design-gatelisted

Decides the architecture for an AI feature before it is built, whether the work should be a deterministic workflow, a workflow with a model at the judgment points, a single model call, a single agent with tools, or multi-agent. Use when someone is scoping or designing a new AI use case, asking whether it needs an agent, reviewing a PRD, ticket, or spec for an agent, or pushing back on a design that reaches for an agent where a workflow would do. Also covers Salesforce Agentforce designs, deciding whether a use case should be a Flow, a prompt template, an Agentforce topic with actions, or multiple agents. Reads the design input, extracts the deciding signals with evidence, runs a deterministic verdict, and writes an architecture decision record that biases toward the simplest thing that works.
hamidettefagh/two-gates · ★ 0 · AI & Automation · score 73

Install: claude install-skill hamidettefagh/two-gates

# Agent, or workflow? Most agents that struggle in production were the wrong shape from the start. The work was predictable and got handed to a model anyway, or a single call was grown into a multi-agent system nobody can test. This skill runs the decision that happens before any of that: given a use case, what should you actually build, and where does reasoning genuinely earn its seat. It is the design-time gate. Its sibling, the [agent-production-readiness](https://github.com/hamidettefagh/agent-production-readiness) review, is the ship-time gate. This one decides what to build; that one decides whether it is ready. ## When to run this Run this when the user is: - Scoping or designing a new AI feature, and deciding what shape it should take. - Asking whether something needs an agent, or whether a workflow would do. - Reviewing a PRD, ticket, design doc, or spec that proposes an agent. - Pushing back on a design that feels over-engineered, or under-specified. Do not run it to review an agent that is already built and heading for production. That is the readiness review. Ask if you are unsure which gate the user needs. ## The bias Reach for the simplest thing that works, and make anything more earn its place. In order of increasing cost and risk: a single model call, a deterministic workflow, a workflow with a model at the judgment points, a single agent with tools, multi-agent. Default down this list, not up it. A model in a path that did not need one buys cost, late