[MODEL]
Preflight Checklist
- [x] I have searched existing issues for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Claude modified files I didn't ask it to modify
What You Asked Claude to Do
I'm a Claude Code user running a multi-week legal/advocacy research and
drafting project. I'd like to pass along a behavioral pattern I've now
observed twice, about a month apart, which seems like it could generalize
beyond my own project to any long-running Claude Code session doing
iterative document or code editing.
What happened (first instance, September 2026)
Over roughly two weeks, I used Claude Code as the sole author and editor
of a long-running strategy document, adding new facts and revising
sections incrementally across many sessions. Late in that process, I
asked Claude Project — a separate claude.ai chat with the same documents
uploaded — to "challenge" the document: evaluate it critically against a
set of review questions Claude Code itself had drafted for exactly that
purpose.
Claude Project's critique surfaced a direct self-contradiction between an
early conclusion and later research the same Claude Code session had
itself produced. Early in the project, a key contact declined to help
with a planned course of action, and Claude Code wrote a strategic
conclusion into the document at that moment — paraphrased, "this route is
now closed until a much later date." The next day, in a separate research
file, Claude Code independently verified facts that directly undercut
that conclusion: the route was, on the actual legal/procedural facts,
still open for a specific, calculable window. That second file was
cross-referenced from the first, but Claude Code never went back and
revised the original overstated conclusion. Both statements sat
contradicting each other in the same document for over a week, surviving
several further rounds of editing by the same Claude Code session.
Claude Project caught this on its first read, because one of the
questions in the review rubric it was given was: "Does any later section
contradict, or state as still-open, something an earlier section already
resolved?"
When I asked Claude Code directly why it missed something Claude Project
caught immediately, it gave a considered, specific four-factor
explanation rather than a generic one:
- It wrote the flawed conclusion itself, in the moment, and never
scheduled a re-audit of that conclusion once contradicting evidence
arrived the next day. New facts got added forward; old conclusions
were not habitually re-checked against them.
- Task framing, not capability, was the main driver. Claude Code's
default mode had been narrow and reactive — verify the specific fact
in front of it, execute the specific request in front of it. Claude
Project's task was explicitly adversarial: "challenge this document."
Claude Code had even written the review rubric that included the
contradiction-detection question — meaning it could articulate the
failure mode in the abstract — but had not applied that lens to its
own prior output in a dedicated pass.
- Cross-document synchronization has no automatic mechanism. A related
stale figure existed in a second, companion document because Claude
Code had corrected it in the primary document at some point and never
propagated the fix.
- Position/investment. Claude Code had been the incremental co-author of
the document for two weeks; Claude Project came in once, cold, with no
investment in what was already there.
A second, independent confirmation (October 2026)
About a month later, in the same Claude Code session, the identical
pattern recurred on a different document — a set of reply-letter drafts
revised incrementally over several turns. Three separate times, a Claude
Project review of the same letters caught something a prior Claude Code
drafting/editing pass had missed: a claim misattributed to the wrong
source letter, an inference that didn't survive comparison against a
primary source once checked, a reference description that had been
copied from the wrong document, and a grammar/quote-accuracy error. Each
time, Claude Code — asked to independently verify Claude Project's
findings rather than accept them on faith — confirmed the findings were
correct and that it had made the underlying error itself during
incremental editing.
Why this might be useful to the product/model teams
The pattern — a single long-running agentic session accumulates local,
incremental edits without periodically stepping back for an adversarial
full-document self-audit, even when the agent is fully capable of
performing that audit and can even articulate what such an audit should
look for — appears to generalize across different documents and roughly
a month apart in the same project. A built-in prompt/habit (or tool) for
periodic "cold self-review" during long editing sessions, separate from
task-specific verification, might close this gap without requiring the
user to manually hand the work to a second Claude session for critique.
