Assistant editing a passage
claude-opus-4.8 · temp 0
What the user sends
<pasted-artifact-f8458437a6a7>
</pasted-artifact-f8458437a6a7>
THE SEAM: one newline, no marker
↑
What the model returns
newline · no marker
ABSORBED
The afterthought is now part of the returned artifact.
blank line
STILL ABSORBED
A blank line does not significantly reduce absorption in any of the 20 models.
explicit boundary + instruction
SEPARATED
Marked seams fix absorption in 19 of 20 models.
We release SEAM
A paste boundary is invisible to the model.
Unmarked, the user's own words get absorbed into the artifact.
7.7–66.7%
absorption on a bare newline, across 20 models
19–81%
when the comment matches the artifact's register (code edits)
19/20
models improved by explicit boundary markers
1,800 test prompts · 20 models · 6 source datasets
Can LLMs Separate Pasted Artifacts From User Speech? Absorption at Unmarked Prompt Seams
Panthi, Yeamin, Abdelfattah