DeepProgrammer

LabExperimentPrototype

The starter workflow, on paper first

A personal experiment: can a four-move loop make AI use repeatable without turning it into a product?

4 Nov 2026

Status
Prototype
Kind
Experiment
Role
Personal practice. I am the only user.
Scope
Ordinary writing and review tasks. No team rollout, no integration, no claim that this changes anyone else’s output.

  1. 01

    Frame

    The job, in one sentence.

  2. 02

    Draft

    A pass you expect to mark up.

  3. 03

    Interrogate

    Assumptions, gaps, refusals.

  4. 04

    Keep

    Only what you would sign.

Problem

My use of AI was opportunistic. A prompt here, a rewrite there, nothing I could hand to my future self on a busy day.

Approach

I wrote the loop on a page before opening a tool: Frame, Draft, Interrogate, Keep. I ran it on a few real tasks and threw away anything I would not sign.

Stack

A written checklist · A prompt shape · Whatever model is at hand · No application

What changed

The first prompt got slower. The second pass got more honest. I stopped saving prompts that had only worked once.

What I learned

The constraint list and the refusal list do more than the choice of model. A workflow that cannot say no is just a shorter path to a fluent mistake.

Personal practice only. Nothing here is employer process, customer data, or a shipped product.

Related note: A practical AI workflow for a normal workday · Build