How you make a new hire good

Two ways to work with AI

Build a skill

The prompt — build the skill

Prompt“

Turn what we just did into a skill using the skill creator for this task.

Keep it lean and minimal — every line must earn its place. Capture the judgment and nuance we used, and keep the skill general so it applies to any proposal of this kind.

Two ways people test — and fool themselves

✕

You read one output and it looks fine

✕

You test it on the same example you built it with

Test it blind — against work it's never seen

The prompt — write the checklist

Prompt“

I want a checklist that tells me if a drafted proposal is good.

Here are three different proposals I wrote and won, and the skill that drafts new ones [SKILL NAME].

Write the checklist as plain yes-or-no questions, each with exactly one clear answer. Pull out the standard that runs through all three proposals, so the questions apply to any proposal of this kind and hold true for every one of them, not just one example. Aim for seven to ten questions that capture what really makes these win.

Turn your standard into yes / no

Does it open with the client's real problem?yes · no

Is the price on the first page?yes · no

Is it under two pages?yes · no

Let a subagent grade it

The prompt — add the grader

Prompt“

Add a grading step to this skill.

After the skill writes a proposal, have a separate subagent grade it with fresh eyes, looking only at the finished proposal and the yes-or-no checklist. It grades the proposal against each question and returns a short report: each question, a yes or no, and one line of proof for every no.

Keep this lean, and keep the writing step exactly as it is — only add the grading subagent and the report.

The prompt — one round of the loop

Prompt“

Here is the proposal the skill wrote, my real proposal for the same client, and the checklist questions it failed.

Show me where the skill's version falls short of mine, question by question. Then make the smallest change to the skill that fixes the repeated patterns behind those misses, and leave everything that already works exactly as it is. When you are done, show me exactly what you changed.

A pass is a photo — not a promise

Three things to remember

1Don't just use your AI — train it.

2Prove it's good — test it blind against your real work.

3Coach it in a loop until it matches you every time.