Task 3.3: Apply path-specific rules for conditional convention loading

• AA reminder in the pull request template asking reviewers to check migration conventions

  Incorrect

  Review catches problems after the code is written; the conventions should guide Claude while it writes the files.

• BA .claude/rules/migrations.md rule file whose YAML frontmatter sets paths to "**/*.migration.sql"

  Correct

  A path-scoped rule loads only when Claude works on matching files, wherever they are in the tree.

• CA migrations section in the root CLAUDE.md, so that Claude always knows the conventions in every session

  Incorrect

  Always-loaded text spends context in every session, including the many that never touch a migration.

• DA CLAUDE.md file in each of the 25 service folders that repeats the migration conventions

  Incorrect

  Directory-bound files have to be duplicated in every folder and loaded for unrelated work in those folders too.

For conventions that follow a file type across many directories, use a .claude/rules/ file with a paths glob. It loads only when matching files are read or edited, so it costs nothing elsewhere.

• Afrontend.md stays off until a developer turns it on for the session with the /memory command

  Incorrect

  No command switches a path-scoped rule on; matching files trigger it.

• BBoth rule files load only when they are imported from CLAUDE.md with an @path line

  Incorrect

  Files in .claude/rules/ are discovered automatically; they do not need to be imported.

• Csecurity.md loads at launch in every session, with the same priority as the project's .claude/CLAUDE.md

  Correct

  Rules without a paths field are loaded unconditionally at the start of the session.

• Dfrontend.md loads once Claude reads, writes or edits a file that matches its glob

  Correct

  Path-scoped rules load on demand when Claude works with matching files.

• EThe paths value is read as a regular expression, so the pattern has to escape the dots in .tsx

  Incorrect

  paths takes glob patterns such as **/*.ts, not regular expressions.

Rules in .claude/rules/ are discovered automatically. Without paths they load at launch; with a paths glob they load when Claude reads, writes or edits a matching file.

• AIt is a bug; only the Read tool should ever trigger path rules

  Incorrect

  The documented behavior includes recognized Bash reads.

• BThe rule loaded because .sql rules are always global

  Incorrect

  File type does not make a rule global.

• CViewing the file with cat counts as a read, so the rule loads

  Correct

  Path-scoped rules also load when a single file is viewed with a recognized Bash read command.

• DPath-scoped rules load for any Bash command in the session

  Incorrect

  Only reads of matching files trigger them.

A single-file Bash read such as cat or head on a matching file counts as a read for path-scoped rules.

• AThey reload the next time Claude touches a matching file

  Correct

  Path-scoped rules reload when a matching file is touched again after compaction.

• BThey become global and load in every later turn

  Incorrect

  Their paths scoping still applies.

• CThey are deleted from the project and must be recreated

  Incorrect

  Compaction affects context, not files.

• DThey load again only after Claude Code is restarted

  Incorrect

  A restart is not required.

The project-root CLAUDE.md is re-injected after compaction, while nested CLAUDE.md files and path-scoped rules load again on demand.

• AThey load only once Claude works on files inside .claude/rules/frontend/ or backend/

  Incorrect

  Rule subfolders are just organization; loading still depends on each file's own paths field, not on the folder name.

• BThey load only if an index.md in .claude/rules/ imports each subfolder

  Incorrect

  No index or import is needed; the folder is searched as a whole.

• CThey stop loading, because rules must sit directly in .claude/rules/

  Incorrect

  Subdirectories are supported.

• DThey load as before: .md files are discovered recursively, and each keeps its own paths frontmatter

  Correct

  Rules can be organized into subdirectories, and scoped and unscoped rules behave as they did.

All .md files under .claude/rules/ are discovered recursively, so you can organize rules into subfolders such as frontend/ and backend/. Each file's paths frontmatter, or its absence, still decides when it loads.

• AListing the symlink in .mcp.json

  Incorrect

  MCP configuration has nothing to do with rules.

• BYour approval of external imports for the project

  Correct

  Symlinks pointing outside the working directory are treated like external imports and wait for approval.

• CCopying the files instead, since symlinks are not supported

  Incorrect

  .claude/rules/ supports symlinks.

• DRenaming the files so that they end in .rule

  Incorrect

  Rules are markdown files.

Shared rules can be symlinked into .claude/rules/; links that resolve outside the working directory need external-import approval before they load.

• AEscape the brackets, as in 'app/\[locale\]/**/*.tsx', so that they match literally

  Correct

  In a glob, [locale] is a bracket expression that matches one character; escaping the brackets matches the folder's literal name.

• BUse single quotes around the pattern in the YAML instead of double quotes

  Incorrect

  Quoting style changes how YAML reads the string, not how the glob interprets the brackets.

• CWrite the pattern as a regular expression, such as app/\[locale\]/.*\.tsx$

  Incorrect

  paths takes glob patterns, not regular expressions.

• DUse "**/*.tsx" instead, which also covers the localized pages

  Incorrect

  That loads the rule for every .tsx file in the project, not only the localized pages.

Glob syntax treats [ as the start of a bracket expression, so a folder named [locale] does not match itself. Escape the bracket to match it literally; an invalid bracket pattern matches nothing while the rule's other patterns keep working.