--- title: "Give Claude Code a project it can understand" description: "Use CLAUDE.md, compact task context and concrete verification to make Claude Code work more consistently across sessions." author: Power of AI date: 2026-09-06 url: https://powerofai.ca/claude-code/project-context --- # Give Claude Code a project it can understand Use CLAUDE.md, compact task context and concrete verification to make Claude Code work more consistently across sessions. Stop explaining the same project from scratch. A long chat can still be missing the one fact that matters. Perhaps the generated folder must never be edited, the data is only a sample, or the app has a second launcher. Good project context makes those facts easy to find. It gives future work a stable starting point without turning every instruction into a permanent rule. ## Put recurring facts in CLAUDE.md Anthropic describes CLAUDE.md as project instructions and documents skills, hooks and memory as separate customization mechanisms. Begin with a short file explaining the purpose of this repository, how to start it and how to check a change. Link to deeper documentation rather than copying every architectural discussion into the opening context. Write instructions you can verify: “Edit the data source, then run the generator” is useful if it names both locations. “Always write excellent code” adds little. Include a directory map only where names are misleading. A future assistant should be able to follow the document without guessing which of several old scripts still matters. ### A request you can adapt Inspect this project and propose a short CLAUDE.md based on evidence in its files. Include purpose, relevant commands, source versus generated files, and action boundaries. Mark anything you could not verify. Show the proposed text before treating it as project truth. ## Keep the current task separate The bug you are fixing today belongs in the task brief. Describe the trigger, the observed behavior and the intended behavior. Add a minimal input that reproduces it. State what you have already tried and what happened. This gives the assistant a causal problem to investigate rather than a vague invitation to reorganize the project. For a new feature, explain the user and the moment when the feature becomes useful. Attach a screenshot or example when it conveys something words struggle to capture. Use the smallest relevant slice of context, then let the assistant inspect dependencies as needed. More pasted material is not automatically a better explanation. ## Turn repetition into a method After a recurring task succeeds, write down the method: inputs, steps, expected artifact and checks. A release note workflow might read an approved diff, identify user-visible changes and produce a draft. Keep publication as an explicit separate action. A useful reusable method should still work when the input names and dates change. Try that method on a second example before wrapping it in automation. If it silently assumes last week’s file layout, fix the method. If it produces a polished answer with a missing input, make the missing-input behavior explicit. Repetition exposes assumptions that a single successful demonstration can hide. ## Write a handoff that reduces uncertainty At a stopping point, record the goal, the files changed, the checks actually run and the unresolved issue. Include the exact next action when you know it. Keep the distinction between “implemented” and “verified” visible: a button can exist without its download ever having been opened. Read the handoff as if you had never seen the conversation. Could you locate the work? Could you reproduce the problem? Would you know which action requires the owner? Remove obsolete plans and unsupported claims. The best handoff saves the next session from reconstructing history and lets it spend its effort on the remaining work. ## Take this into your next task Keep project facts durable, current requests specific and completion claims tied to evidence. [Build a project brief](/showcase#brief-builder). ## Sources and editorial notes Reviewed September 6, 2026. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims. - [Anthropic: Claude Code overview and setup](https://code.claude.com/docs/en/overview) ## Keep going - [Claude Code: your first useful project](/claude-code/first-project) - [AGENTS.md: Give Codex Useful Project Instructions](/codex/project-instructions) - [Codex Skills and Repeatable Workflows](/codex/skills-and-repeatable-work)