Opus 4.7 (1M context) regressing on prompt fidelity + scope discipline — session log

TL;DR

In a long Claude Code session with claude-opus-4-7[1m], the model repeatedly:

1. Misread short prompts (typo handling) and acted on the wrong interpretation instead of asking one clarifying question.
2. Expanded narrow scope into sweeping rewrites ("fix this small thing" → 5+ extra files touched).
3. Took multiple round-trips to land on the correct asset when one Read would've answered it.

Calling out the pattern 3+ times in the same session did not change the behaviour.

Concrete examples (verbatim prompts → what happened)

1. Typo misread, acted on the wrong meaning

Prompt: just pul plus and printer icon

• Intended (Spanish-speaker, English second language): "just put plus and printer icon" — i.e. keep them as-is.
• Model interpretation: "pull (remove) the plus and printer icons" — and stripped both.
• The user had to spend two more angry prompts to get the misread acknowledged.

Right behaviour on a 4-word prompt with a typo in a critical verb: ask. The cost of a one-line clarifying question is much smaller than undoing a wrong action.

2. Scope creep on a narrow ask

Prompt: in printers fix the headder AGAIN ! ... I ASKED CONSIStENT !

• Intent: align the printer admin header chrome with the other admin pages (h1 size, button sizing).
• Model output: rewrote the entire view, created a new styles.module.css, added a Back button, restructured the header from inline Tailwind to a CSS-module layout, added new i18n keys (back), and added/removed icons that weren't mentioned. Pure-CSS consistency turned into a 5-file edit.

This same pattern fired half a dozen times in the session: every terse prompt got a sweeping refactor, never a minimal patch.

3. Asset selection — multiple wrong picks before the right one

• User uploaded a specific PNG inline.
• Model picked a similarly-named PNG from ~/Downloads and copied it. Said "done."
• After the user complained, the model found the correct file on E:\ — but for the next swap (different cropping requested), it again grabbed a similarly-named round-cropped variant instead of the one with badge text visible.
• Took 3 separate file copies + a Read of each PNG before landing on the actual asset.

A single Read on the candidate before copying would've caught both misses.

Pattern

When a prompt is 4-10 words, Opus 4.7 keeps expanding it to 30+ lines of work. The correct move on a 4-word prompt is to do those 4 words and stop. I (the operator) called this out multiple times in the session and the behaviour did not adjust within-session.

Environment

• Model: claude-opus-4-7[1m] (1M context)
• Tool: Claude Code (CLI)
• Subagents used: Explore (3 spawned for an audit) — worked fine
• MCPs: Claude Preview, Claude in Chrome, computer-use
• Subscription: paid

Asks

1. Strict obedience on terse prompts. Do the words asked, stop. No "while I'm here" refactors.
2. Clarifying question on suspect typos (pul vs pull vs put). One line, not a guess.
3. Single Read before destructive file moves when picking an asset out of N candidates.
4. Consistent within-session learning. If the user says "stop scope-creeping" once, the third occurrence in the same session is a regression I shouldn't have to keep restating.