From knowing.
To building.

Write for two minutes before you prompt anything. Then, if you want, hand your writing to your chat assistant with this:

Here is a description of my HR work and the things I wait on. Read it and tell me: (1) the three dependencies that most slow me down, in my own words; (2) which one is a bottleneck of information, which of skill, and which of permission; (3) one of them that a small, single-purpose tool could plausibly remove. Ask me one clarifying question first if you need to.

[paste what you wrote]

Use these four questions to evaluate an AI opportunity before you build.

Take two minutes to score or annotate the dependency you identified. Which opportunity looks attractive, and which looks less attractive after these questions?

Look for opportunities that are valuable, frequent, feasible, and appropriately low-risk. Possible does not always mean worth doing. Evaluate risk by the consequences of a wrong answer.

Use this ladder to decide how much authority the AI should have in a given situation. Higher rungs mean more action and less human review before the output matters.

Before you build, ask: which rung is appropriate for this task, this data, and this reviewer? Most HR work starts at Suggest or Create and only moves up after it earns trust.

Five levels: Asker, Operator, Composer, Builder, Steward. Circle where you are today, honestly. Then look one step to the right. That is the level you are reaching for by the end of the day.

Booklet: p. 13.

H Human Accountability

You own the outcomes. AI assists, but you are responsible.

U Understand the Source

Know where the data came from and where it goes.

M Maintain Context

Give the model your rules, plan, and philosophy; it does not know them.

A Audit the Output

Facts, math, dates, citations, tone. Assume one thing is wrong until checked.

N Never Delegate Judgment

Decisions about people are made by people.

Booklet: p. 14.

"Grounded does not mean correct. Grounded means traceable."

When a tool gives you an answer, ask: can I follow this back to the source it used? If it quotes the policy line, the leveling text, or the data row, you can check it. If it cannot show you, treat it as a guess, even a confident one.

1. F

   Frame the job.

   Give the AI a role and a clear definition of done. Not "help with this policy," but "You are an HR communications specialist. Rewrite this policy for frontline employees at a ninth-grade reading level."

2. R

   Real context.

   Add the constraints the model cannot know: your audience, your plan year, your comp philosophy, what is off-limits. This is the step people skip, and it is the one that matters most.

3. A

   Add your source.

   Paste the actual messy input: the real job description, the real policy, the real de-identified data. Grounding beats guessing every time.

4. M

   Model the output.

   Say exactly what you want back: the format, the length, the tone, and one example of "good." Showing one example beats describing ten rules.

5. E

   Evaluate and iterate.

   Treat the first answer as a draft, never a deliverable. Refine it in plain language, then audit it against the Human Stewardship Model before it goes anywhere near an employee.

Booklet: pp. 8 to 9. Full prompt library under Resources.

1. 01

   Name the dependency.

   What, or who, are you waiting on? Write it in one sentence. If you cannot name it, you are not ready to build.

2. 02

   Front-load the reasoning.

   In a regular AI chat, think the whole thing through before you touch a builder: inputs, output, edge cases, and where the data lives.

3. 03

   Let one AI write the build brief.

   Have the chat model turn your reasoning into a clear brief you can hand to the builder.

4. 04

   Build in small moves.

   Paste the brief into the builder, then work in small, testable steps. Tip: If you build with an MCP connection, you can have your LLM direct the build in small steps. This may surface items for you to address while building, helping you build something more aligned to your overall vision.

5. 05

   Test with real, messy data.

   Not the clean example. The ugly real thing (de-identified). That is where it breaks.

6. 06

   Audit with HUMAN.

   Run the output through the Human Stewardship Model. Facts, math, tone, bias, policy fit.

7. 07

   Ship to one real person.

   Hand it to a single colleague who has the actual problem. Watch them use it. Then widen.

Booklet: p. 10. Also listed under Resources.

Gemini Notebook is one way to turn a single trusted source into many formats: a summary, an FAQ, an audio overview. The lesson is the pattern, not the product.