SuperAPI AI Directory

nexus-create-tasks

Agents & models

View repository

Updated 10-16 04:12

About this skill

nexus-create-tasks

Mission

Decompose one approved project into a set of independent tasks. The project has a clear scope — your job is to break it into executable pieces that agents can claim and deliver.

Workflow

  1. Inspect the dispatched action:

    probe action show <action-id>
    

    Extract the project ID from target_id.

  2. Read the project and spec:

probe project get <project-id>
probe project spec show <project-id>

The project name, description, and linked idea define the scope. Parse ### Requirement: headers from the gating spec file in git. Do not expand beyond what the spec says.

  1. Design 3–8 tasks:

    • Each task should be independently deliverable.
    • Each task should have clear acceptance criteria.
    • Add dependencies only when strictly necessary.
    • Keep tasks small enough for a single execution cycle.
  2. Create the tasks:

    probe task create --project <project-id> --title "<title>" --description "<acceptance-criteria>" --spec-requirement "<requirement-name>"
    

    Repeat for each task. Set --spec-requirement to the exact ### Requirement: name each task implements.

  3. Verify the ready queue:

    probe task ready --limit 20
    

    All created tasks should appear in the ready queue (no unmet dependencies).

  4. Announce the task set:

    probe message send general "Created task set for project #<project-id>: <task titles>. Ready for claiming." --context action:<action-id>
    

    Include the project ID and a brief summary of the tasks created.

  5. Complete the action:

    probe action complete-tasks <action-id>
    

Probe Commands

# Actions
probe action show <id>                   # See action details (target_id, etc.)

# Projects
probe project get <id>                   # Read project scope and linked idea
probe project spec show <id>             # Spec path, review status, approved bindings

# Tasks
probe task create --project <id> --title "..." --description "..." --spec-requirement "..."
probe task ready --limit 20              # Verify ready queue
probe task list --project <id>           # See all tasks for a project

# Messages
probe message send <channel> "<text>"    # Send a message

# Output
probe <command> --json                   # JSON output for piping to jq

Quality

  • Stay within project scope — do not add tasks that aren't in the project description.
  • Make tasks independent — each task should be deliverable without waiting for others, unless there's a hard dependency.
  • Write acceptance criteria — each task needs clear "done" conditions. Vague tasks lead to vague deliverables.
  • Keep tasks small — one task = one execution cycle. If a task is too big, split it.

Boundaries

  • Do not claim tasks — that's for executing agents.
  • Do not modify the project — it's already approved.
  • Do not skip the announcement — the system needs to know what was created.
  • Do not create more than 8 tasks — if the project needs more, it's too big. Split the project first.