@meir-labs/skill-loop-engineering
v0.1.9
Published
The meirlabs loop-engineering skill — drive a big goal end-to-end as a multi-agent orchestrator. Fan out as many haiku/sonnet/opus subagents as the goal needs, pick the model tier per task, plan first and get the plan approved before any large fan-out, ad
Maintainers
Readme
@meir-labs/skill-loop-engineering
The meirlabs loop-engineering skill — drive a big goal end-to-end as a multi-agent orchestrator. Invoking it is the explicit opt-in to large subagent fan-outs: the agent restates the goal, produces a phased plan (what fans out, which model tier per task, where the human decides), and only executes after the plan is approved. Model tiering is per task — haiku for bulk/mechanical work, sonnet for standard research and drafting, opus for hard reasoning and judging other agents' output. Anything that must be right gets an adversarial-verify pass; anything outward-facing (emails, posts, publishing) is drafted for approval, never auto-sent.
The philosophy behind this working style is written up at meirlabs.com/loop-engineering.
Install
# into the current project (./.claude/skills/loop-engineering)
npx @meir-labs/skill-loop-engineering
# into every project (~/.claude/skills/loop-engineering)
npx @meir-labs/skill-loop-engineering --globalPass --force to overwrite an existing copy. Claude Code picks the skill up automatically when the user says "my goal is…", asks for "a workflow with subagents", or says "be the orchestrator" / "orchestrate this".
The skill is plain markdown — read it before you run it: files/.
