mike-diff/agent-team

Ask numbered clarifying questions before proceeding with a task

Implement a feature end-to-end with external verification gates. Implements directly by default, delegates to subagents when work genuinely decomposes. Supports spec-backed runs, autonomous multi-phase sweeps, and an unattended mode (--unattended) for headless/CI runs.

Explore a rough idea through conversation and parallel background research. Produces a validated plan with an adversarial check. Hands off to /spec or /dev.

Decides between the Claude Code /goal and /loop commands for a rough task, then emits a copy-paste-ready command block. Use when the mechanism is NOT yet chosen, for asks like should I use /goal or /loop, is this a goal or a loop, what is the best way to run this until it is done, set this up to run autonomously until X, or keep working on this unattended until it converges. Picks /goal (verifiable end-state provable from the transcript), /loop (cadence or waiting on an external clock), BOTH in sequence (terminal deliverable plus ongoing watch), or NEITHER (one-shot, needs human judgment, or unobservable). Triggers on keywords goal, loop, autonomously, until done, unattended, converge. Does NOT claim plain set up a recurring task / cron / poll every N minutes (that is loop-patterns) or schedule work that survives the session (that is schedule); and does NOT fire once the user has already typed /loop or /goal.

Fetch a GitHub issue, explore the codebase, ask clarifying questions, and produce an implementation plan. Stops before building.

Recommended /loop patterns for the agents shipped in this repo (explorer, implementer, reviewer, qa, skill-author). Activates when the user asks about /loop, polling, watch-mode, periodic tasks, or autonomous iteration.

Build a comprehensive understanding of the codebase. Maps tech stack, architecture, and patterns. Findings persist through project memory.

Enumerate all native Claude Code tools, modes, and configuration primitives available in the current session

Code review patterns for spec compliance, quality verification, and reference integrity. Use when reviewing code changes, verifying implementation against requirements, or assessing code quality. Covers three-pass review method, severity levels, scope creep detection, security checks, and stale reference detection.

Create a new Agent Skill in .claude/skills/ following the agentskills.io specification. Use when creating skills, writing SKILL.md files, setting up skill directory structures, or validating skill frontmatter. Delegates to skill-author teammate using TDD methodology.

Run tool-driven code quality analysis then apply LLM judgment for cleanup. Use when cleaning up a codebase, removing slop, dead code, or improving code quality. Two-phase approach: static analysis tools find problems, agent judges what to fix.

Generate a right-sized feature specification with acceptance criteria, self-contained phases, and transcript-verifiable goal conditions. Use when you know what to build.

Delegation and orchestration patterns for Claude Code — when to work directly vs delegate, subagents vs Workflow fan-outs, parallel write isolation, verification topology, and effort/model tiering. Use when coordinating subagents, running multi-agent work, or deciding whether to delegate at all.

Testing patterns for running and writing tests across project types. Use when running lint, typecheck, or tests, writing new tests, setting up test infrastructure, or diagnosing test failures. Covers auto-detection of test frameworks, error categorization, and structured reporting.

Create a GitHub issue through a guided interview. Explores the codebase first, asks batched questions, previews, then runs gh issue create.

Generate detailed, actionable developer tasks using TaskCreate with rich descriptions, dependency tracking, and owner assignment.