Project memory, written once
Each project keeps a short guide to its stack, conventions and decisions that Claude reads at the start of every session. No one re-explains the project, so fewer tokens go on context.
Plan first, then build
Claude proposes a plan and we approve it before any code is written. One clear plan replaces many corrective prompts.
One task, one precise prompt
Each prompt names the files, the expected result and how to check it. Precise prompts get it right the first time.
Reuse over regenerate
Shared components and data files mean new pages are assembled, not written from scratch: less output, fewer tokens, consistent results.
The right model for the job
Larger Claude models for planning and tricky code, faster ones for routine edits, with prompt caching for long-running work.
Review the diff, don't start over
We correct small diffs instead of regenerating whole files, which keeps changes small and cheap to review.