· AI · 2 min read
Notes on Using Claude: Tips & Architecture
Key takeaways and best practices for using Claude and multi-agent architectures, including prompt design, skills, and error handling.
AI landscape is quickly moving, as well as Claude is rapidly evolving.
This is a note to catch up as to how better we can use Claude.
Since these could be quickly updated, I will go with bullet point rather than sentences.
Take away bullet points:
description What + When + Negative trigger
negative tigger: prefer “Do NOT use for…”
skills description: MAX 1024 characters
SKILL.md length: MAX 5000 words
Subagents does NOT share the memory with the coordinator; It’s in an isolated context, only information is given explicitly in the propmpt
Do NOT rely on prompt; You have to enforce mission or secrurity critical rules via hooks and prerequisite gates
With MCP, when selecting a tool, descriptions are the primary trigger for Clause to decide which ones to pick; better descriptions are the key
For each subagent, scope the number of tools by giving only 4 - 5 MCP tools for the task; otherwise decision making process would degrade
instruction hierachy:
- ~/.claude/CLAUDE.md: top/user level
- .claude/CLAUDE.md: project level
- yourproject/subfolder/CLAUDE.md: directory level
be mindful, ”~/.claude/CLAUDE.md” cannot be shared via github repo automatically
path-specific rules: rather than directory or project base, it can be applied for specific file or path
globs:
- "**/*.test.tsx"
instructions: |
Use React Testing Library.
Avoid enzyme.- when reviewing code, independent session is better than doing it continuously in the same session where it wrote the code
- be explicit: “be conservative” or “give me high-confidence finding” does not decrease low quality or false positive
- give it 2 - 4 example: this wouls increase ouput quality
- do not summarize long context, isorate important facts and findings and attach it every prompt; longer the context, blurer their attenstion
- key facts and datas should be at the begenning, not the middle
- be mindful error validation; give explicit, concreate steps to propagate error, else risk skipping the errors
- design and architect for a human escaration process; when a chat agent is dealing with a customer, do not rely on ai’s self-report or some subjective criterior to trigger human agent; rely on an explicit, objective trigger like “request human agent” button, or “policy”, or “infinite loop”
References:
I want to become a Claude architect yohey-w / multi-agent-shogun Copare Shogun Skills Architecture with Clause Official Skills Guide (Japenese)
