Agentic-Driven Development (add)

A de-hooked clone of the Agentic Development Skills workflow, tuned to run fully unattended. You boot it explicitly, it researches, decides for itself, builds, and hands you a pull request — no human in the loop. Built to run as one of several background leaves in parallel.

spec → plan → implement → verify → PR, with the agent making and recording every decision instead of asking.

There is no SessionStart hook — nothing is auto-injected. Start a run on demand:

/agentic-driven-development add react to this project

The agentic-driven-development skill is a short router (progressive disclosure): it holds the workflow map + the Unattended Operation Contract, and loads each sub-skill on demand.

The workflow

agentic-driven-development   (router — invoked explicitly via /agentic-driven-development)
  │  Unattended Operation Contract: never pause; decide + record; end with a PR
  │
  ├─ add-brainstorming ─────────►  spec doc   docs/add/specs/…   (no approval gate; records Assumptions & Decisions)
  │     └─ add-writing-plans ───►  plan doc   docs/add/plans/…   (auto-continues to execution)
  │           └─ execute the plan, task-by-task:
  │                • add-subagent-driven-development   (REQUIRED — fresh subagent + 2-stage review, continuous)
  │                     ├─ add-test-driven-development        (red → green → refactor, per task)
  │                     ├─ add-requesting-code-review         (reviewers are subagents)
  │                     ├─ add-verification-before-completion (evidence before any "done")
  │                     └─ add-finishing-a-development-branch (auto: push + open PR)
  │                • add-executing-plans               (fallback if subagents unavailable)
  │
  ├─ add-using-git-worktrees      (DEFAULT ON — isolate each leaf so parallel runs don't collide)
  └─ add-systematic-debugging     (any bug / test failure — root cause before fixes)

What "unattended" changes vs. the original

• No approval gates. brainstorming writes the spec and moves on; no design/spec sign-off.
• No execution-path question. writing-plans hands straight to the continuous subagent path.
• No merge/keep/discard menu. finishing always pushes and opens a PR.
• Decisions are recorded, not asked. Every non-obvious call goes in an Assumptions & Decisions log that travels spec → plan → PR body, so you can review the agent's judgment after the fact.
• Blockers don't stall the run. Only unrecoverable external blockers (missing creds/services) stop a task — they become deferred TODOs surfaced in the PR; the rest of the plan continues.
• Isolation on by default. Each leaf gets its own branch/worktree.

For the involved, human-in-the-loop counterpart, use human-driven-development (/human-driven-development).

Install (npx skills)

npx skills add ./AgenticDrivenDevelopment

Skills install flat, so every skill is prefixed add- to avoid colliding with the hdd- skills from Human-Driven Development when both are installed.

Conventions

• Skills live in skills/<name>/SKILL.md (YAML frontmatter + markdown body).
• Skills reference each other by their prefixed name, add-<skill>.
• Workflow artifacts go in docs/add/specs/ and docs/add/plans/ (user preferences override).

Ported from obra/superpowers via Agentic Development Skills.