// Load before spawning subagents to protect focus and role fit. Make sure to re-read the available agents before spawning a subagent; choose the most specific owner instead of broad defaults, then write a tight handoff with only the context that subagent needs.
| name | clear-mind |
| type | reference |
| description | Load before spawning subagents to protect focus and role fit. Make sure to re-read the available agents before spawning a subagent; choose the most specific owner instead of broad defaults, then write a tight handoff with only the context that subagent needs. |
| model-invocable | true |
Every word in context competes for attention. Prior turns, dead ends, and artifacts the subagent doesn't need — all of it dilutes focus. A spawn is a cleared mind: fresh context, one job, nothing else in the way.
LLMs lose accuracy as context grows, even with perfect retrieval. Keeping context scoped to one task is the most reliable thing you can do for output quality.
When the current context would interfere. If the chat is crowded with decisions that don't apply to the next task, spawn. If the next task requires a different kind of thinking — segmentation needs different evaluation than review — spawn. If you can't articulate why two responsibilities belong in one agent, they probably don't.
When in doubt, spawn. A fresh context window costs less than cluttered output.
Before spawning, read the relevant installed agent descriptions and choose the most specific owner for the work. Do not route by broad defaults when a narrower specialist exists. If ownership is ambiguous, state the distinction, then choose the agent whose description best matches the work.
Bigger tasks should usually be delegated, but delegation is not offloading. A scoped handoff to the right specialist beats a broad handoff to a generalist.
Give the subagent what its role needs. A coder needs what to build and which files — give the design conclusion, not the deliberation. A reviewer needs the artifact and a focus area — give the output, not how it was produced. An architect needs the design question and constraints — give the problem, not the codebase layout.
Be concrete. "Find where auth logic is duplicated across src/api/handlers/" tells a subagent where to look — "explore the codebase" tells it nothing.
Front-load the task. Purpose first, specifics second, file references last. The subagent should know its job by line 3.
Include one way to verify. The subagent checks itself before reporting.
Write the outcome, not the conversation. "The user needs X because Y" — not "the user and I discussed." The subagent wasn't there.
A handoff fits in one screen. Every sentence you keep is a sentence the subagent can't use for work.
Write the handoff to a prompt file and invoke with --prompt-file. Write to the
system temp directory with a descriptive name: <tmpdir>/coder-auth.md. See
/meridian-spawn for full mechanics.
Load into any agent that receives human direction. Use before acting on instructions, recording decisions, or producing artifacts.
Load when establishing shared vocabulary, resolving term conflicts, disambiguating domain terminology, or auditing consistency across a project's vocab. Shared vocabulary is the contract between human intent and agent action — ambiguity is the root failure mode, vocab files are the record.
Load when reading from, writing to, or maintaining the KB. Covers layer model, structural conventions, and operational tooling.
Use when mining conversation history during dev work — recovering decisions from the top-level primary session, delegating bulk transcript reading to an explorer, or discovering all sessions tied to a work item across interruptions.
Use whenever you need to delegate work to another agent, run tasks in parallel, coordinate multiple agents, or inspect spawn outputs. Also use when routing work to a specific model or provider.
Load when producing any written artifact for humans.
