Fix the “I asked for one small edit and it rewrote everything” loop with a scope-lock workflow that non-technical users can run in minutes.
A common trust-break moment: you ask for one tweak, then your agent changes tone, structure, or whole sections you didn’t ask for.
This is usually a scope control problem, not a “bad model” problem.
“I just wanted one sentence changed, but now my whole doc feels different.”
You are editing with strict scope control.
Task: Change only [exact sentence/section].
Target file: [path]
Boundaries:
- Do NOT rewrite structure
- Do NOT change tone/voice
- Do NOT edit any other section
Step 1: Show a 3-bullet edit plan only (no edits yet).
Step 2: Wait for my approval.
Step 3: Apply only approved changes.
Step 4: Return:
a) exact changed lines/snippets
b) “unchanged sections” confirmation list.
That invites broad rewriting. Use concrete change requests with explicit “do not touch” constraints.
Old context can push style drift. Start a fresh thread for high-precision edits.
If you skip plan approval, you lose control early. Plan-first dramatically reduces over-editing.
Always ask for changed snippets + unchanged confirmation so you can validate quickly.
For critical docs, do edits in micro-batches: one section at a time. You’ll get better control and fewer surprises.
Don’t paste a full 200-message history and ask for “one quick edit.” Large context increases drift risk and makes precision harder.