Output must be directly editable

What it is

The result appears in a bubble you cannot type into. The legal path to fix one wrong character is another prompt, waiting for the model to happen to fix that character. Direct output editability requires the artifact to be a document: a caret can go in, people can change words, numbers, delete a sentence, without passing through the model. The first band of control grain is not a finer prompt. It is putting hands back on the text.

Retry-only equals no control — that is the next diagnosis. This entry only nails: there must not be a membrane of mandatory regeneration between the person and the artifact. It also does not treat whether a later generate round will overwrite those characters — that is subsequent respect.

Why it happens

The object of direct manipulation should be the task object. If the task object is that letter or that table, a bubble is the wrong object: it makes the artifact a line of dialogue, and lines of dialogue are by default uneditable; you can only take another turn. An editor makes the artifact an object you can act on. Changing a character costs a keystroke, not a sample. Sampling also brings new peripheral motion; direct edit, by default, only moves characters that were hit.

There is also a competence mismatch. People can fix typos, fix their own proper names, delete a sentence they would never say. Those acts do not need a model; routing them through a model is slower and more dangerous. Forcing them through a prompt voids a clerical skill people already have.

Studying it

Give a generated draft with a few surface errors (typo, wrong proper name, one extra courtesy sentence). One group can only change via prompt; one group can type in the artifact. Dependent variables: time to finish those surface fixes, whether new peripheral changes appear, whether they abandon the fix. Independent variables: whether the editor is the same surface as the generate bubble, whether save is immediate, whether a second confirm to “take over editing” is required.

Do not use a structural overhaul as the task. A large rewrite will force a prompt or a regeneration and drown the question of whether surface errors should pass through the model.

Where it stops holding

Read-only display cards (weather, a stock price) are not clerical tasks; editability is a fake act. “Direct edit” for image, audio, and video needs the matching object layer (layers, a timeline), not a text caret. Regulated output that must keep a generation trace should mark direct edits, not turn editing off because of it. On mobile, a keyboard makes long-form editing painful; that does not license removing editability, it licenses better selection and voice-replace. This entry does not require grain to switch with expertise. It requires that the document band exist.

Applying it

• Treat generated results as a focusable, typeable document by default, not a chat bubble. If the bubble stays, at least offer “open in document.”
• Changing characters must not trigger regeneration. Save should have document semantics: this piece is this piece.
• Proper names, numbers, sentences the user would never say — let people change them directly. Do not offer only “generate again” beside them.
• Check: leave one wrong proper name on purpose. If fixing it requires another generation, and neighboring sentences move after the fix, the artifact is not yet a document. If the caret can enter and neighbors stay still, direct edit holds. Put “surface errors do not go through the model” in acceptance.

Related

• Same group: L2.06.2 An interface that only offers retry is an interface with no control · L2.06.3 Grain should match the user’s expertise
• Nearby: L2.05 Iterative Revision · L2.13 Control Granularity and Editability · L3.12 Editing and Taking Over Generated Content
• Search terms: direct output editability · artifact as document · bypass the model for typos