skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
robsonrung/rar-skills84 installs

to-tasks

Turn an approved PRD into canonical task drafts and an approved dependency queue. Use when planning implementation tasks; supports optional combined task and model approval.

How do I install this agent skill?

npx skills add https://github.com/robsonrung/rar-skills --skill to-tasks
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill processes project requirements (PRDs) and source code to generate task lists. It identifies and uses local repository commands for verification, which could lead to executing malicious scripts if the repository is untrusted. This risk is mitigated by a required human approval step before tasks are finalized.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

To Tasks

Turn one approved PRD into an executable queue. Each slice fits one focused implement-and-review run and has a checkable acceptance contract. You draft the queue yourself with repository tools. One already accepted task with a complete contract goes straight to implement-and-review; do not manufacture a PRD and queue for it.

Read .ai-workflow/work/<feature-slug>/prd.md; a status other than approved goes back to to-prd. Write tasks.md and one T<N>-<slug>/brief.md per task, as shared/references/work-state.md lays out and references/task-contract.md details. Approval promotes those same files.

Draft the slices

  1. Read the relevant code and find the repository commands that can verify each slice. Never invent a command.
  2. Cut vertical slices: each is one observable path through the layers it needs, ideally a tracer bullet that verifies on its own. Put a behavior-preserving prefactor first only when it removes a shared blocker, and name the characterization test or suite that protects it. A codebase-wide mechanical change uses expand, migrate, and contract slices.
  3. Assign stable IDs T1, T2, …. IDs never change, because approval refers to them: a split gets a new ID, a deletion leaves a gap, and reordering keeps IDs. Dependencies name IDs; no self-blockers or cycles. Mark a slice HITL only when a human decision or authorization truly remains; otherwise AFK.
  4. Write each slice's contract:
    1. The end-to-end behavior, without file paths.
    2. Verified commands and observable acceptance behavior. For stateful work, the meaningful failure sequences, and which adapters need realistic stored-value or integration checks. A non-feature slice may state Test expectation: none; <reason>. Never delete, skip, weaken, or mock away a check to make the contract pass; a wrong contract goes back before approval.
    3. Design: assess shared design, security, interface, and acceptance decisions once across the queue (shared/references/workflow-stage-routing.md) and cite them by decision ID in each task that relies on them. Run design-gate only on an unresolved or changed boundary. Every published slice has a proceed verdict and no open decision.
    4. security-gate with its matched trigger. A test-lens conclusion only when it settled a real test-design choice.
    5. Rollback, review focus, product blockers, and write conflicts, kept separate. Name the owner of each shared contract and what releases its consumers. Serialize slices that share a migration, contract, security surface, or likely write scope unless a merge plan says otherwise. Do not invent a foundation phase just to create parallel work.
  5. For a new feature, mark each task critical or deferred for feature mode. A task is critical when it carries high risk, a shared contract, or a deep security level. Each task names its focused regression checks and the contract tests its dependents rely on.
  6. Check the queue yourself: every dependency exists, there is no cycle, owned paths overlap only under a merge plan, and every task's gates read proceed with no open decision.

Do not call design lenses directly, a panel, a runner, or models-consensus. An open decision or an unanswered material requirement goes to the user before approval.

Approve

Present each ID, title, path, dependencies, type, acceptance summary, gate verdict, security level, and ownership. Resolve material changes first. Ask: "Approve this task queue, or specify changes." Silence or a generated sentence is not approval.

On approval, record the user's response in tasks.md as shared/references/work-state.md describes, set each task's status to pending, and mark the queue approved. Never rewrite the PRD, or a started, accepted, or blocked task; a change to its scope, acceptance, gates, or ownership gets a new ID and approval.

Task approval is not model approval. When the request includes execution and the user wants one decision, have implement-tasks prepare its model plan and present both together; one explicit response then covers both.

The next step is implement-tasks.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/robsonrung/rar-skills/to-tasks">View to-tasks on skillZs</a>