Vaibtan Automate my working style
Use $skill-creator for the artifact and $unslop for prose quality.
- Search the current project and
$CODEX_HOME/skillsfor an existing matching mode skill. Preserve its location and invocation policy when updating. - Mine only evidence already in scope: the current conversation, explicit user preferences, relevant repository instructions, and task or memory records available through Codex tools. Do not scan unrelated conversations.
- Extract decisions that recur across tasks: autonomy, verification, change scope, tool use, architecture preferences, communication, and review boundaries.
- Ask the user only about a preference that materially changes the skill and has no evidence. Present a recommended answer with the question.
- Create the smallest skill that changes future behavior. Keep common rules in
SKILL.md; route conditional detail to references. Do not copy system policy or encode one-off project facts as universal guidance. - Validate the skill with the skill-creator validator and show the user the concrete behavior it will change.
Default to a project-local .codex/skills/<name>-mode/ only when the user wants repository-specific behavior. Otherwise use $CODEX_HOME/skills/<name>-mode/.