Huge - thanks Rob. I also found that trying to use claude to manage an agentic workflow on a project simply for risk/action tracking, drafting, etc - became a chore to stay on top of contex windows, migrating markdown files, and remembering to inform the AI of changes for diminishing benefit.
Yeah, that's the failure mode most people hit. You end up spending more energy managing the AI than doing the actual work.
The thing that changed it for me was stopping trying to fight the context window and designing around it instead. Treat the AI as stateless by default, then build the state management in files alongside it. Markdown files that persist decisions, config that reloads conventions at session start, a file-based task queue the AI reads rather than waits to be told about.
The markdown migration chore disappears when file state is the source of truth from the start, not an afterthought you're copying out of chat history.
The "remembering to inform the AI" problem is really a missing session-start protocol. If the AI reads its own context files before anything else, you're not briefing it -- it briefs itself.
None of it is magic. It's just treating the workflow like any other stateful system that happens to have a stateless component in the middle.
Huge - thanks Rob. I also found that trying to use claude to manage an agentic workflow on a project simply for risk/action tracking, drafting, etc - became a chore to stay on top of contex windows, migrating markdown files, and remembering to inform the AI of changes for diminishing benefit.
Yeah, that's the failure mode most people hit. You end up spending more energy managing the AI than doing the actual work.
The thing that changed it for me was stopping trying to fight the context window and designing around it instead. Treat the AI as stateless by default, then build the state management in files alongside it. Markdown files that persist decisions, config that reloads conventions at session start, a file-based task queue the AI reads rather than waits to be told about.
The markdown migration chore disappears when file state is the source of truth from the start, not an afterthought you're copying out of chat history.
The "remembering to inform the AI" problem is really a missing session-start protocol. If the AI reads its own context files before anything else, you're not briefing it -- it briefs itself.
None of it is magic. It's just treating the workflow like any other stateful system that happens to have a stateless component in the middle.