compares it and undo becomes archaeology.
✕Hiding the affordance in a toolbar far from the selection, so nobody discovers that scoped redo exists.

In the wild

Photoshop: Generative Fill is the pattern’s namesake mechanic: select a region, regenerate inside it, and the rest of the image is untouched by construction.
Midjourney: Vary Region opens an editor to mask part of a finished image and regenerate only that area, optionally with a new prompt for the masked zone.
ChatGPT: Canvas lets people select a passage and ask for a scoped rewrite, revising the selection in place while the rest of the document stands.