- Updated
- Version
- v1.1
- Sources
- 3 sources
- Status
- Published
context · context limit
Context window
The content capacity available to the model to process a request and produce its response, measured in tokens.
Everyday phrases
- my agent forgets the beginning
- resume a long conversation
Key difference
It is neither the entire visible history nor persistent memory across sessions.
Example
After compaction, you supply a dated decision and its source again before resuming work.
Counter-example
An exhausted message quota does not prove that the context window is full.
Continue with a guide
Related concepts
Sources
- https://ai.google.dev/gemini-api/docs/tokens
- https://ai.google.dev/gemini-api/docs/long-context
- https://platform.claude.com/docs/en/build-with-claude/context-windows
Revision glossary-v1.1:context-window · 2026-09-08
- Sourced content digest
- sha256:898d7c20a7458a7c2922a43e732c11ef28179a401fcd0e9aa35f80b493fe8f27
Understand → Recognize → Choose → Compare
Pack for your agent
Pre-written instruction by SkillCodex — your request is neither sent nor used to adapt this text; no content is generated, and copying executes nothing.
Implement correctly
Use Context window deliberately
# Ground Context window in your agent ## Objective Use “Context window” precisely in code, documentation, and conversations: The content capacity available to the model to process a request and produce its response, measured in tokens. ## Prerequisites - Inspect how the term is used today in the repository, documentation, and prompts. - Identify the neighboring notions already named: Token, Prompt. - Preserve existing correct usages. ## Working definition The content capacity available to the model to process a request and produce its response, measured in tokens. ## Decisive difference It is neither the entire visible history nor persistent memory across sessions. ## Example that counts as Context window After compaction, you supply a dated decision and its source again before resuming work. ## Counter-example An exhausted message quota does not prove that the context window is full. ## Acceptance criteria - Every use of the term matches the working definition. - Edge cases are settled with the decisive difference, not by intuition. - Neighboring notions (Token, Prompt) are named with their own term. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
- Requires · The agentic vocabulary already used in the repository and documentation
Why it works
- The notion is framed by a sourced working definition, not ambient usage.
- The decisive difference settles edge cases reproducibly.
- The example and counter-example bound the term on both sides.
Try next
Map the neighboring notions
# Map the neighboring notions ## Objective Extend the same rigor to the related notions (Token, Prompt) so the project vocabulary stays coherent. ## Checks - List every neighboring notion used without a definition. - Apply its documented decisive difference. - Record the still-ambiguous cases for the next entry. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
sha256:16e316001e46958f3564f220599dcc41f3a89b4c542c9a46a2ab19ba89b213eb
Diagnose a problem
Fix a Context window mix-up
# Fix a mix-up around Context window ## Observed symptom [DESCRIBE THE SYMPTOM HERE] ## Observable checks - Compare the disputed usage with the working definition: The content capacity available to the model to process a request and produce its response, measured in tokens. - Apply the decisive difference: It is neither the entire visible history nor persistent memory across sessions. - Confront the case with the documented counter-example: An exhausted message quota does not prove that the context window is full. ## Possible causes - The term is used for a neighboring notion (Token, Prompt). - The observed form (interface, script, service) is confused with the actual behavior. - An alias (context, context limit) carries a different meaning in the project. ## Bounded fixes - Rename or reclassify only the usages that contradict the working definition. - Document each choice by citing the decisive difference. ## Final verification - Reread every corrected occurrence against the documented example and counter-example. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
- Requires · The agentic vocabulary already used in the repository and documentation
Why it works
- The diagnosis compares usage with documented criteria instead of an opinion.
- Fixes stay bounded to genuinely contradictory usages.
- The final verification runs back through the sourced example and counter-example.
Try next
Anchor the vocabulary in the project
# Anchor the vocabulary in the project ## Objective Turn the clarification into a reusable rule so the mix-up does not return. ## Checks - Add the working definition to the project glossary. - Link every neighboring term to its own entry. - Check new documents against the decisive difference. ## Guardrails - Show the proposed changes before any external action. - Do not publish, send, delete, pay for, or change remote state without explicit authorization. - Preserve unrelated changes and stop if the scope becomes ambiguous. ## Output format - Outcome or verdict. - Files or actions involved. - Checks run and observable evidence. - Remaining blockers or limitations.
sha256:74683bc6191baeefa33582ba3ff4fbed1623af497837ebbed7cb4183256e0b61
Pack digest: sha256:190e2718326ac91edf275a31f414a510bd58b2ba2015eb7201cebf2e4af83afd