Disruptor

Disruptor

Eleven battle-tested skills that carry an AI-agent-built project from idea to running server, wired into one ordered flow. This skill is the router: it names the stage sequence, the artifact each stage hands the next, and the gates that decide where you go — don't reinvent the skills here.

Core principle: the agent's own sense of "done" is the enemy. Every method exists to outlast a plausible-but-premature stop — an unread green check, a spec with a hidden default, a rubber-stamp review, a loop that only ever adds. Trust the method over the feeling, and the gate over the guess.

The pipeline

0 Setup ─► 1 Design ─► 2 Spec ─► 3 Guardrails ─► 4 Breakdown ─► 5 Build&converge ─► 6 QA ─► 7 Deploy
 (once)     7w3         spec       arch           slices          ├ 5a reviewers
                                                     │            └ 5b unvibe
                                                    G3F ──► 4–5 fleet lane ───────► 6 QA

Stage 5 is the engine: converging-and-polishing orchestrates spawning-reviewers (each review-and-fix iteration) and unvibe-review (each cut pass) — load all three for the loop. Every stage skill opens with a ## Stage contract block (Inputs · Outputs · Entry gate · Done when · Next): when Done when is met, read its Next — not your intuition — to move on.

Start here — the router

Don't pick a stage by feel: find your on-ramp, then walk the gates — each a checkable condition with an explicit destination, so the route is a rule, not a guess.

On-ramps (how work enters):

• New idea / RFC / ticket (greenfield) → gate G0, then stage 1.
• Bug / incident (something's broken) → stage 6 (qa-demo-stand reproduces + root-causes) → fix through stage 5 → re-QA → stage 7 if it ships.
• "It feels over-engineered" / architecture drifting (existing codebase) → stage 5b unvibe-review for the cut, and audit stage 3 guardrails; the fixes re-enter stage 5 as slices.
• Already have a design or a spec → skip ahead: settled 7w3 design → stage 2; build-ready spec → gate G3.
• Vague product idea, speed-to-testable-MVP matters more than full design rigor → the delivering-mvp-fleet lane (own interview → plan → board → fleet), rejoining at stage 6 QA.
• A concept/PRD produced outside the flow (business validation, research hub) → stage 1 designing-with-7w3; if it is already a build-ready spec → gate G3.

Gates (which stage next):

• G0 — Is the repo set up? No CONTEXT.md / conventions recorded → run stage 0 first.
• G1 — What's the trigger? Resolved by the on-ramp above; a fresh build → stage 1.
• G2 — Is the design ready to spec? All ten facets of every leaf subject grounded, no unacknowledged holes? No → stay in stage 1. Yes → stage 2.
• G3 — Is this more than one slice? Yes → stage 4, then gate G3F. No → one stage-5 loop right here.
• G3F — Is the build parallelizable across independent nodes AND is there budget for a worker fleet? Yes → delivering-mvp-fleet (it builds its own board from the spec/slices). No → run converging-and-polishing once per slice, each in a fresh session.
• G4 — Has it converged? Two consecutive near-zero passes from cross-family critics + a whole-artifact pass, settled after the last change? No → keep looping 5. Yes → 6. (Declared from raw findings — never because a reviewer said so.)
• G5 — Does this touch shared/prod infra? Before destructive QA (stage 6) or irreversible provisioning (stage 7) → stop and confirm with the human; read the prompt's hard invariants first.

Two kinds of skill — load-and-follow vs fill-and-hand-off

Mixing them up is the main misuse — don't paraphrase a prompt template into your own words; don't hand a method to an agent as a task prompt.

Cross-cutting invariants

The canonical list — other skills cite it ("per the disruptor invariants"), never restate it.

• Deny-by-default, fail loud. No hidden default for required data or config — demand it, fail loud; rights denied unless the spec grants them. Why: a silent fallback hides a broken assumption until it ships.
• Independence beats agreement. A review is only worth the independence of the critic — a different model family, or a fresh no-context subagent. Why: you cannot catch the gaps you introduced by re-reading your own reasoning; agreement is an echo, not a signal.
• Two forces, kept apart. Iteration adds (the ratchet); a separate pass removes (the cut). Never fuse them. Why: an add-only loop grows into bloat; a deleting review muddies both signals.
• Architecture is executable, not prose. Module boundaries are enforced by linter + import-graph + filesystem checks in CI, not a doc. Why: an unchecked rule gets "temporarily" violated and never restored.
• The document is the source of truth. Design, spec, and tests are the record; code realizes them. If reality diverges, update the document. Why: stale intent is how a team loses the ability to reason.
• Verify effective state, not the write. Trust sshd -T, docker inspect, a re-run, a real test — not that a file was saved or a command exited 0. Why: masked exit codes, stale caches, and reload-vs-restart gaps manufacture false greens.
• Base rules travel with every subagent prompt. DRY, KISS, SOLID, YAGNI, and file-size caps are restated in each subagent's prompt, never assumed to inherit. Why: an orchestrator authoring child prompts silently drops its own standing rules — one real run produced a 5.5K-line monolith.

Provenance

From the @aostrikov_agents_chat Telegram community: Nambr — designing-with-7w3 and the review-loop suite (converging-and-polishing, spawning-reviewers, unvibe-review); Pukh — the prompt templates (write-spec, architecture-guardrails, qa-demo-stand, setup-server); Serge Shima — the router flow, setting-up-domain-model, slicing-into-tracer-bullets, and delivering-mvp-fleet.

Source: Serge Shima (@aostrikov_agents_chat), the Disruptor pipeline and router.