What our web team learned using Claude Code for a month
Key Points
- Best for knowledge retrieval and well-specified tasks
- Session/context management and skills are fragile
- Use MCPs, tests, and git worktrees to scale safely
Summary
Our web team used Claude Code intensively for a month to understand where it helps and where it needs human support. We found clear wins in knowledge retrieval, executing well-specified tasks, and parallelizing routine work. Major limitations remain around session/context management, skills consistency, and the need for human engineering judgment. Below are practical takeaways and recommendations you can apply immediately.
Key Points
-
Strengths
- Knowledge retrieval: effective for exploring unfamiliar parts of a codebase and unblocking tasks that require system context.
- Well-specified tasks: produces best results when given clear requirements and acceptance criteria; use Plan mode for ambiguous features.
- Strong repo patterns help: linting, type checks, tests, and MCPs (e.g., Linear, Sentry, Figma, Graphite) guide Claude toward reliable solutions.
- Parallel development: use Git worktrees plus tools like Conductor or Claude Desktop to safely delegate and parallelize changes.
-
Limitations and gotchas
- Session onboarding: Claude starts fresh each session—add a concise CLAUDE.md system prompt to reduce repeated instructions.
- Skills are brittle: pre-built skills are useful but often need manual invocation to be consistently applied.
- Context window decay: long sessions degrade output quality; use /clear or export progress to markdown, clear, then reload to reset context.
- Always review: LLM output can be confidently wrong—maintain engineering standards, code review, and tests.
Recommendations (practical)
- Keep system prompts short and focused; store onboarding in CLAUDE.md.
- Use Plan mode to surface clarifying questions for complex features before coding.
- Enable MCPs and continuous checks (lint/type/tests) so the agent can self-validate.
- Parallelize work with Git worktrees and local branches to let AI work concurrently without blocking developers.
- After finishing a discrete task, export state, /clear the session, and re-import to avoid context drift.
- Invoke skills explicitly when you need a specific recipe applied, and always run tests and reviews before merging.
Bottom line
Claude Code is already improving developer workflows for routine, well-scoped tasks and knowledge discovery, but requires disciplined context management, concise onboarding, and human oversight to be production-safe.