TECHNICAL TALK · UPAM3D
Best practices building
agents with Claude Code
Skills · Subagents · Slash commands · MCPs
MaximoFN · 2026
Starting right: the /init command
Generate the context Claude will use throughout the session
$ /init
- Creates the
CLAUDE.mdfile at the project root. - Indexes the structure, dependencies and conventions of the repo.
- That file is loaded in every conversation: it's the base context.
- If the context changes, run it again to refresh it.
Subagent claude-code-guide
The in-house expert that knows how to build with Claude Code
$ @claude-code-guide
- Subagent that knows skills, subagents, slash commands, hooks and MCPs.
- Invoked with
@claude-code-guide. - Ask it and it guides you step by step to build them correctly.
- Avoid reinventing the wheel: it applies proven patterns.
- Useful when starting out and when reviewing your own design.
Skills · “I know kung fu”
Knowledge packaged for a specific task
$ /skill
- A skill teaches Claude how to perform a specific task.
- Made of instructions + scripts + examples inside a folder.
- Manual invocation with
/skill; otherwise the agent calls it when relevant. - Reusable across projects:
read-quality-excel,read-quality-pdf, etc. - If the task repeats, turn it into a skill.
Subagents
Child processes with their own context window
$ @subagent
- Each subagent has its own context: it doesn't pollute the main one.
- Can be launched in parallel to speed up independent tasks.
- Return only a summary → keep the conversation clean.
- Manual invocation with
@subagent; otherwise the agent calls it when relevant. - Ideal for research, heavy reading or noisy tasks.
MCPs · Model Context Protocol
Standard for connecting external tools
- Open protocol to expose external tools to an LLM.
- Losing ground to skills for everyday tasks.
- Current niche: running long-lived / persistent programs or services.
- Cost: they take up a lot of context window when loading all their tools.
- Recommended pattern: a skill that explains to Claude how to use the MCP.
Slash commands
Reproducible task sequences one command away
$ /command
- A command = a well-defined sequence of steps.
- Invoked with
/command— always launched by the user. - Orchestrate skills + subagents + writing files.
- Accept arguments:
/create-quality-doc "Hand Hygiene". - Turn an hours-long process into a single call.
Next steps · Learn
Go deeper into each piece to design them better
- Skills: when to create one, what triggers to define, how to structure it.
- Subagents: isolation criteria and context distribution.
- Slash commands: orchestration, parallelism and error handling.
- Best practices for internal prompting and output validation.
Next steps · Iterate
What gets used improves; what isn't used is excess
- Test skills with real cases and measure where they fail.
- Refine instructions and triggers with each use.
- Remove redundant steps; add only what brings value.
- Share useful skills with the rest of the team.
END OF THE TALK
Thank you
Questions?