# Workflow Builder Template

Document one repeatable AI workflow. Fill out one copy per task from your Stage 1 leverage map — five copies total.

**Workflow name:** [NAME — one task, one output]
**Leverage task (from Stage 1):** [TASK]
**Built:** [DATE]   **Last revised:** [DATE]

## Part 1 — Trigger

What starts this workflow? Pick one. Be specific enough that you can't miss it.

- [ ] Calendar: [E.G., FIRST MONDAY OF EACH MONTH]
- [ ] Event: [E.G., CLIENT FILE ARRIVES, GRANT NOTICE POSTS]
- [ ] Request: [E.G., PARTNER ASKS FOR A CASE SUMMARY]

My trigger: When [THIS HAPPENS], I run this workflow.

## Part 2 — Input prep

List everything you gather before the first prompt.

1. [INPUT — e.g., this month's P&L export]
2. [INPUT — e.g., last month's summary as a style example]
3. [INPUT — e.g., word limit, deadline, audience]

**Format check:** [HOW INPUTS MUST BE PREPARED — e.g., PDF converted to text, tables pasted as plain text]

**Confidentiality line:** These never go into a prompt: [CLIENT NAMES / STUDENT RECORDS / SSNs / PRIVILEGED MATERIAL / OTHER]. Replace with: [PLACEHOLDERS YOU USE INSTEAD].

## Part 3 — Prompt chain

Most tasks need 2-4 prompts, not one. Pull starting prompts from the prompt library for your profession, then adapt them. Write the full prompts here — the page should work without your memory.

| Step | Job | Prompt | Output feeds |
|---|---|---|---|
| 1 | [E.g., extract key facts] | [FULL PROMPT] | Step 2 |
| 2 | [E.g., draft the document] | [FULL PROMPT] | Step 3 |
| 3 | [E.g., cut to length and format] | [FULL PROMPT] | Verification |

## Part 4 — Verification

AI invents facts, mangles numbers, and misreads dates. Name at least two specific checks. "Review it" is not a check.

1. [CHECK — e.g., tie every number to the source spreadsheet]
2. [CHECK — e.g., open and confirm every citation]
3. [CHECK — optional third]

**Who verifies:** [YOU / NAMED COLLEAGUE]   **Time budgeted:** [MINUTES]

## Part 5 — Output standard

Define done.

- **Format:** [LENGTH, FILE TYPE, TEMPLATE USED]
- **Quality bar:** [WHAT A GOOD ONE ALWAYS CONTAINS]
- **Delivery:** [WHERE IT GOES, WHO GETS IT, BY WHEN]
- **Reject if:** [CONDITIONS THAT SEND IT BACK THROUGH THE CHAIN — e.g., any unverifiable claim, any number that doesn't tie out]

## First-run notes

Run the workflow once on real work before you call it built.

- What broke: [NOTES]
- What I changed: [ONE REVISION]
- Run time: [MINUTES]

## The four-question test

The workflow is done when you answer yes to all four:

1. Could a competent colleague run this from the page alone?
2. Is the confidentiality line written, not assumed?
3. Are the verification checks specific enough to fail something?
4. Has it produced one real, delivered output?

Part of the ADAPT framework — vetslearn.com/framework