The Process Prompt Stack: Where the Cohort Has Landed

What the cohort is now doing

The summer arc of AIMM converged on something the room is calling the process prompt. The shape is stable enough now to write down.

A process prompt is a single, reusable system instruction that interviews the user, infers the variables for a specific job, plans the work, and then executes against a generalized framework. It is the layer above a super prompt, which is the layer above a structured PROMPTS prompt, which is the layer above conversational prompting.

The shift the cohort is making, mostly without naming it, is from writing prompts to writing the prompts that write the prompts.

What is sticking

Three patterns have shown up across enough member work to call them stable:

1. Voice and style files as durable assets. Don’s third revision of his style prompt is qualitatively different from the first. Lou’s voice file has been running unchanged across three model generations. The style file is now treated as infrastructure, not as a draft.
2. Project folders as context containers. Most members have moved from chat-by-chat work to project-by-project work, with a folder of context files per major workstream. The folder is the unit of leverage, not the chat.
3. Frameworks encoded as prompts, not as documents. Members who used to keep frameworks in Notion are now keeping them in markdown files that the AI can ingest, edit, and execute against. The framework file is a tool, not a reference.

What is not sticking

Two things have tried and failed to land for the cohort:

1. Prompt libraries with hundreds of entries. The members who tried to build comprehensive libraries stopped using them within weeks. The members who keep five to ten well-tuned prompts use them daily.
2. Aggressive workflow automation before the workflow is stable. Several members built n8n flows for processes they had not yet run manually three times. Every one of them has been rebuilt.

The move most members are skipping

Almost no one is interviewing themselves before writing the process prompt. They are writing the prompt the way they think they want the work done, instead of asking the AI to interview them about how they actually do the work and producing the prompt from that conversation.

The members who have made this move report a noticeably different output. The prompts feel more like them, because they were derived from how they actually work, not from how they would describe their work in a single sitting.

What to try this week

Pick one piece of work you have done at least five times. Open a fresh chat. Ask the model to interview you about how you actually do that work, with permission to push back when your answers are too clean. Once the interview is done, ask the model to write the process prompt based on the conversation. Use that prompt next time you do the work. Compare.