Editing a generated artefact can take more skill than writing from scratch, because the existing text must be understood first

What it is

Writing twenty lines of a function, the structure is in the hand. Changing forty lines a model emitted means first seeing which abstractions it used, which names are live, which stretch is never called. Many people find: changing is harder than writing from a blank. Comprehension before edit means editing a generated artefact is reading plus revision, not continuing one’s own draft; the skill required can exceed writing that task from scratch.

A low threshold still does not get the edit done, because what stalls is understanding, not the button.

Why it happens

When people write, intent, names, and roads not taken sit in working memory; revision is local. A generated artefact throws those internals away and leaves the finished surface. Editors must reverse-engineer: what this stretch is doing, whether changing one place breaks another, which comments that “look right” are false. Reverse-engineering load rises with length, unfamiliar style, hidden coupling, and soon exceeds “write from a blank in my own habits.”

Generators also like idioms the editor does not use and an extra layer of indirection, adding another rung of reading cost. “We wrote it, you tweak it” then hands the heavier cognitive work to the person who was supposedly saved.

Studying it

Same task, write from blank versus edit a middling generated draft. Dependent variables: time to finish, final correctness, self-reported load, rate of abandoning edit to rewrite. Independent variables: length of the generated draft, distance of its style from the participant’s habits, whether a structure map or comments are provided.

“Abandon edit, rewrite” is the key behaviour: the edit path has already been judged more expensive than authoring. Reporting only “someone edited” will hide rewrites.

Where it stops holding

Short, structurally transparent artefacts in the user’s idiom have low reading cost; editing can truly be cheaper. When the user is the original author (editing yesterday’s generation, still remembered), load drops. Cosmetic polish barely needs understanding the argument; skill demand is low. This entry does not treat settling: people may send because they cannot edit, threshold and skill stacked.

Applying it

• Do not default “generated, please edit” as saved labour. For code and long argument, supply a structure summary, call relations, which blocks can be changed on their own.
• Make “discard and rewrite” a first-class action; do not pin an already-unreadable artefact in the editor.
• On generate, prefer names and structure already in the user’s files, to cut idiom translation while reading.
• Check: the same people, same task, half from blank, half editing a generated draft. If editing is slower or more people delete and rewrite, edit skill already exceeds authoring. Add a structure summary and measure again; the gap should shrink.

Related

• Same group: L3.12.1 Whether a generated artefact is framed as a draft or a finished piece decides whether the UI should lean editor or viewer · L3.12.2 Regenerating after the user has taken over will overwrite human edits unless a merge rule is explicit · L3.12.3 Human edits should be marked, or afterwards no one can tell which stretch came from whom · L3.12.4 The higher the takeover threshold, the more people settle for a result they dislike
• Nearby: L3.13 User Feedback Loops on Generation Quality · L2.12 Iterative Edits and Local Regeneration
• Search terms: comprehension before edit · editing vs authoring · reverse-engineering generated code