agents-stack

Use when a request will produce frontend UI, a visual asset, an image prompt, or a video prompt, AND the intent/style is not already fully specified by the user or an existing design system in the repo. Forces a minimal, checkable design-context object to exist before generation so output does not default to generic/templated. Do NOT use for: single-property edits (color, spacing, copy tweak) with no ambiguity; bug fixes; refactors; or when the user already gave complete visual specs and you only need to execute. For those, just do the work.

Root orchestrator. Reads durable state, routes to one phase, dispatches fresh workers.

Pre-implementation alignment gate. Checks spec / plan / tasks consistency and plan vs codebase reality before code is written.

Define the goal: what problem we solve, for whom, and what success looks like. The directional anchor that architecture and specs must serve.

RED-GREEN-REFACTOR implementation per task. Each task passes before next. Produces handoff.md with reproducible evidence.

Design architecture: components, API schema, DB model, impact analysis, test strategy.

Independently verify implementation against SPEC acceptance criteria. Generator ≠ Auditor.

Write changelog, update reference docs, archive completed workstream.

Define what we build: goal, user stories, edge cases, BDD acceptance criteria.

Break plan into minimal verifiable tasks with 5-dimension verification metadata.

Checkpoint

Extract implicit knowledge and audit AGENTS.md against four principles. Derives sections from domain, not from a fixed template.

On-demand mistake/learning logger. Records errors, corrections, and insights to `.agents-stack/learnings/` for cross-session recall. Also queries past learnings on request.

Use when designing APIs, making architectural decisions, adding features, writing code, code review, UI design, process design, or any design decision. Provides a cross-level (code → system → process → practitioner → UI → API) simplification discipline: identify what is essential, protect it from the transient, remove everything else.

Use when design.md is in design_contracted phase and the design artifact must be implemented.

Use when a design deliverable must be produced as a durable HTML artifact — prototypes, slide decks, animations, wireframes, or UI mockups — and the work must follow a file-first, adversarially reviewed, harness-compatible workflow.

Integrated 5-phase reasoning workflow: calibrate state, frame the problem, analyze with lenses, recommend with evidence, verify before acting.

Use when a request is analytical and could plausibly benefit from integrated reasoning, reality-check, or scenario-planning templates.

Use when starting a new product from a raw idea, pressure-testing startup viability, or moving through blueprint, requirements, and system-design stages that produce durable reference docs.

Use when a raw product idea needs a structured blueprint, testable requirements, and implementation design. Covers the full greenfield definition pipeline: vision → PRD → architecture.

Use when creating or upgrading reusable agent manifests or agent teams with explicit model profile metadata and harness-style contracts.

Use when creating or upgrading a reusable skill package — leaf or router. Covers scaffolding, SKILL.md authoring, resource bundling, and validation. Does not cover agent manifest creation (use create-agents for that).

Use when creating, upgrading, or packaging reusable skills and you need to choose between a leaf skill and a router family.

Use when you need to break a system, find edge cases, challenge design assumptions, or validate product stability under adversarial conditions. Triggers: "red team", "adversarial testing", "edge case hunt", "break this system", "stability test", "attack this", before high-risk releases, or when pipeline QA passes but you sense blind spots remain.

Use when extracting the implicit methodology from a deep-dive conversation — reverse-engineering the analytical path and operational principles into a reusable Framework + Workflow + SOP. Triggers include "extract the methodology", "derive a framework from this conversation", "what method did we use here", "turn this into a repeatable process", "提煉出方法論". Do NOT use for summarizing content; this extracts the method, not the subject.

Standalone complexity audit specialist — on-demand only. Not a lifecycle phase. Not gated on review.md. Dispatched when the user explicitly requests a complexity audit.

Use when validating backend behavior in real execution paths, including APIs, jobs, queues, webhooks, auth boundaries, data integrity, failure handling, observability, performance, or adversarial backend conditions.

Use when orchestrator has reconciled a design sprint outcome and queued the feature id in compound_pending_feature_ids.

Use when a design sprint is starting and no trusted design spec exists yet for this feature.

Use when design.md exists with phase design_spec and the design sprint needs a bounded scope, output contract, and human approval gate before building begins.

Use when design-builder has produced a handoff and the artifact must be evaluated adversarially before orchestrator processes the verdict.

Use when validating frontend behavior in a real browser, including functional flows, visual quality, accessibility, responsive behavior, perceived performance, or adversarial edge cases.

Transform raw domain materials into structured domain knowledge with mental models, debate maps, SOPs, stress tests, and execution checklists.

Use when extracting a production-ready brand identity and design system from curated reference images, mood boards, or brand descriptions. Triggers include requests to "extract brand identity", "create brand design system", "define brand DNA", "generate design tokens from images", "build a brand from moodboard", "create AI image prompts from brand", "design AI prompt methodology for brand", or "systematize brand identity for front-end handoff". Also triggers when the user provides reference images and asks for a structured brand output or AI image generation prompts.

Use when an AI agent must hand off in-progress work to a successor agent and needs a structured, high-signal payload that prevents hallucination, re-litigation of settled decisions, and duplicate work.

Use when a sparse or under-specified prompt for text, image, or video generation needs enrichment, clearer direction, or useful variants while preserving the user's core subject.

Use when pressure-testing a startup, launch plan, GTM thesis, or early unit economics with pessimistic but credible assumptions about acquisition, retention, monetization, burn, and runway.

Use when a request could plausibly be an informational site, a web application, or a browser game and the agent must choose the narrowest website-building child.

Use when public posts, threads, newsletters, interviews, transcripts, or creator/company notes should be turned into evidence-backed business opportunities by extracting recurring pains, DIY workarounds, and unmet jobs-to-be-done.

Use when building or evaluating model-generated interactive UI in the browser, including schema-driven component rendering, streamed UI, sandboxed HTML experiences, or agent-controlled interface updates.