Task: Capture Conversation Workflow into Structured Document

Task: Capture Conversation Workflow into Structured Document

Objective

Given a conversation where the user guided an agent step-by-step through a workflow — invoking skills, running ad-hoc commands, making decisions, discovering edge cases — produce a WORKFLOW.md that captures the complete workflow in a structured, generalizable format. The result is a document that anyone (human or orchestrator-creator skill) can read to understand every step, every decision point, every feedback loop, and every input/output relationship in the workflow, without needing to replay the conversation. When an existing WORKFLOW.md is present, merge new findings with the existing document rather than replacing it.

Inputs

1. Conversation history (auto-discovered) — the full conversation in the current context window. This is the primary input. The agent reviews everything: user messages, agent actions, tool calls, skill invocations, outputs, failures, corrections, and explanations.
2. Existing WORKFLOW.md (auto-discovered) — if a WORKFLOW.md already exists in the working directory, read it before starting. This triggers refinement mode: merge new conversation findings with the existing document instead of writing from scratch.
3. User's final guidance (optional) — any instructions the user provides alongside the skill invocation, such as "focus on the deployment steps" or "this was a retry of the analysis phase."

Workflow

Workflow extraction follows four phases: inventory, extract, generalize, and write.

Phase 1: Inventory

Scan the full conversation from beginning to end. Build a chronological list of every meaningful action. An action is meaningful if it changed state, produced output, consumed input, or represented a decision.

For each action, note:

• What happened — the command, skill invocation, file edit, or explanation
• Who initiated it — user instruction or agent initiative
• What it consumed — files read, data referenced, prior outputs used
• What it produced — files written, state changes, information discovered
• Whether it succeeded or failed — and what happened next

Also inventory:

• Decision points — moments where the user chose between alternatives. Record the options considered, the choice made, and the reasoning given.
• Corrections — moments where the user said "no, not that" or "go back and do X instead." These reveal important constraints.
• Edge cases — unexpected situations discovered during execution. How were they handled? Were they resolved or left open?
• Project context — any information the user shared about the tech stack, architecture, conventions, team practices, or constraints.

If an existing WORKFLOW.md is present, read it now. Note which steps it already covers and which are new, changed, or contradicted by the current conversation.

Phase 2: Extract