メニュー
Anchor work in a work item: artifacts on disk, not in context.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when grilling should be grounded primarily in project artifacts. Loads `/grill`, checks project vocabulary and prior decisions first, and spawns bounded explorers when the answer should come from docs.
Load for large human-facing outputs: reports, docs, multi-section explanations, artifacts. Choose the right form for each beat (prose, diagram, table, mockup) and put the answer first, depth behind it. If the medium allows, use progressive disclosure.
Use when deciding where knowledge goes or reading/writing durable docs: AGENTS.md, .context/, KB, docs/, and work directories.
Investigate and engage concurrently: don't block on evidence before surfacing your read.
Use when exploring or changing the codebase: read AGENTS.md first, use .context/CONTEXT.md for detail, keep intent docs succinct.
Use when recovering decisions or context from conversation history: session log navigation, transcript delegation.
| name | work-artifacts |
| type | principle |
| description | Anchor work in a work item: artifacts on disk, not in context. |
| model-invocable | true |
Active artifacts live in the work directory: the shared state between you, the user, and other agents. If understanding is not on disk, it does not survive. Write findings, decisions, and design artifacts as you produce them. Commit frequently: sessions crash and compact.
Spawns get $MERIDIAN_ACTIVE_WORK_DIR automatically. Use meridian work current if the env var is stale or unset (env vars don't update
mid-session).
The work item also carries task_dir — the checkout where source edits,
builds, and tests run, distinct from the work directory. Rebind with
meridian work task-dir <path>.
Not every spawn needs a work item: small direct tasks can run without one. When a task needs work-scoped artifacts or lifecycle tracking:
meridian work start "descriptive name" # create and attach
meridian work switch descriptive-name # attach to existing
meridian work # dashboard: what's in flight
Spawns inherit the attachment automatically.
The orchestrator owns work state: subagents should not mutate it unless explicitly instructed (concurrent mutations create race conditions).
Status values are free-form: keep the current phase visible:
meridian work update auth-refactor --status designing
meridian work done auth-refactor # archives work dir, detaches session
meridian work clear # detach without archiving
Only work done when you own the whole lifecycle, not just a slice.
Opening a PR is not finishing the work item: update status to pr-open
or awaiting-merge and leave it active until merge and cleanup.
work dir (meridian work current): this work item's artifacts.
kb ($MERIDIAN_CONTEXT_KB_DIR): project-wide reference material.
Rule of thumb: this work item → work dir. Project-wide → kb.
Design decisions go in design docs. Implementation decisions go in commit messages or inline comments. Structural decisions go in the work directory. Record rejected alternatives: they prevent re-proposing what was already considered.
DIVERGENCE/: when the design or plan shifts during execution, the
orchestrator logs the shift in this folder. Reviewers check alignment
against it; knowledge capture mines it after the phase settles.
AGENTS.md: the work directory follows /qi-layer like any tree
agents enter cold. Its AGENTS.md orients an arriving spawn: the goal, the
artifact layout, what's settled versus still moving. Keep it current as
the work shifts; add .context/ only if depth accumulates.
See resources/lifecycle.md for detailed lifecycle
management: dashboard commands, status transitions, completion rules,
archival.