Claude Code Best Practices: Lessons from Real Projects
Practitioner lessons (ranthebuilder.cloud) from shipping three projects with Claude Code. Thesis: “the tool does not matter nearly as much as the person holding it” — domain expertise is the real bottleneck, and AI handles the typing, not the judgment.
Recommendations
- Model selection — Opus for complex work, Sonnet for iteration (claude-opus-4-8); the cost/capability tradeoff also central to agentsys. This is only the model axis; Anthropic later separated out a second control, effort-level — how much work the agent does per request (claude-code-model-effort-levels) — with the rule of thumb to raise effort when Claude does too little (skipped files/tests) but upgrade the model when it’s confidently wrong with good context.
- Process by stakes — BMAD methodology (a spec-driven-development flavor) for projects with real users / security surfaces; plan mode for simpler features. BMAD surfaced 36 user flows + security risks before coding on one project; plan mode missed security issues on another.
- Focused
CLAUDE.mdfiles (<200 lines): stack, conventions, constraints. - Seek community agent-skills to fill blind spots rather than relying on built-ins.
- Specs before code.
Why it routed here
Evidence-from-practice for this wiki’s process themes — spec-driven-development (BMAD), agent-skills, and the agentic-coding-harness view (Claude Code as one component in governance-first, spec-driven development, not a standalone generator). (Caveat: one practitioner’s experience.)
Related
spec-driven-development · agent-skills · agentic-coding-harness · claude-opus-4-8