A toolkit built around Claude Code commands /loop and /goal has been presented as a way to run 25 workflows across marketing, product, operations, and research on a schedule. The source article says the setup was used to keep Fable 5 focused on the same task for hours, with each workflow packaged with a prompt, attached tools, a definition of done, and explicit stop conditions.
In the framework, a loop is a scheduled task that repeats until it reaches a stop point, while a goal is a one-time finish line the model keeps working toward on its own. The structure follows a simple formula: define the objective, define the metric, define the boundaries, then let the agent work. What matters is not endless autonomy, but clear rules on when the model wakes up, what it is allowed to change in each round, how it records its history, and when it must stop.
How the loop model is structured
The article breaks every loop into five parts: a schedule, one change per round, the same recurring check, a state file, and a hard stopping point. The state file is treated as the key piece. Before each new round, the model reads its own written history so it does not redo completed work, then logs what it changed and what should happen next.
Risk controls are built in through three color bands. Green loops can run alone because they only read information and write to their own files. Yellow loops are draft-only and require human approval before anything is sent. Red loops are not allowed to run unattended, covering money, production systems, outbound communications, and anything visible to customers. The article also advises users to run every loop manually before scheduling it and to route routine rounds through cheaper models, only escalating to Fable when the lower-cost option fails.
Goals require proof inside the conversation
A goal works differently from a loop. It is not about timing. It is about reaching a finish line with evidence. The article says a smaller model can act as a judge, but that judge only sees the chat. It cannot open files, run tests, or inspect a website directly.
That limitation shapes how completion must be written. “Done when tests pass” is not enough, because the judge cannot verify it. “Done when the full green test output is pasted into the chat” is a valid contract. The article stresses this point repeatedly: an agent claiming success is not proof. Proof must be visible in the conversation, and if it cannot be produced, the model should post a failure message and stop.
Four categories of workflows
The 25 workflows are grouped into four business areas. Marketing and content include answer-engine gap detection, quality checks for programmatic pages, brand visibility tracking in AI answers, competitor content monitoring, a content-outline backlog, X idea mining, ad creative fatigue drafting, expectation-gap audits, and pre-sales question capture. Product workflows include feature radar from brand mentions, review mining for roadmap input, and churn-point copy fixes.
Operations workflows cover inbox triage, month-end reconciliation prep, SOP drift detection, proposal and RFP drafting, KPI anomaly monitoring, and overdue invoice follow-up. Research and decision workflows include overnight intelligence refreshes, regulatory and source summaries, escalation queues for hard problems, elimination-condition loops, premortems, repeat-failure summaries, and shadow prompt testing. The article also explains tool access through MCP connectors and API keys, with a preference for hosted MCP systems that support OAuth.
Suggested starting point
The recommended rollout is simple: start with one loop tied to the biggest operational gap, such as inbox handling, KPI monitoring, content work, or error review. Run it once manually and inspect the state file. Then schedule it. A week later, add a second loop or hand a painful one-off task to /goal, but keep the completion rule tied to posted evidence.
The article spends less time selling automation and more time warning about cost and control. Its point is blunt. A model that never gets tired will not stop by itself, and Fable is described as one of the most expensive models available. Without budget limits, stop rules, and human review on sensitive actions, the bill and the operational risk rise together.