Guide · Agent skill
How to use the Context Engineering agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The context-engineering skill gives an AI agent a reference framework for deciding what information sits in its working memory, the context window, at each step. It is meant for people designing or debugging a multi-step agent that drifts, repeats itself or forgets constraints. After this guide you will know how to install it, how to ask for it in plain language, and what it will and will not help with.
Context Engineering Guidemvschwarz/openrig · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Get the context-engineering ZIP from the directory page. It contains SKILL.md.
- 2Unzip the skill. Extract the archive so the skill files are in a folder on your machine.
- 3Place it where your agent looks. Put it in ~/.claude/skills/ for personal use, or in the project's .claude/skills/ folder for one project.
- 4Or upload it in the Claude apps. In the Claude apps, go to Settings, then Capabilities, then Skills, and upload the ZIP.
- 5Describe your agent problem. Ask in plain language about drift, repetition, forgotten constraints, handoffs, or what should load at startup.
- 6Apply suggestions as tests. Make one change at a time based on the review and check whether the agent's behavior improves.
What the skill makes the agent do
The skill tells the agent to reason about its context window as a limited resource. It covers what should load at startup versus on demand, how to keep always-loaded files lean, when a single agent is enough versus when subagents make sense, and how to plan handoffs and summarization between steps.
It also explains why agents drift or forget constraints: the source describes attention as a budget rather than free storage, and says reliability can degrade as input grows, often before the window is full. The skill is built around one goal stated in the source: find the smallest set of high-signal tokens that maximizes the chance of the desired outcome.
- Reviews what fills the window: system instructions, tool definitions, retrieved data, message history and tool outputs
- Suggests keeping stable prompt prefixes and append-only context to support caching
- Explains recitation: re-surfacing the current objective near the end of long contexts to limit goal drift
Install the context-engineering skill
The skill ships as a single file, SKILL.md. Download the ZIP from the directory page and unzip it. For a personal install, place the unzipped skill in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.
You can also upload the ZIP in the Claude apps under Settings, then Capabilities, then Skills.
- Personal install: ~/.claude/skills/
- Project install: the .claude/skills/ folder inside your project
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
~/.claude/skills/How to trigger it with a natural request
You do not need a special command. The skill's description lists the moments it applies to, so ask in plain language about the problem you are having. Name the agent behavior you see, such as drifting, repeating or forgetting, and say what you want reviewed or designed.
The skill is not meant for every prompt. It says it does not cover rewording a prompt's tone, choosing which model to pin, or debugging business logic.
- "My agent keeps repeating the same tool calls. Help me figure out what is filling its context."
- "Should this setup load all of our docs at startup, or fetch them on demand?"
- "Review our always-loaded instruction file and tell me what is bloated or contradictory."
- "Plan the handoff between my planner agent and the coding agent so context is not lost."
- "Should this task use one agent or several subagents?"
Examples of what you get back
Typical output is a review or plan that walks through the components of the context window and flags which are missing, stale, bloated or contradictory. The skill asks the agent to check each one against the step it is trying to complete.
Treat the output as a set of questions and design options to test, not a verified fix. The source says its guidance on compaction, subagents, and fixed context limits needs re-verification for current models.
- A list of always-loaded content that could move to on-demand loading
- A note on where durable instructions should sit and where the current objective should be restated
- A recommendation on single agent versus subagents, with the trade-offs named
- A summarization or handoff plan that keeps constraints and failed actions visible
What it needs and how it behaves
The source lists no requirements. The skill runs no scripts and makes no network calls. It is a single SKILL.md file of guidance that the agent reads when the request matches.
Because it is guidance rather than code, nothing about it runs on your machine. Your own review still matters, especially for any file you edit based on its suggestions.
- Requires: nothing listed in the source
- Runs scripts: no
- Uses network: no
- Files: SKILL.md
Tips for getting useful results
Give the agent a concrete symptom and a concrete setup. Describing which step failed and what the agent was given at that point produces better analysis than a general request to improve the agent.
Keep the source's caveats in mind. The skill labels its specific numeric context limits as perishable and says they should be re-verified for the model you use. Check any number against current documentation before you rely on it.
- Name the symptom: drift, repetition, or a constraint the agent forgot
- Share the always-loaded files so the review can judge their size and content
- Ask for the trade-offs, not just a single recommendation
- Verify any numeric limit it mentions against your model's current documentation
What people use it for
- Finding bloated or contradictory content in an always-loaded instruction file
- Deciding which reference material should load at startup and which should be fetched on demand
- Planning handoffs and summaries for a long multi-step agent task
- Choosing between one agent and several subagents for a given workflow
- Diagnosing why an agent repeats tool calls or loses track of its goal
Frequently asked questions
Is the context-engineering skill free?
The source material for this page does not state a price or licence terms for the skill in the body text. The directory lists its licence as Apache-2.0, so check the licence file in the repository before you redistribute it.
What do I need to run the context-engineering skill?
The source lists no requirements. The skill runs no scripts and makes no network calls, and it ships as a single SKILL.md file.
Where do I install the context-engineering skill?
Unzip the downloaded ZIP into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings, then Capabilities, then Skills.
Why does my agent drift or forget constraints?
The skill's explanation is that context is a limited budget and performance can degrade as input length grows, often before the window is full. Relevant content can also get buried in the middle of a long context. The skill suggests reviewing what is loaded and where durable instructions sit.
Does this skill choose which model I should use?
No. The skill says choosing which model to pin is outside its scope, and it does not cover rewording prompt tone either.
Are the numbers in the skill reliable for current models?
Not as fixed facts. The file marks its specific context ceilings and cost figures as perishable snapshots that need re-verification for the model you use, and it says it is not normative for current frontier models.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are designing or debugging a multi-step agent that drifts, repeats itself or forgets constraints, and you want a structured way to decide what loads and when. Pick something else if you mainly need help rewording prompts, choosing a model, or fixing business logic, since the skill explicitly excludes those tasks and treats its numeric limits as out of date until verified.
Download the skill