Context Engineering for Multi-Agent Systems workflow
Nine markdown files turn a Claude Project into a multi-agent context engine with plan checks and an execution trace
- 292 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- Setup comes down to text instructions and markdown files in a Claude Project; nothing runs on your computer
- The model sees the documents uploaded to the project, so avoid putting personal or confidential data into knowledge
Install
- Works in
Manual install
Create a Claude Project. Take the claude_projects folder from the Denis2054/Context-Engineering-for-Multi-Agent-Systems repository. Paste the content of 01_PROJECT_INSTRUCTIONS.md below the horizontal rule, starting at the CONTEXT ENGINE KERNEL heading, into the instructions box. Upload files 10, 11, 12, 13, 14, 20, 21, 22 and 23 to project knowledge. Do not upload files 30 and 40. Start a new chat for each new goal and request the plan first with PLAN.
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- Setup comes down to text instructions and markdown files in a Claude Project; nothing runs on your computer
- The model sees the documents uploaded to the project, so avoid putting personal or confidential data into knowledge
This is third-party code. Review the repository files before installing.
What it does
The claude_projects folder in Denis Rothman's repository ports the Universal Context Engine to a Claude Project with no code. One instruction block is pasted into the project instructions box, and nine markdown files are uploaded to project knowledge. The model then follows a five-step pipeline: first gate, plan, second gate, execution and trace. The pack contains an engine constitution, planner and governance protocols, a context library and two knowledge packs, marketing and legal. Five agents do the work, and the critic is off by default. The second gate applies a domain topology and forbids, for example, handing work from Marketing to Legal. The planner builds a directed acyclic graph of steps, and the trace shows what ran and what governance rejected. The rest of the repository is a companion to the book: chapter notebooks, LangChain, NVIDIA NIM and DeepSeek-R1 editions and a Gradio interface. Only claude_projects needs installing; the rest is for reading and running in Colab.
Who it is for. For architects and analysts who design multi-agent systems with handoff control and an execution log.
Good fit when
- You want to see how a multi-agent architecture holds up when code is removed and only protocols remain
- You need a document from several domains, such as legal facts and marketing copy, with checks on permitted hand-offs
- You need an execution trace: which steps were planned, what the gate rejected and where text was condensed
- You want to stand up a new domain in an evening using the templates in 30_NEW_DOMAIN_PACK_TEMPLATES.md
Not a fit when
- You need a hard guarantee that rules are enforced: the second gate here is a protocol in prose that a model follows, with no condition in code
- You need real concurrency: waves in the plan show step independence and do not speed anything up
- You need open-ended exploration where the plan changes along the way: the plan is fixed before execution
- You have no Claude plan with Projects: the feature is available on Pro, Team and Enterprise
Example request
PLAN: draft a short press release for a product launch and check the wording from a legal standpointLimitations
Requires Claude with Projects on a Pro, Team or Enterprise plan. Governance relies on a protocol the model mostly follows, without a guarantee. Token counts are the model's own estimates. Retrieval works without embeddings and is sized for roughly 40 documents per knowledge pack.
By the fortieth turn of a chat the kernel sits further from the model, so a fresh chat per run is better. 30_NEW_DOMAIN_PACK_TEMPLATES.md and 40_CONTROL_DECK.md stay local and are not uploaded to the project.
How to disable. Delete the Claude Project or clear its instructions and knowledge. Nothing is installed from the repository, so only downloaded files remain on your computer.
FAQ
How do I build the context engine in a Claude Project?
Create a project, paste 01_PROJECT_INSTRUCTIONS.md below the horizontal rule into the instructions and upload the nine files numbered 10, 11, 12, 13, 14, 20, 21, 22 and 23. No terminal is needed.
Why must 01_PROJECT_INSTRUCTIONS be pasted into the instructions box?
That file has to act at all times. The model finds uploaded files by search, while the pasted instruction is always in context.
What is the second gate, Gate 2?
It checks the plan against the domain topology. Handing work from Marketing to Legal is forbidden, while Legal and Marketing may pass results to General because they report findings back without commissioning work.
How is this edition different from the code notebooks in the repository?
In the notebooks the checks run as code that cannot be bypassed. In the Claude Projects edition the same rules are written as text, and the model itself follows them.
Which agents does the context engine have?
Librarian, Researcher, Summarizer and Writer work by default and Critic is disabled. Researcher comes in three forms: General, Marketing:Researcher and Legal:Researcher. They behave identically and differ only in the knowledge pack and the hand-off rules.
Discussion
No questions or reviews yet.
Related
WorkflowsA 7-rule CLAUDE.md template and session handoff commands so Claude Code keeps numbers and decisions through compaction
An environment for Claude Code that keeps context across sessions via hooks, ledgers and handoffs, orchestrates agents and saves tokens
Multi-agent orchestration for Claude Code: teams, roles, autopilot and task routing across skills and subagents
Workflows for Claude Code: /do, /omo, /bmad-pilot and more commands, with Codex, Gemini or OpenCode writing the code through codeagent-wrapper
A set of /sc slash commands, agents and behavior modes for Claude Code, from brainstorming and research to implementation and tests
A meta-skill for Claude Code that turns a domain description into an agent team and the skills it uses, using six team-architecture patterns
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| Context Engineering for Multi-Agent Systems (this entry) | no | 292 | |
| Context Engineering Kit | no | 1,793 | a plugin set for agents with reflection, review and git work, without a plan engine or hand-off checks |
| oh-my-agent | no | 1,354 | a multi-agent harness with stop hooks and artifact checks, installed as a CLI |
README in short
The main page introduces a book on context engineering and engine editions for different models, from GPT-6 Astra to NVIDIA NIM Nemotron. A no-code Claude Projects edition and a sovereign-AI guide on DeepSeek-R1 are highlighted separately.