Do not index
Do not index
I thought I had a prompt problem. Some drafts sounded like me, and some did not. I kept rewriting instructions and
style_hints. Quality improved for a run or two, then drifted again.The issue was not effort. The issue was where judgment lived. My workflow was repeatable, but style capture was not.
Manual style hints were useful, but they were memory-based. Memory-based rules look clear and still leave room for interpretation. When output was off, I could edit the draft, but I could not reliably debug the system.
If quality depends on how well you prompt in the moment, quality will vary with your energy.
Shift to sample-derived style assets
I stopped tuning prompts and changed the architecture. I added a style-asset generation endpoint and an MCP tool that takes 2+ real writing samples and derives reusable style assets.
I feed it writing that already reflects my voice: Scope docs, task templates, and state-of-work updates. The system runs extraction and normalization, then outputs structured
template_schema plus style_hints with:- voice DNA
- negative constraints
- writing rules
- generation contract
- template sections
Those assets persist at the project level, and Scope loads them automatically on future runs.
This was the proof point:
Before:
- style_hints: manually written
- quality source: prompt skill at generation time
- consistency risk: high drift between runs
After:
- style_hints + template_schema: derived from 2+ real samples
- quality source: normalized asset loaded by default
- consistency risk: lower drift, faster first good draftBefore, each run asked, "Can I prompt this well right now?"
After, each run starts with, "Load the right asset."
How to apply this pattern
This pattern works beyond writing. If AI output is inconsistent in a repeated workflow, move critical behavior out of ad hoc prompts and into reusable assets.
Use this sequence:
- Pick one repeated task where quality varies.
- Collect 2-5 examples of your best real work.
- Extract patterns from examples, not memory.
- Encode those patterns as constraints and structure.
- Save the asset and load it by default.
- Review misses at the asset layer before rewriting prompts.
The principle is simple: move judgment from runtime improvisation to system design.
That is how you increase trust without adding more effort.
If you want to follow along, I am building PailFlow in the open and sharing how I am using AI to scale a one-woman business.
Written by
.png?table=block&id=9ba33ac6-8e12-48f6-b980-4333b612ec56&cache=v2)
Lola
Lola is the founder of Lunch Pail Labs. She enjoys discussing product, app marketplaces, and running a business. Feel free to connect with her on Twitter or LinkedIn.