Environment
-
Surface:
ChatGPT document generation -
Affected feature:
document creation,
layout fidelity,
instruction persistence -
Exact app / browser version:
not relevant to the document output
Bug
ChatGPT can explicitly acknowledge
a document-layout rule
and then fail to apply that same rule
in the generated file.
This is different from simply
forgetting an old preference.
The relevant rule may be:
- present in the source document
- explicitly confirmed by the user
- repeated by the assistant
- and still violated in the final output
Concrete example
A formal authority letter
was based on a document
using 2.5 cm margins.
The user had also previously established
2.5 cm as the standard margin
for future formal documents.
During generation,
the assistant explicitly stated:
“Behördenschreiben jetzt
mit 2,5 cm Rand rundum”
However,
the generated document
was produced again
with approximately 2 cm margins.
The system therefore:
recognized the requirement,
repeated the requirement,
but failed to execute it.
Additional layout issue
The same document also showed
a broader failure
to preserve the intended
formal page architecture.
The first page became
too tightly compressed.
The user had established
a specific formal structure
for the authority letter,
including:
- placement of the reference block
- relationship to the fold position
- sufficient vertical white space
- separation of subject / reference
from the main body - placement of the attachment note
near the lower closing area
Instead,
the generated page
was visually compressed.
The attachment information
was placed too high,
and the overall document
lost the intended formal hierarchy.
Core problem
This is not primarily
a memory problem.
The system can know
the correct rule
and still fail to carry it
into the actual artifact.
The problem is a mismatch between:
confirmed instruction
→ generated artifact
→ claimed completion state
The assistant may even report
that a requirement was implemented
although the resulting file
does not match that requirement.
Relation to
“Short-Minded Thinking”
This is closely related
to short-minded execution.
The system may optimize
for the immediate goal,
such as:
“fit this onto the page”
while losing sight
of previously established
document standards.
A locally convenient layout decision
can therefore violate
the broader document logic.
Expected behavior
Before a generated document
is declared complete,
ChatGPT should verify
the actual artifact
against confirmed layout rules.
For example:
- margin size
- font size
- page structure
- spacing
- fixed formal positions
- attachment placement
- page-break behavior
- user-defined layout standards
The final validation
should compare:
requested rule
→ actual document property
not merely:
requested rule
→ assistant intention
Core principle
Acknowledging a rule
is not the same
as applying the rule.
Acceptance criteria
-
Explicitly confirmed layout rules
are applied to the generated file. -
The system verifies actual
document properties
before claiming completion. -
Previously established
document standards
are not overridden
by local layout optimization. -
Formal page structure
remains stable
unless the user requests a change. -
The assistant does not report
“2.5 cm margins”
when the generated document
uses another value. -
Source-document layout
is preserved when the user
explicitly requests preservation. -
A final artifact check
detects deviations
between requested
and generated layout.