Prompt File Method

# The Prompt-File Method

There is one workflow that outperforms inline conversation prompts, verbal briefings, and direct agent coordination. It's also the simplest: write the task as a standalone markdown file, point the agent at the file, receive a structured report, verify, move forward.

The prompt file is the unit of work.

---

## Why Markdown Files, Not Conversation

When you prompt an agent inside a conversation, the task inherits all the ambiguity of the conversation — unstated assumptions, context from 40 messages ago, things that were decided but never written down. The agent guesses at gaps. The gaps compound.

When you write a prompt file, you're forced to make the task self-contained. Every file path is explicit. Every constraint is stated. Every DO-NOT is written. The agent executing the file doesn't need to read your conversation history. It reads one file and executes.

Additional properties that matter:

**Git-able.** The prompt file lives in the repository. It can be diffed, reverted, reviewed. You can see exactly what instruction was given when a bug traces back to a bad prompt.

**Replayable.** If an agent session crashes mid-task, you don't reconstruct the prompt from memory. You point a fresh session at the same file.

**Diff-able.** When you update a prompt for a follow-up task, the diff shows exactly what changed — which means you can review whether the change is complete and correct before dispatching.

**Cold-start readable.** An agent with zero prior context can execute a well-written prompt file. This is the key property. Every session is effectively cold-start. Design for cold-start.

---

## The Workflow