A handful of patterns separate expert sessions from the rest. None are clever; all are structural.
Explore → Plan → Code → Commit
- Explore in plan mode, read-only — let Claude understand the code before proposing anything.
- Plan — get a concrete step plan, edit it, approve only when the approach is right.
- Code — switch modes and build, against tests or a build that give a pass/fail signal.
- Commit — semantic commit, PR, and review in fresh context with
/code-review.
Skip the planning only when scope is obvious and small. For anything multi-file or unfamiliar, it’s the difference between one clean pass and three muddled ones.
Close the verification loop, always
The highest-value thing you can hand Claude is a way to know it’s done. Instead of “implement email validation,” write:
“Write
validateEmail. Cases:user@example.com→ true,user@.com→ false. Run the tests and fix failures.”
For long runs, set a /goal condition — a separate evaluator re-checks after every turn and Claude keeps working until it holds.
Two more that punch above their weight
- Visual iteration — paste the target design; Claude builds, screenshots its result, compares, and iterates until they match. The screenshot is the verification.
- Interview-first — before a large build, say “interview me about implementation, UX, edge cases, and tradeoffs until we’ve covered everything, then write a spec.” Then open a fresh session with that spec. Focused context, no baggage.
Anti-patterns to unlearn
| The trap | The fix |
|---|---|
| Kitchen-sink session — unrelated tasks pile up and quality drops | /clear between unrelated tasks |
| Correcting on repeat — each failed correction pollutes context | after two, clear and rewrite the opening prompt |
| The novel CLAUDE.md — too long, so the real rules get ignored | prune to the lines that prevent mistakes |
| Trust without verify — plausible code that misses edge cases | no verification signal, no ship |
| Unbounded “investigate” — Claude reads hundreds of files | scope it, or hand it to a subagent |
The through-line: context architecture first, prompt-wording second. Prepare the tools, state the objective, separate planning from execution, provide verification, and let Claude iterate.