Setup AI Context

Automatically fill in the 3 template files in .ai/ by probing the project instead of filling them in manually.

Do not write placeholders, do not guess. Only fill in what can be found in the project.

Run

Step 1 — Run the detection script

bash .claude/skills/setup-ai-context/detect.sh

Read the entire output before proceeding to the next step.

Step 2 — Fill in .ai/commands.md

Based on the SCRIPTS section in the output:

• Find the test command: prioritize scripts named test, test:ci, or direct test framework invocations
• Find the build command: scripts named build, compile, or equivalent
• Find the lint command: scripts named lint, check, or direct eslint/ruff/clippy calls
• Find the typecheck command: scripts named typecheck, type-check, tsc
• Find the dev command: scripts named dev, start, serve
• Find the deploy command: scripts named deploy, or related Makefile targets

For each command: if found → write the actual command to the file. If not present → remove that section (do not leave placeholders).

Step 3 — Fill in .ai/context/architecture.md

Use the README EXCERPT and TOP-LEVEL STRUCTURE sections:

What the project is: taken from the beginning of the README (usually the first line after the title).

Stack: from the STACK, FRAMEWORKS, DATABASE SIGNALS sections.

Module structure: from TOP-LEVEL STRUCTURE — describe the 3-5 most important directories, skip node_modules, dist, .git, etc.

Data flow: if the README describes a flow → write it down. If not → leave this section with the comment <!-- TODO: describe the main request/event flow --> and notify the user.

Module boundaries and External dependencies: if the README does not mention them → leave as TODO and notify the user.

Step 4 — Fill in .ai/context/conventions.md

Use the LINTING / FORMATTING CONFIG section:

• Has biome.json → write: "Formatter and linter: Biome. Run biome check --apply"
• Has .eslintrc* + .prettierrc* → write: "Lint: ESLint. Format: Prettier."
• Has ruff.toml → write: "Lint + format: Ruff."
• Has .rubocop.yml → write: "Lint: RuboCop."
• Has .editorconfig → read that file, extract indent_style and indent_size

Create an Error Handling rule skeleton based on the stack (see Gotchas below). The Naming, Data Access, Testing sections → leave empty with <!-- TODO --> comment and notify the user they need to fill these in.

Step 5 — Report results

Print:

• What was successfully filled in (with specific commands/values)
• Remaining TODOs that the agent cannot fill in automatically (data flow, business logic, naming conventions)

Gotchas

node not available: The script uses node -e to parse JSON. If the project is non-Node and node is not available → the script will skip those sections. That is fine; the output is still useful from other sections.

Monorepo: Running detect.sh from the root will surface too much. If TOP-LEVEL STRUCTURE contains apps/, packages/, services/ → ask the user which sub-project they want to set up .ai/ for, then re-run from that directory.

Sparse README: Many projects have READMEs with only a few lines. In that case architecture.md will have many TODOs — this is expected behavior, not an error.

Error Handling rule by stack:

• Node.js/TS → example with try/catch and typed errors
• Python → example with except Exception as e and logging
• Rust → example with Result<T, E> and ? operator
• Go → example with if err != nil