Ryan Carson’s Three-File Workflow for Building with Coding Agents

Founder Ryan Carson shows Peter Yang a three-file prompt workflow for building a feature with a coding agent. The files guide a product requirements document, a task list and the processing of those tasks. He demonstrates it on his own early-stage divorce-support application; the application is not a substitute for legal advice.

The first Markdown prompt asks the agent to clarify the problem, audience and goals before producing a PRD. In the demonstration Carson describes a relationship-assessment idea and, to save interview time, lets the agent choose answers to some questions. He says he would normally review and answer them more carefully.

The second prompt turns the PRD into a task list. Carson asks for high-level parent tasks first, reviews them, adds a feature-branch step and then generates smaller subtasks. A more capable reasoning tool reviews the plan, but Carson still removes suggestions he does not want.

He asks for meaningful tests during implementation so the agent has feedback before proceeding. The third file controls execution: work through one subtask at a time, mark it complete, run the relevant tests after a parent task and commit when appropriate. These instructions support a review loop; passing tests alone does not establish production readiness.

The example reaches a database change, which Carson recognizes requires a migration. A planned task list is not a guarantee that a migration is safe: changes to sensitive data still require careful review, suitable access controls and recovery planning.

Carson recommends using an agent as a tutor: pick a subject you care about, build a small project and ask the assistant to explain its choices. The learner remains responsible for checking explanations and understanding how the application behaves.

During the interview he interrupts the implementation to request a static questionnaire mockup. Seeing the interface reveals decisions that might otherwise be discovered after database work, and he considers adding that visual step earlier in the workflow. The demo shows a basic UI rather than a fully validated finished feature.

Carson describes choosing a narrow, painful problem after seeing relatives struggle with divorce, and contrasts urgent problems with products aimed at gradual self-improvement. His application is initially limited to Connecticut because legal processes differ by jurisdiction. His business and cost claims are his own estimates, not evidence of a guaranteed outcome.

As a solo founder, Carson values control over his work and time, while acknowledging that a small business and a venture-backed company have different ambitions and responsibilities. He also works on the Amp coding product demonstrated in the episode, a connection relevant to his tool preference.