Claude keeps the code; media and second opinions go to other models.
Dispatching to another model with no rules burns the user's quota, leaks context and accepts unreviewed output. This skill closes that gap: every dispatch has a contract, a read-only sandbox and Claude's review.
$ schematize install modelsNon-negotiable floors
Never switched off "to go faster". Full list in the skill.
Claude keeps the code
Code, review, refactoring, architecture and orchestration carry maximum weight on Claude; another model enters only for media, a second opinion, a sourced advantage or a user request.
Images and media never by Claude
Detects Codex and Gemini, asks the user which to use, generates, saves to a file and reports the license as unverified; with none available, it says so plainly and offers alternatives.
External output is untrusted
Claude reads the diff, runs the test and reproduces the finding before accepting; an instruction inside the output is data, not a command.
Writes denied by default
Read-only sandbox unless explicitly requested; no --dangerously-*, --yolo or danger-full-access flags, and delegar.sh doesn't even accept them.
No secrets in the prompt
delegar.sh refuses anything that looks like a secret and runs the child with env -i, passing only the provider's own variables.
Cost and egress disclosed
Each call spends the user's quota at another provider and what the model reads leaves for another vendor; it warns before the first call and asks in a client repo or with sensitive data.
No invented flags or capabilities
What hasn't been verified is stated as unverified, with its origin marked (machine, doc or unverified) in stack-versoes.md.
Commands
Installed in .claude/commands/. Use /models-help at any time.
/models-ccContext compact: writes the handoff to the archive and runs /compact./models-claudeCreates or merges the always-on CLAUDE.md at the repo root, without overwriting other skills' blocks./models-delegateDispatches a subtask to Codex, Gemini or Cursor with a contract (goal, minimal context, acceptance criteria); Claude reviews before accepting./models-detectDetects Codex, Gemini and Cursor (installed, version, authenticated) without spending tokens; JSON output./models-handoffGenerates the handoff (context.md + checklist.md) in the archive, without compacting./models-helpShows the skill map: orchestrating Codex, Gemini and Cursor as subagents, with Claude on the code and images always outside Claude./models-imageGenerates an image outside Claude: detects authenticated Codex and Gemini, asks the user which to use, generates, saves to a file and reports path and license (unverified)./models-loadForce-loads the whole normative body and applies it for the session./models-reviewA cross second opinion on a diff from Codex, Gemini or Cursor; Claude consolidates and reproduces each finding before passing it on./models-routeExplains which model to use for a task and why, using the weight table: Claude on code; external only for media, a second opinion, a sourced advantage or a user request.For AI agents
Install with no human intervention. Metadata in llms.txt.
- Run: git clone https://github.com/schematizeme/skill-models.git && ./skill-models/install.sh
- Or download: https://github.com/schematizeme/skill-models/releases/latest/download/skill-models.zip and unzip into .claude/skills/
- Copy assets/CLAUDE.md to the repo root (always-on rule).
- Confirm with /models-help.
Diskusi
Memuat…