From Random Prompts to a Real Workflow
Over the last 6 months, Hoang Nguyen has been evolving his AI coding workflow while building AI DevKit. The most important shift was not about getting better code generation — it was about moving from reusable commands and templates into a workflow that can carry context, trigger the right behavior, and verify work automatically.
A Concrete Example That Made the Difference
A recent feature made this evolution obvious. He used Codex to build interactive skill selection for the ai-devkit skill add command. With just one sentence of instruction, the workflow carried the task through:
- Requirements gathering
- Design
- Planning
- Implementation
- Verification
- Tests
- Code review
The entire session took under an hour, with the actual feature flow taking about 30 minutes. What made this remarkable was not just that AI wrote code — it's that the workflow left behind requirements docs, design artifacts, tests derived from requirements, and verification against the spec, rather than just a code diff.
Key Observations from the Workflow
Several things stood out in practice:
- Memory pulled back an old CLI rule that the author had forgotten he stored — the system remembered so he didn't have to.
- Review phases could loop backward instead of blindly moving forward, catching problems before they compounded.
- Verification caught drift between implementation and design — ensuring the final code actually matched the spec.
- Human judgment remained essential — he still made the product decisions and fixed the last failing test himself.
The Bigger Question: Prompts vs. Workflow Layer
The author poses a thought-provoking question to the community: Are you mostly optimizing prompts, or are you now trying to optimize the workflow layer around the model?
This distinction matters. Optimizing prompts is a local improvement — you get better output from a single interaction. Optimizing the workflow layer is a systemic improvement — you build a system that orchestrates AI across multiple steps, preserves context, enforces structure, and verifies results.
This connects directly to the "Engineer A vs. Engineer B" framing Hoang explored in his earlier post on agentic engineering: Engineer A uses AI faster; Engineer B uses AI at scale by coordinating multiple agents and workflows. The exponential effect doesn't come from typing speed — it comes from accumulated workflow leverage.
Why This Matters for Every Developer
As AI tools evolve rapidly — Cursor, Claude Code, Codex — the interface changes but the underlying insight remains: don't couple your discipline to a single tool. Build your workflow so it works across interfaces. A workflow that only functions inside one UI is fragile.
The shift from "AI as a prompt target" to "AI as a workflow participant" is one of the most important mental model changes an engineer can make right now.
Original article: How My AI Workflow Evolved from Prompts to Workflow — Hoang Nguyen, codeaholicguy.com, April 11, 2026
Also discussed on: Hacker News
]]>