Working with Claude
As I work through building the three "Ways to" pages, I keep surfacing how to actually do each move with Claude Code. This page gathers those notes in one place, grouped by what you are doing with Claude rather than by which page it sits on. Each one links straight back to where it lives in full. It grows as I go.
Get it clear first
Before Claude builds anything, get the ask sharp.
| Make it ask questionsHave Claude interview you before it builds, so the gaps close first. | clarifying questions sharp answer |
| Ask for optionsGet several genuinely different drafts side by side, then pick. | option generation sharp answer |
| Show by exampleShow the result you expect instead of describing it in words. | worked examples sharp answer |
Plan and explore
When the change is big enough that the approach is the hard part.
Say it another way
Plan mode is a setting where the AI works out an approach and shows it to you before writing any code, so you agree the plan first. It is like asking a builder for the drawings before they start knocking down walls.
| Plan modeSettle the approach with nothing committed, then build it. | plan mode build with AI |
| Explore subagentPush noisy discovery into its own pass so the main thread stays lean. | subagents build with AI |
Make the ask checkable
Pin the result to something you can test, not a feeling.
| Testable criteriaPin the ask to something you could actually test. | specific prompts sharp answer |
| Test firstWrite the test first, then let Claude work until it passes. | test-driven build with AI |
Make the output trustworthy
Get structured, honest output you can rely on without re-checking it by hand.
Say it another way
A schema is just a fixed shape you ask the AI to fill in, like a form with set boxes, rather than letting it write whatever it likes. Handing it the form means nothing gets missed and nothing made up creeps in.
| Ask for a shapeHand Claude a schema to fill, so the shape is guaranteed and nothing creeps in. | schemas and tool_use trustworthy |
| Keep extraction honestWorked examples stop it inventing values when pulling data from a messy source. | worked examples trustworthy |
| Validate and retryValidate the output and hand back the specific error so it self-corrects, not a blind rerun. | validation loops trustworthy |
| Review with fresh eyesA separate instance with none of the generation context catches what the maker glossed over. | independent review trustworthy |
| Guard against false positivesStop a check crying wolf, because one noisy category poisons trust in all of it. | noise control trustworthy |
Stay the driver
Keep your hands on the wheel while Claude does the work.
| Group your fixesDecide if fixes are related or independent before you send them. | batching feedback sharp answer |
| Small steps and diffsShort leash, small steps, check each diff before the next. | incremental review build with AI |
| Permissions and review gatesDecide up front what Claude does alone and what pauses for you. | permissions run for real |
Hold the context
Keep the decisions where you and Claude can both find them.
Say it another way
Project memory, a file called CLAUDE.md, is a note the AI reads every time it starts, holding your project's decisions and rules. It is like leaving a briefing on the desk so each new shift picks up where the last one left off instead of guessing.
| CLAUDE.md, memory and sessionsKeep decisions and state where the next session can find them. | project memory run for real |
| Project memory and rulesLayer CLAUDE.md by scope, keep it modular, and scope rules to where they apply. | CLAUDE.md and rules set up |
Keep it honest
Make a live system surface trouble instead of hiding it.
| Fail wellMake failures surface clearly instead of being guessed around. | fail loudly run for real |
| Fresh-eyes reviewHave a clean instance review what the one that built it cannot see. | fresh session run for real |
Pick a build approach
When the change is small and clear, just build it.
| vibeCodeFor small, clear and low-risk work, just describe it and build. | direct prompting build with AI |
In plain words
In plain words: this page gathers the practical moves for working with Claude, the AI coding tool, sorted by what you are doing at the keyboard rather than by stage. AI here means a tool that writes and changes code when you describe what you want. If you know the moment you are in, getting the ask clear, planning a big change, checking the result or keeping a live system honest, you can jump straight to the help for it.
These notes live across three pages: Ways to get to a sharp answer, Ways to build it with AI and Ways to run it for real.
The soft skills behind them: The soft skills of working with AI.
Why Claude. These notes use Claude Code because it is one of the most widely used AI tools for building software. It is also what we build with day to day. The moves themselves are general and carry across to other capable AI tools.