- name
SKILL.md
Le fichier du skill reste dans le dépôt de l'éditeur.
Collez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add jamditis/claude-skills-journalism --skill context-engineering-fundamentals -a cursor -yCollez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add jamditis/claude-skills-journalism --skill context-engineering-fundamentals -a claude-code -yCollez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add jamditis/claude-skills-journalism --skill context-engineering-fundamentals -a codex -yCollez la commande copiée dans un terminal. Elle installe ce SKILL.md dans cet agent.
npx skills add jamditis/claude-skills-journalism --skill context-engineering-fundamentals -a grok -yEnvoyez le message copié dans un chat Grok Bot. Le Bot récupère ce SKILL.md sur GitHub et l'ajoute aux skills que partagent tous vos Bots.
Install this skill: https://github.com/jamditis/claude-skills-journalism/tree/HEAD/dev-toolkit/skills/context-engineering-fundamentalsDe l'éditeur
Context engineering is the practice of managing an LLM's limited attention budget. Use this skill to keep instructions, evidence, and state available during long work.
Context windows are constrained by attention mechanics, not only token capacity. A large context limit does not guarantee equal use of every item.
The "Lost in the Middle" experiments show that retrieval quality can change with information position. The result depends on the model, task, context length, and number of documents.
Implication: Keep critical constraints easy to find and repeat them near the decision that uses them. Do not assume position alone predicts recall.
Information in the middle of long context gets lower attention weight.
Mitigation: Structure with explicit sections. Put critical constraints at start AND end.
Errors compound when incorrect information enters context (from tool outputs, summaries, or earlier mistakes).
Mitigation: Validate intermediate outputs. Don't blindly trust previous responses.
Irrelevant information forces attention allocation away from relevant content. Models can't "skip" irrelevant context.
Mitigation: Be selective about what goes into context. More isn't better.
Multiple task types or conflicting instructions create ambiguous responses.
Mitigation: One task per interaction when possible. Clear task boundaries.
Contradictory information from multiple sources causes derailing conflicts.
Mitigation: Resolve contradictions explicitly before asking Claude to use the information.
Do not use a fixed token threshold to decide when context is reliable. Measure retrieval and reasoning quality on your own model and task. Test representative evidence at several positions, then compare the result before and after summarization.
Compress only when the measured result or the agent's behavior shows a problem. Preserve exact constraints, decisions, source links, unresolved questions, and verification evidence.
Do not rely on the agent to remember across turns. Write important state to files, but agree the path with the user first. Prefer a gitignored workspace so you never overwrite project-owned content:
With the user's approval, after each major step write progress to an agreed scratch file (for example a gitignored PROGRESS.md or a path they choose)
Before starting, read that file back to restore context
Filter irrelevant context before loading:
Instead of: "Here are all 50 files, find the bug"
Do: "Here are the 3 files involved in the error"
Summarize while maintaining signal:
Instead of: Full 1000-line file
Do: Key functions and their signatures, with context on the specific area
For complex tasks, use subagents with focused contexts rather than one agent with everything.
Manages attention and evidence in long agent sessions. Use for lost instructions, dropped evidence, or large multi-agent contexts.
---name: context-engineering-fundamentalsdescription: Manages attention and evidence in long agent sessions. Use for lost instructions, dropped evidence, or large…---# dev-toolkit/skills/context-engineering-fundamentals/SKILL.mdFAQ
Téléchargez dev-toolkit/skills/context-engineering-fundamentals/SKILL.md auprès de l'éditeur, ou copiez le chemin et ouvrez le dépôt. PlugMCP répertorie le skill et n'héberge pas le fichier.
L'éditeur le maintient sur https://github.com/jamditis/claude-skills-journalism.
Claude, Cursor, Codex et Grok Build peuvent lire un SKILL.md que vous ajoutez depuis le dépôt de l'éditeur. Grok Bot l'installe lorsque vous envoyez dans un chat le lien du dossier GitHub du skill.