A Systematic Workflow for Writing and Testing Prompts
Chapter 3's self-check assumed a prompt had already been through a real testing process. This is that process — a repeatable sequence that turns a first draft into something you can actually stand behind, applied consistently to every prompt you build rather than done once and forgotten.
The Workflow
Start from a specific goal, not a vibe
Write down exactly what output you want, for whom, before typing a single word of the prompt itself. "A good image" isn't a goal; "a consistent, warm-lit product photo style for handmade jewellery listings" is.
Draft v1, then run it many times
For image prompts, run it across several different seeds. For text prompts, run it against several different realistic inputs (a different product, a different name, a different scenario each time). One good output proves nothing — consistency across many runs is the actual test.
Log where it breaks
Note every run that went off-topic, ignored an instruction, or produced something unusable. These failure modes — not the successes — are what tell you what to fix next.
Add one guardrail at a time
For each failure mode found in Step 3, add a specific constraint that closes it — per Chapter 3's own guardrails pillar. Change one thing at a time rather than rewriting the whole prompt at once, or you won't know which change actually fixed the problem.
Re-test after every change
Repeat Step 2 after each guardrail added. A fix for one failure mode can quietly introduce a new one — the only way to catch that is testing again, not assuming the fix worked.
Document the final version
Record the exact final wording, the model/tool and version it was tested against, and any known limits (input types it doesn't handle well, settings that matter). This is what Chapter 3's documentation pillar actually produces.
File it into a personal prompt library
Keep every finished, tested prompt organised by niche and category in one place (a spreadsheet or a document works fine to start). This is what Chapter 6 draws on for packaging, and what eventually lets you bundle related prompts into themed packs rather than starting from zero each time.
Testing Differs by Category
Image Prompts
- Vary the seed, keep the prompt fixed
- Check consistency of style, composition, and quality
- Model checkpoints are usually stable once released — less ongoing drift risk
Text / Chat Prompts
- Vary the input (different names, topics, scenarios), keep the prompt fixed
- Check consistency of length, tone, and factual accuracy
- Underlying chat models update more frequently — higher ongoing drift risk