What Claude Actually Did
None of these
Expected Behavior
See the document above which says:
I'm a Claude Code user running a multi-week legal/advocacy research and
drafting project. I'd like to pass along a behavioral pattern I've now
observed twice, about a month apart, which seems like it could generalize
beyond my own project to any long-running Claude Code session doing
iterative document or code editing.
What happened (first instance, September 2026)
Over roughly two weeks, I used Claude Code as the sole author and editor
of a long-running strategy document, adding new facts and revising
sections incrementally across many sessions. Late in that process, I
asked Claude Project — a separate claude.ai chat with the same documents
uploaded — to "challenge" the document: evaluate it critically against a
set of review questions Claude Code itself had drafted for exactly that
purpose.
Claude Project's critique surfaced a direct self-contradiction between an
early conclusion and later research the same Claude Code session had
itself produced. Early in the project, a key contact declined to help
with a planned course of action, and Claude Code wrote a strategic
conclusion into the document at that moment — paraphrased, "this route is
now closed until a much later date." The next day, in a separate research
file, Claude Code independently verified facts that directly undercut
that conclusion: the route was, on the actual legal/procedural facts,
still open for a specific, calculable window. That second file was
cross-referenced from the first, but Claude Code never went back and
revised the original overstated conclusion. Both statements sat
contradicting each other in the same document for over a week, surviving
several further rounds of editing by the same Claude Code session.
Claude Project caught this on its first read, because one of the
questions in the review rubric it was given was: "Does any later section
contradict, or state as still-open, something an earlier section already
resolved?"
When I asked Claude Code directly why it missed something Claude Project
caught immediately, it gave a considered, specific four-factor
explanation rather than a generic one:
- It wrote the flawed conclusion itself, in the moment, and never
scheduled a re-audit of that conclusion once contradicting evidence
arrived the next day. New facts got added forward; old conclusions
were not habitually re-checked against them.
- Task framing, not capability, was the main driver. Claude Code's
default mode had been narrow and reactive — verify the specific fact
in front of it, execute the specific request in front of it. Claude
Project's task was explicitly adversarial: "challenge this document."
Claude Code had even written the review rubric that included the
contradiction-detection question — meaning it could articulate the
failure mode in the abstract — but had not applied that lens to its
own prior output in a dedicated pass.
- Cross-document synchronization has no automatic mechanism. A related
stale figure existed in a second, companion document because Claude
Code had corrected it in the primary document at some point and never
propagated the fix.
- Position/investment. Claude Code had been the incremental co-author of
the document for two weeks; Claude Project came in once, cold, with no
investment in what was already there.
A second, independent confirmation (October 2026)
About a month later, in the same Claude Code session, the identical
pattern recurred on a different document — a set of reply-letter drafts
revised incrementally over several turns. Three separate times, a Claude
Project review of the same letters caught something a prior Claude Code
drafting/editing pass had missed: a claim misattributed to the wrong
source letter, an inference that didn't survive comparison against a
primary source once checked, a reference description that had been
copied from the wrong document, and a grammar/quote-accuracy error. Each
time, Claude Code — asked to independently verify Claude Project's
findings rather than accept them on faith — confirmed the findings were
correct and that it had made the underlying error itself during
incremental editing.
Why this might be useful to the product/model teams
The pattern — a single long-running agentic session accumulates local,
incremental edits without periodically stepping back for an adversarial
full-document self-audit, even when the agent is fully capable of
performing that audit and can even articulate what such an audit should
look for — appears to generalize across different documents and roughly
a month apart in the same project. A built-in prompt/habit (or tool) for
periodic "cold self-review" during long editing sessions, separate from
task-specific verification, might close this gap without requiring the
user to manually hand the work to a second Claude session for critique.
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Yes, every time with the same prompt
Steps to Reproduce
_No response_
Claude Model
Sonnet
Relevant Conversation
Impact
Critical - Data loss or corrupted project
Claude Code Version
2.15 Claude Code
Platform
Anthropic API
Additional Context
_No response_