Context
6 current practices, and 7 retired.
StandardContextCX-001Treat context as a scarce budgetDeliberately curate everything the model sees on each call rather than stuffing the window.
Reviewed Aug 2026WatchContextCX-002Don't fill the window past roughly 40%Quality degrades as the context window fills. Keep usage low rather than treating the advertised limit as usable.
Reviewed Aug 2026ContextCX-003Keep identifiers, load the data at runtimeInstead of pre-loading everything, hold lightweight references — file paths, stored queries, links — and fetch on demand via tools.Reviewed Aug 2026StandardContextCX-004Write state to files before the window fillsHave the agent persist plans and intermediate results outside the context window, and re-read them as needed.Reviewed Aug 2026StandardContextCX-005Give agents a repo-level instruction fileA plain Markdown file at the repo root telling agents how to build, test and contribute. Nest one per package in monorepos — closest file wins.
Reviewed Aug 2026ContextCX-006Compact rather than resetCondense older turns into a summary as the window fills, keeping recent detail intact.Reviewed Aug 2026
Reviewed Aug 2026WatchContextCX-002Don't fill the window past roughly 40%Quality degrades as the context window fills. Keep usage low rather than treating the advertised limit as usable.Reviewed Aug 2026ContextCX-003Keep identifiers, load the data at runtimeInstead of pre-loading everything, hold lightweight references — file paths, stored queries, links — and fetch on demand via tools.Reviewed Aug 2026StandardContextCX-004Write state to files before the window fillsHave the agent persist plans and intermediate results outside the context window, and re-read them as needed.Reviewed Aug 2026StandardContextCX-005Give agents a repo-level instruction fileA plain Markdown file at the repo root telling agents how to build, test and contribute. Nest one per package in monorepos — closest file wins.Reviewed Aug 2026ContextCX-006Compact rather than resetCondense older turns into a summary as the window fills, keeping recent detail intact.Reviewed Aug 2026Retired in this pillar
RetiredContextRE-002Tell the model to think step by stepChain-of-thought prompting to elicit reasoning — the most repeated prompting tip of 2023.
Retired Dec 2024RetiredContextRE-003Split documents purely to fit the context limitChunking documents only because the window was four to eight thousand tokens.
Retired 2024RetiredContextRE-005Put a retrieval pipeline in front of every queryRAG as the default architecture regardless of corpus size or freshness needs.Retired 2025RetiredContextRE-007Add few-shot examples inside agent loopsAdding uniform examples to an agent's context to improve output quality.Retired Jul 2025RetiredContextRE-008Scrub failed actions from context to keep it cleanRemoving errors and wrong turns from history so the agent isn't distracted by them.Retired Jul 2025RetiredContextRE-011Treat prompt engineering as the primary disciplineOptimising a single prompt as the main lever on quality.Retired Sep 2025RetiredContextRE-013Assume the full advertised context window is usableTreating a stated window size as the amount of context a model actually reasons over evenly.Retired 2025
Retired Dec 2024RetiredContextRE-003Split documents purely to fit the context limitChunking documents only because the window was four to eight thousand tokens.Retired 2024RetiredContextRE-005Put a retrieval pipeline in front of every queryRAG as the default architecture regardless of corpus size or freshness needs.Retired 2025RetiredContextRE-007Add few-shot examples inside agent loopsAdding uniform examples to an agent's context to improve output quality.Retired Jul 2025RetiredContextRE-008Scrub failed actions from context to keep it cleanRemoving errors and wrong turns from history so the agent isn't distracted by them.Retired Jul 2025RetiredContextRE-011Treat prompt engineering as the primary disciplineOptimising a single prompt as the main lever on quality.Retired Sep 2025RetiredContextRE-013Assume the full advertised context window is usableTreating a stated window size as the amount of context a model actually reasons over evenly.Retired 2025