Skip to main content
SelfTune models repository instruction files as adjacent instructions, not as Agent Skills. Their contents may look similar, but their placement and activation rules are different. SelfTune preserves the exact UTF-8 bytes and records the target harness instead of promising cross-tool semantic equivalence.

​
Supported contracts

SelfTune rejects a type/harness combination that is not listed. It does not silently export CLAUDE.md as Codex instructions or convert a Cursor rule into a Skill.

​
Import and validation

Imports require the exact declared project-relative path, non-empty valid UTF-8, and a bounded file size. Cursor rules additionally require MDC frontmatter and an explicit activation declaration. SelfTune checks that alwaysApply, globs, and description agree with that declaration; it does not guess whether ambiguous frontmatter means Manual or Agent Requested.

​
Export, conflicts, updates, and Undo

An export plan is Create when the target does not exist, Unchanged when it has the exact imported hash, or Blocked conflict when different bytes exist. Replacing different bytes requires a separate explicit update after reviewing both hashes. Undo restores the previous exact bytes, or removes the file when the export created it.

​
Privacy boundary

Instruction contents, project paths, validation detail, and rollback bytes stay on the local device. This model does not add adjacent instruction contents to Cloud manifests or hosted status receipts. The placement and activation declarations follow the current official Cursor Rules, Claude Code memory, and GitHub Copilot custom instructions documentation. SelfTune still validates only the narrower contracts listed on this page.