ankitclassicvision/agent-automation-creator

AAC, short for Agent Automation Creator, is a written framework for designing workflows that include AI. The author's argument is that most AI projects fail because teams pick the AI tool first and then bend a process around it. AAC flips the order. You map the process, then assign the right kind of runtime to each piece of work, and only then add AI discipline where AI is actually doing the work. The framework has three layers and the author says you must build them bottom up. Layer 1 is the process map, drawn at three levels of detail: small work elements with one verb each, nodes that group work with a single owner, and the full process graph with triggers and queues. Layer 2 is runtime assignment, where every work element is tagged as Deterministic Code, Closed-Loop AI, Assisted AI, or Human Only. The cheapest option that meets the constraints wins. Layer 3 applies only to the Closed-Loop AI parts and lists five rules: Bounded, Grounded, Gated, Observed, and Governed. If any rule fails, that piece runs as Assisted AI instead, with a human in the loop. The repository itself is mostly documents. The main artifact is a 34 page PDF spec with a 57 item checklist that builders can walk through to grade a workflow. There are two infographic PNGs that explain the ideas visually, and a sample HTML report you can open in a browser to see what the output looks like. The repo also ships a companion tool called AgentTwin. It is a skill you drop into an AI assistant such as Claude Code. You point it at an agent, automation, or vendor pitch, and it walks the 57 item rubric and produces a two view HTML report: a plain English summary with a letter grade for executives, and a detailed process map view for builders. Install steps and a test prompt are included in the README.