The fix isn't a longer prompt. It's editing the skill's instructions. We add an interview step: before drafting, the skill asks which client need the proposal should center on and which engagement we're proposing, leading each question with its own recommended answer plus a few alternates, so you can confirm in a word or steer hard when you want to. Then we run it again on the same prospect and the same transcript, and put run one and run two side by side. The skill got smarter because it asked.
What's covered:
- Iteration means editing the skill, not prompting harder
- Designing interview questions that recommend, then confirm
- Reading the SKILL.md diff
- Watching the interview fire on the second run
- Comparing run one and run two from the same inputs
- The lens: your version is whatever document you keep rewriting from calls