Add Claude guidance page

[jpmckinney]

Add a "Claude" page (docs/agents/) covering task selection, tool and model
choice, context management, prompting, sycophancy, and security/responsibility
for Claude tools in general, with Claude Code specifics where relevant. Link
to official Anthropic documentation throughout.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_0125gpj95EzqfeMYBMWscJae

---

My original outline to compare against:

• Task selection
  • Claude differentiates itself with its tooling (like paper Felix shared); better results when using than when acting as pure LLM; you can tell from the UI feedback when it is doing one or another
  • Offer office hours for anyone who got a bad experience
• Tool choice
  • Chat
  • Design
  • Cowork
  • Code
  • Word/Excel
  • Browser extension (Google docs) [risk!]
  • …
• Model choice
  • make sure Opus is selected, and xhigh for Claude Code
• Context: An LLM itself has no “conversation history”. As your conversation grows, the full transcript is processed each time. There is a limit: the content window. If too long, CC will “compact” the context which might lose quality (give steps to see context if possible) You want to have a high quality context, and a shorter context generally yields better results. Managing context:
  • start new conversations. The downside is that Claude “forgets” everything. Mitigations: Claude.md, /init , projects, skills, custom system prompt - these are essentially the same as simply pasting a boilerplate first prompt (only slight difference is that the default system prompt gives them a bit more priority)
  • “Thinking” models use more context, because they generate text that they then reingest before a final answer. use simpler models for simpler tasks
  • CC: use BTW feature; restore an earlier part of the conversation.
  • explain how memory and projects work with respect to context (eg don’t overdo it on memories or uploaded docs ).
  • Non English reportedly uses more tokens than English
• Prompting :
  • per above about context: it is better to provide upfront as much relevant detail as is practical, to avoid a lot of back-and-forth and changing of direction, etc. which can end up “confusing” the LLM and reducing the quality of its output;
  • That said, dont worry about making your prompts super precise and detailed, or about adding a bunch of memories and skills and using “advanced” techniques from Claude “experts” - commercial products do A LOT of work to improve your prompts, and they are continually iterating their own system prompts ; (give example of that open source tool to refine prompts)
  • With experience, you’ll learn what details are relevant and which you can leave out
  • If you don’t really know what you’re asking for, you can always use one session to refine your idea, then ask Claude to prepare a prompt for a new conversation
  • Give example outputs to follow; Claude is not an Oracle
• General
  • be wary of sycophancy - genAI has been trained to be overly positive, and there’s no reliable way to stop this behaviour
• Tech-specific :
  • Risks: autonomous modes, permissions, secrets/confidential; related: skill loss
  • mitigations:
    • largely mitigated by existing practice (centralized secrets etc, dangerously accept permissions disabled).
    • Mainly about limiting permissions (note that shell can read ssh;
    • if using to debug logs, ensure logs don’t leak security/privacy/confidentiality details - they shouldn’t already, like with Sentry).
    • Add deny rules for pillar/private, salt/private;
    • Never give agents access to production
  • responsibility
    • disclose if unreviewed, via __generated_by etc. tags
    • when it’s okay not to review: PoC, prototype, one time use, non critical code path
  • Admin page: link to Claude GitHub permission page