For almost a year, I've been spending more time with Claude than with any other entity in my life. A few months ago, after I'd gushed about Claude Code one time too many, my husband finally declared that he's jealous. Here are some of the things I've been doing to give Claude the support it needs to do its best work.
CLAUDE.md
I add a CLAUDE.md file to the root of every project. This gets loaded in the context for each session and serves as core instructions for Claude. I generally keep this short and reference my PRD/Extended PRD and design documents as appropriate.
PRD and Extended PRD
I start every new project with a PRD (product requirements document): this has my goals, preferred architecture, reference projects, APIs, design direction, anything Claude needs to understand what we're building and why. For larger projects, I'll draft the initial concept in the PRD and then develop an Extended PRD with Claude — architectural details, project phases, the full picture. CLAUDE.md references the extended PRD so the session file stays lean but has depth available when it needs it. I keep the original PRD as the source of truth.
.env.example
In addition to my .env or .env.local, I maintain a .env.example. Claude generally won't read gitignored files, so this is how it knows what environment variables exist in the project and what they're for.
Models
I currently use Opus 4.7 for higher-level tasks, including creating PRDs and CLAUDE.md files (see above), for conversations about architecture, and for the evaluation of Sonnet-generated code that needs an extra look. I use Sonnet 4.6 for implementation -- it's a little more than half the cost.
Design
Claude is a decent designer but tends to the generic, and tends to make poor choices/perform poorly as a critic if not given a strong design direction. For me, Tailwind and shadcn are part of the equation. Luckily, Claude can work with you to develop a design direction, as in this design doc that we developed for a website featuring biographies of powerful women. Once given some structure (and provided the DESIGN.md file is referenced in CLAUDE.md) Claude is able to generate consistent, excellent designs, or at least more excellent than many developers, this one included.
** update**: now that Claude Design is out, I'll try designing there then exporting to Claude Code.
Skills
Claude skills are useful for all sorts of things, but here are some of the recent ones that I've added:
Playwright Skill
Rather than acting as Claude's secretary — conveying screenshots and other data from the outside world for its review - I give Claude access to Playwright, a browser automation library, so it can view the UI it's creating and run basic integration tests before I take a look.
Tavily Skill
Claude has native WebSearch and WebFetch but often needs to be prompted to do so, and Tavily offers a superior web search with generous free credits, so I occasionally add that as a local skill to a project, and ask "Please use Tavily to . . . "
Claude's Local Memory
Claude Code has a persistent local memory system that survives between sessions — preferences you've stated, feedback you've given, decisions you've made. Memory files live in ~/.claude/projects/[project]/memory/ and are local to your machine, not shared with your team (that's what CLAUDE.md is for). Claude writes them automatically when it learns something worth keeping, but does so sparingly. Occasionally I will ask it to remember something.
A Note on MCP Servers
A lot of people use MCP servers to connect Claude to tools. I'm not generally a fan. MCP servers load a lot of information into context and often don't do everything that the API can. Last I looked, even the official Github MCP Server is guilty of this. I'd rather write tools directly, to my specifications.
/clear and /compact
I use /clear freely between tasks. In addition to being expensive, a long session can accumulate errors and wrong turns that Claude starts to treat as fact. By clearing I'm preventing drift. /compact is another way to minimize context while keeping past session memory and often functions as a light code review -- while summarizing, Claude sometimes finds omissions or errors in its recent refactors.
I'm still begging my husband, who runs a university research institute and is tech-cautious, to let me set him up with the AI sandbox his work offers. We'll get there. In the meantime, Claude as a coworker needs all the things we all do -- goals, context, shared memory, and the tools to do the job.
Top comments (0)