Context Window Hoarding
Teams fill context windows maximally with documents, history, and examples without understanding what actually helps, leading to unpredictable behavior, high cost, and debugging nightmares.
- Also known as
prompt stuffingcontext inflationover-contextualizingthe kitchen sink prompt
- First noticed by
ai engineerplatform engineerstaff engineer
- Mistaken for
- thorough AI system design
- Often mistaken as
- comprehensive AI system design
Why it looks healthy
Concrete external tells that make the pattern read as responsible behavior.
- Prompts include extensive relevant-looking context
- The system has access to "everything it could need"
- Teams call the design "thorough" or "comprehensive"
- Demos on canned inputs look strong
Definition
What it is
Blast radius product operations cost
Context windows are stuffed with every potentially relevant document, instruction, and example under the assumption that more context produces better outputs.
How it unfolds
The arc of the pattern
-
Starts
A team adds more context to improve model output quality.
-
Feels reasonable because
In many cases, more relevant context does improve output.
-
Escalates
Context grows uncritically. Old instructions accumulate. Retrieval dumps everything. The prompt is enormous.
-
Ends
Behavior becomes unpredictable, cost is high, debugging requires reading thousands of tokens, and nobody can explain why output changed.
Recognition
Warning signs by stage
Observable signals as the pattern progresses.
EARLY
Early
- Context length grows over time without review.
- Instructions accumulate without removing obsolete ones.
- Retrieval returns everything above a low similarity threshold.
MID
Mid
- Behavior changes after minor context additions that should not matter.
- Token cost is high relative to task complexity.
- Debugging requires reading the full prompt to understand a failure.
LATE
Late
- Nobody knows what all the context instructions do.
- Removing context causes unpredictable behavior changes.
- The team treats the context as brittle and stops modifying it.
Root causes
Why it happens
- More context sometimes helps and the heuristic overgeneralizes
- Context additions are easier than context removals
- No ownership of the full context window
- Cost and complexity of long contexts are underestimated
Response
What to do
Immediate triage first, then structural fixes.
First move
Take one recent production prompt, remove half of it, and measure the quality difference on real cases - most hoarding folds under this test.
Hard trade-off
Accept removing content that might occasionally help, in exchange for behavior that's debuggable and cost-predictable.
Recovery trap
Compressing the context (summarization, chunking) rather than questioning whether the content should be there at all.
Immediate actions
- Audit every section of the current context for demonstrated value
- Remove instructions that cannot be explained or traced to a decision
- Set a budget for context length and require justification to exceed it
Structural fixes
- Version and review context window contents like code
- Measure output quality at different context lengths
- Use dynamic retrieval with relevance filtering instead of static large contexts
What not to do
- Do not assume longer context always produces better output
- Do not add context to fix a problem without understanding why the problem exists
AI impact
How AI distorts this pattern
Where AI-assisted workflows accelerate, hide, or help with this failure mode.
AI can help with
- AI can help analyze which parts of a context window are actually referenced or influential in outputs.
AI can make worse by
- Native mode: the failure mode is in the structure of how AI systems are built. Adding context feels like improvement even when it is not.
Relationships
Connected patterns
Causal flows inside Failure Modes, and related entries across the site.
Easy to confuse with
Nearby patterns and how this one differs.
-
Prompt chaos is prompts changing without version control. Context hoarding is prompts growing without evidence.
-
RAG-without-ground-truth adds retrieval without validation. Context hoarding adds content without validation.
- Adjacent concept Legitimate context engineering
Legitimate context engineering measures what helps. Hoarding measures what fits.
Heard in the wild
What it sounds like
The phrase that signals the pattern is about to start, and who tends to say it.
Let's just add the full documentation as context.
Said byai engineer or product manager
Notes from practice
What experienced people notice
Annotations from engineers who have worked this pattern before.
- Best momentWhen intervention actually changes the trajectory.
- Before a bloated context becomes the default and spreading pattern
- Counter moveThe specific action that breaks the pattern.
- Measure what context contributes before adding more of it.
- False positiveWhen this pattern is actually the correct call.
- Rich context can genuinely improve output. The failure mode is adding context without evidence it helps.