Lost in the middle
Evidence can become harder to retrieve when it sits between other material.
Use this skill when
- A task spans many steps, files, tools, or agents.
- An agent loses an instruction or misses supplied evidence.
- You must decide what to preserve, compress, or remove.
- You need a repeatable way to test long-context quality.
Core rule
A large context limit does not guarantee equal use of every item.
Measure retrieval and reasoning quality on the model and task you use. Do not rely on universal token thresholds.
Five failure patterns
Context poisoning
An early error can shape later reasoning when nobody verifies it.
Context distraction
Irrelevant material competes with the evidence that matters.
Context confusion
Mixed tasks and unclear boundaries make the next action ambiguous.
Context clash
Conflicting sources or instructions need an explicit decision before use.
Practical workflow
- Keep critical constraints near the decision that uses them.
- Verify intermediate outputs before later steps depend on them.
- Remove irrelevant material before you add more context.
- Write durable state only to a user-approved path.
- Compare task quality before and after summarization.
Install
Install the dev-toolkit plugin or copy this skill with the standards-based skills CLI.
/plugin install dev-toolkit@claude-skills-journalismnpx skills@latest add jamditis/claude-skills-journalism \
--skill context-engineering-fundamentals --agent codex --copy -g -y