# Ways of Working - Pocket Cheatsheet (Build-a-Skill)

> Print this one page and pin it next to your screen. Three habits, plain examples, nothing technical. Full source and all five ways: [ways-of-working.md](../ways-of-working.md).

The day teaches all five ways. For non-technical work - status updates, triage, summaries, drafts - these **three** carry the load. Get these right and the rest follow.

---

## 1. Describe first - *say what "good" looks like before you ask*

Decide the shape of the answer before you prompt. If you can't describe it, you're not ready to generate it. ([Way #1](../ways-of-working.md))

| 👍 Looks like | 👎 Not yet |
|---|---|
| "Give me a 4-column table: Summary, Decision, Risk (Red/Amber/Green), Action (owner + due). Max 8 rows. Factual tone." | "Summarise this for me." |
| "A 3-bullet update for the steering group, plain English, no jargon, under 80 words." | "Make this sound good." |
| "Rewrite as a polite chase email to a supplier, two short paragraphs." | "Reply to this." |

**Try it:** before your next prompt, write one sentence describing the output. Paste that sentence *into* the prompt.

---

## 2. A human signs it - *AI drafts, you review and own it*

Nothing leaves your hands unread. The AI is a fast assistant, not the approver - your name is on the output. ([Way #2](../ways-of-working.md))

| 👍 Looks like | 👎 Not yet |
|---|---|
| You read every line and catch the date the AI invented. | You paste the AI's text straight into the report. |
| You check the owner it assigned to each action is a real person. | You trust the figures because they "look right." |
| You delete the confident-sounding sentence you can't verify. | You forward it as "the system's analysis." |

**The 30-second habit:** before you trust any AI output, check the three things it most loves to invent - **dates, owners, and numbers.** Fix at least one. There's almost always one.

---

## 3. Mind the data - *treat every prompt like a public postcard*

Not a sealed envelope. No secrets, no client data, no PII, no confidential figures. Anonymise first. When in doubt, leave it out. ([Way #3](../ways-of-working.md))

| Before you paste | Swap it for |
|---|---|
| Real person's name | Person A / Person B |
| Client or supplier name | Client 1 / Vendor 1 |
| Confidential budget or figure | [REMOVED] or a rounded placeholder |
| ID numbers, account refs, emails | Strip them out entirely |

| 👍 Looks like | 👎 Not yet |
|---|---|
| "Person A flagged a delay on the Client 1 contract." | "Jane flagged a delay on the Acme £400k contract." |
| You scrub names from the notes, then paste. | You paste the whole email thread, signatures and all. |

**The rule of thumb:** if you wouldn't write it on a postcard, don't paste it into a chatbot.

---

## Carry these into the template

Each of these three is wired into the [SKILL-template.md](SKILL-template.md):

- **Describe first** → you write the **Output** section *before* the prompt.
- **Human signs it** → the **Guardrails** section names what to double-check.
- **Mind the data** → the **PRIVACY CHECK** sits inside **Inputs**.

> Want the other two? [Way #4 - small steps](../ways-of-working.md) (build one instruction at a time) and [Way #5 - keep it simple, write down the why](../ways-of-working.md) (save it where the team can find it). Full set: [ways-of-working.md](../ways-of-working.md).