Lobster Subagent Dev

Subagent-driven development. Dispatch fresh subagent per task with 2-stage review (spec + quality). Cost-aware model routing. From Superpowers by Jesse Vincent.

Security Scan

The skill's instructions, metadata, and requirements are internally consistent with its stated purpose (coordinating short-lived subagents for a two-stage review workflow) and it doesn't request extraneous credentials, installs, or file access.

Purpose & Capability

Name/description (subagent-driven development with 2-stage review) match the SKILL.md content. The skill is instruction-only and requires no binaries, env vars, or installs, which is proportionate to a process/workflow orchestration skill.

Instruction Scope

SKILL.md contains high-level orchestration steps (extract tasks, dispatch implementer/spec/quality reviewers, handle statuses). It does not instruct the agent to read unrelated files, access secrets, call unexpected external endpoints, or perform system-level operations. The instructions are procedural but remain within the stated coordination scope.

Install Mechanism

No install spec and no code files — lowest-risk model for an instruction-only skill. Nothing will be written to disk or downloaded by the installer.

Credentials

The skill requires no environment variables, credentials, or config paths. There are no requests for unrelated secrets or broad access, which is proportional to the described functionality.

Persistence & Privilege

always is false and there is no indication the skill modifies other skills or system-wide settings. Autonomous invocation is allowed by platform default, which is expected for an orchestration skill; the skill does not request elevated persistence or privileges.

Assessment

This skill is internally coherent and low-risk as provided: it only defines a workflow for dispatching short-lived subagents and requests no secrets or installs. Things to consider before enabling: (1) subagent autonomy — ensure you trust automated dispatching and review the agent's authorization policy for launching subagents; (2) cost/model routing — confirm which models will be used for each task type to avoid unexpected expense; (3) validation — because the process delegates work, make sure human review gates and test-run policies are enforced so automated subagents cannot merge or deploy changes without oversight. If you want stricter controls, restrict the skill to user-invocable only (no autonomous runs) or configure model/endpoint allowlists and rate/cost limits.

Lobster Subagent-Driven Development 🦞

Fresh subagent per task. Two-stage review. Ship with confidence.

Inspired by Superpowers by Jesse Vincent.

Core Principle

Fresh subagent per task + 2-stage review = high quality, fast iteration

Subagents never inherit session context. You construct exactly what they need. This preserves your own context for coordination work.

The Process

1. Read Plan → Extract Tasks

• Read implementation plan once
• Extract ALL tasks with full text
• Create TodoWrite tracking

2. Per Task: Implementer → Spec Review → Quality Review

[Dispatch Implementer]
  ↓ (if questions → answer → re-dispatch)
[Implementer: implement + test + commit + self-review]
  ↓
[Dispatch Spec Reviewer]
  ↓ ❌ → Fix → Re-review
  ↓ ✅
[Dispatch Code Quality Reviewer]
  ↓ ❌ → Fix → Re-review
  ↓ ✅
[Mark task complete]
  ↓
[Next task...]

3. Final Review

After all tasks complete:

• Dispatch final code reviewer for entire implementation
• Verify all requirements met
• Run full test suite

Model Selection

Task Type	Model	Why
1-2 files, complete spec	Cheapest	Mechanical
Multi-file integration	Standard	Needs judgment
Architecture/review	Strongest	Needs expertise

Agent Status Handling

Status	Action
DONE	Proceed to spec review
DONE_WITH_CONCERNS	Read concerns, then decide
NEEDS_CONTEXT	Provide context, re-dispatch
BLOCKED	Change model / split task / escalate

Never

• Skip reviews (spec OR quality)
• Start quality review before spec passes
• Let implementer self-review replace actual review
• Move to next task while review has open issues
• Dispatch multiple implementers in parallel (conflicts)