disruptor-skills

🧨 Disruptor — agent skills for AI-driven software development, from idea to ship **A gated pipeline of 12 agent skills that turns an idea into deployed, production-grade software — design, spec, executable architecture, honest review loops, QA, safe deploy. Works with Claude Code, Codex, Cursor, opencode — any agent that speaks the skills format. No vibecode.** [](CHANGELOG.md) [](#the-pipeline) [](LICENSE) [](https://skills.sh) [](CONTRIBUTING.md) [](https://t.me/aostrikov_agents_chat) [](https://github.com/smixs/disruptor-skills/stargazers) --- What is it **Disruptor** is a set of 12 agent skills for AI-driven software development, stitched into **one idea-to-ship pipeline**. Skills are plain Markdown with YAML frontmatter — an open cross-agent format read natively by Claude Code, Codex, Cursor, opencode and others. Each stage consumes the artifact of the previous one and hands a new artifact to the next, while a **router with Yes/No gates** decides what happens next. The agent acts **by rule**, not by vibe: it doesn't improvise the process and doesn't declare "done" early. The core premise: an agent's own feeling of "done" cannot be trusted — it is the main enemy. Every method here exists to survive a premature stop: the unread green test, the spec with a hidden default, the rubber-stamp review, the loop that only adds and never cuts. Each skill was extracted from real production runs by practitioners, not invented for a README. The pipeline Two kinds of skills. **Methods** trigger automatically when the situation matches (model-invoked). **Prompt templates** are `/commands` you invoke yourself (user-invoked). Stage | Skill | Kind | What it does | ---|---|---|---| — | `disruptor` | `/command` | The router: on-ramps + gates G0–G5, routes any input to the right stage | 0 | `setting-up-domain-model` | auto | Repo conventions, shared glossary, ADRs → `CONTEXT.md` (once per project) | 1 | `designing-with-7w3` | auto | 7w3 design interview: 10 facets per subject, a design tree per subject | 2 | `write-spec` | `/command` | Canonical build-ready specification, incl. golden case cards | 3 | `architecture-guardrails` | `/command` | Module boundaries as **executable** rules: linter + import graph + CI | 4 | `slicing-into-tracer-bullets` | auto | Breakdown into vertical slices (tracer bullets), each independently shippable | 5 | `converging-and-polishing` | auto | The engine: review loop until findings converge to near zero | 5a | `spawning-reviewers` | auto | Independent cross-family reviewers — fresh context, no rubber-stamping | 5b | `unvibe-review` | auto | The counter-force: cut, unify, rethink bloat instead of adding more | 4–5 | `delivering-mvp-fleet` | auto | Fleet lane: parallel subagent workers build an MVP from a board of nodes | 6 | `qa-demo-stand` | `/command` | Demo-stand testing to production quality; also the bug/incident on-ramp | 7 | `setup-server` | `/command` | Safe Ubuntu/Debian web server setup and deploy | **On-ramps** besides an idea from scratch: **bug/incident** → straight to QA, fix through the engine · **"bloated / architecture drifting"** → unvibe + guardrails audit · **existing design or spec** → jump to the matching stage · **vague idea, speed to testable MVP matters most** → the fleet lane. Install **With the skills.sh installer** (interactive, works across agents): **With install.sh** (copies all 12 skills, no `.git`): **With git clone:** Done — skills are discovered by the presence of the folders; no extra registration command. The `npx skills` installer places them into the right directory for each agent you use (`~/.claude/skills/`, `~/.agents/skills/`, …); `install.sh` and the clone target the `~/.claude/skills/` convention. As a Claude Code plugin (optional) The repo ships `.claude-plugin/plugin.json`, so it can also be installed through the /plugin flow. For most users the installs above are enough. Usage You don't drive the pipeline by hand — the **router** does. Two ways in: 1. **Just describe the task.** The model-invoked methods trigger on matching situations, and the router places you at the right stage via its gates. Example phrasings: "design this properly" · "slice this into tasks" · "review loop until it converges" · "what can we cut" · "this codebase feels bloated" · "build me an MVP of X fast" 2. **Invoke a `/command` directly** when you know what you need: `/disruptor` to run the full gated flow from any input, `/write-spec`, `/architecture-guardrails`, `/qa-demo-stand`, `/setup-server`. The router asks Yes/No gate questions (G0–G5: is setup done? is the design build-ready? is the build parallelizable?…) and refuses to skip ahead — that refusal is the point. Composing with domain skill packs Disruptor is a **process** pipeline — it says nothing about what good CSS, good SQL, or good motion design looks like. That's deliberate: domain taste lives in **domain skill packs**, and the two compose in your agent at runtime, not in this repo. With a pack like emilkowalski/skills (UI polish and animation craft) installed alongside, builders in stage 5 pick up the domain taste automatically, and stage 5a can spawn the pack's reviewer (e.g. `review-animations`) as a **standards-conformance review** — running beside, never instead of, the neutral cross-family defect review (see `spawning-reviewers`). The same shape works for any domain pack that ships expertise or a standards catalog. FAQ What is vibecode? Code that "technically works" but is bloated, fragile, and impossible to hold in your head — the default output of an agent left to trust its own sense of "done". Disruptor exists to prevent it: every stage gates on an artifact, not a feeling. How is this different from spec-kit, BMAD, or superpowers? Spec-kit and BMAD own your whole process end-to-end with heavy scaffolding. Disruptor is 12 small composable skills tied by a lightweight router — you can enter at any stage, use one skill standalone, or run the full flow. Compared to obra/superpowers (a broad general-purpose library), Disruptor is one opinionated pipeline for shipping web-facing software, including a parallel worker-fleet lane. Can I use it with Codex, Cursor, or other agents?