Case study
Side project: I developed constraint-first prompting in my own work and conversations first, then froze the 4-D flow and a 6-part anatomy into a skill; AI drafts inside rails I set.
Outcomes
Context
I kept writing prompts that felt sharp in the moment and fell apart the next week. The failure was not the model — it was me. I had no repeatable way to fix role, constraints, and output shape before I asked for words. I developed Prompt Architect to resolve this once and for all by applying constraint-first prompting in my own work and my own conversations first. I drew on a wide range of AI-related material online — structured guidelines, documentation, and community threads about getting real work out of tools. I kept what survived contact with my own use and dropped the rest. I distilled what remained into one cohesive, practical approach that makes sense when you run it, not when you admire it. Only then did I freeze it into a skill — the 6-part anatomy and 4-D flow as the fence, execution underneath.
The question I answered before writing a line of SKILL.md:
What must stay fixed so AI can run fast without rewriting the rules every session?
How it was built
Constraint
Before generation
I locked the structure first
I locked the 6 anatomy headings and the 4-D order into SKILL.md — after distilling many threads into one approach that had already survived my own use. Only after that did I let the model draft explanations and examples under each heading. The rule was clear: structure before prose.
Judgment
Human
What I refused to outsource
I decided what counts as a stop condition, what belongs in Output versus Task, and which failure modes matter for this skill. Those calls shape how the model behaves when a user is vague — I did not delegate them to a first draft.
Execution
AI
What the model did well
Once the rails were set, AI filled sections, tightened wording, and surfaced edge cases in tables I had already defined. Execution — focused on delivery, not direction.
Edit
Cut
Trimming drift
I cut anything that sounded like generic prompt advice or bloated the skill into a blog post. The skill mirrors a design language: opinionated, bounded, reusable.
Implementation
Where AI entered
- —After headings and section order were fixed, AI drafted body copy, examples, and platform notes under each locked heading.
- —Iteration stayed inside the same file shape — no surprise sections, no alternate frameworks sneaking in.
- —When text drifted toward motivational fluff, I deleted it. The skill is operational, not inspirational.
Where judgment stayed mine
- —Choosing the Agent Skill format and file layout — what ships as the contract.
- —Defining BASIC versus DETAIL modes and what each may skip — that is product behaviour, not copy.
- —Naming failure modes in the anatomy table — what "weak" looks like per section.