# The 5 Ways We Work With AI

> **Single source of truth.** This file holds the five "ways of working" that anchor the whole Innovation Day. Both sessions teach the same five; each practises the relevant ones hands-on. Every other document references this file - do not restate the Ways elsewhere, link here instead.

**Opening line for facilitators:** *"AI doesn't change the standards - it changes how fast we hit them. Same destination, faster road. Here are the five rules of the road."*

These five are a plain-language distillation of the company's [corporate standards](../corporate-standards-main/) (engineering + quality principles). Each cites the real `data-cite` ID it comes from, so the standards owners can validate it and nothing drifts.

---

## The five

### 1. Describe before you generate
**Say what "good" looks like before you ask.** A clear spec beats a clever prompt. Write the goal, the inputs, and the shape of the output first - then let the AI fill it in. If you can't describe the result, you're not ready to generate it.

- *Standards anchor:* `ENG-PRIN-DOC-STMT` (Document Decisions, Not Code - decisions must outlive the conversation they were made in) and `QUAL-PRINC-SHIFT-LEFT` (prevent defects at inception, not after).
- *In the room:* the Skill template puts the **Output** section above the prompt; the Build session forbids feature code before the contract is written.

### 2. A human always signs the work
**AI drafts; a person reviews and owns it.** Nothing ships unread. The AI is a fast pair, not the approver - you remain accountable for what goes out under your name.

- *Standards anchor:* `ENG-PRIN-REVIEW-STMT` (no change reaches trunk without review by another human; no self-merge) and `ENG-PRIN-OWN-STMT` (You Build It, You Own It).
- *In the room:* the Skill session's "find one thing the AI got wrong" beat; the Build session's mandatory human review gate (AI may assist the review, but a human approves).

### 3. Mind what you feed it
**Treat every prompt like a public postcard, not a sealed envelope.** No secrets, no client data, no PII, no confidential figures. Anonymise first - swap real names for "Person A", vendors for "Vendor 1". When in doubt, leave it out.

- *Standards anchor:* the standards' security & data-classification posture (least-privilege, protect sensitive information, maintain trust and compliance).
- *In the room:* the Skill template's "PRIVACY CHECK" line inside Inputs; the optional "run the script yourself" path so real data never leaves your machine; synthetic-data-only rule for both sessions.

### 4. Small steps, not big leaps
**One change at a time - easy to check, easy to undo.** Don't ask AI for a giant deliverable in one shot. Build in reviewable increments; verify each before the next.

- *Standards anchor:* `ENG-PRIN-INCR-STMT` (many small commits over large batches; easier to review and revert) and `ENG-PRIN-TRUNK-STMT` (short-lived branches, frequent integration).
- *In the room:* the Build session's "stub + failing test first, then implement"; the Skill session's iterate-one-instruction-at-a-time guided build.

### 5. Keep it simple, write down the why
**The simplest thing that works, and a note on why you chose it.** Don't accept speculative complexity from the AI (YAGNI). Record the decision so the next person - and future you - isn't guessing.

- *Standards anchor:* `ENG-PRIN-SIMPLE-STMT` (KISS, YAGNI, don't reinvent the wheel) and `ENG-PRIN-DOC-STMT` (capture the decision in a durable artifact).
- *In the room:* the Skill's "Reuse notes" section; the Build session's one-page ADR capturing a real trade-off.

---

## The take-away card (front / back)

**FRONT - the five imperatives:**
1. Describe before you generate.
2. A human always signs the work.
3. Mind what you feed it.
4. Small steps, not big leaps.
5. Keep it simple, write down the why.

**BACK - good vs. not-yet:**

| Way | 👍 Looks like | 👎 Not yet |
|---|---|---|
| Describe first | "Output = 4-col table: Summary, Decisions, Risks (RAG), Actions (owner, due)." then prompt | "Summarise this for me" |
| Human signs | You read every line and fix the one made-up date | You paste the AI output straight into the report |
| Mind the data | Names replaced with Person A/B before pasting | Real client name + budget in a public chatbot |
| Small steps | Stub → test → implement → review | One mega-prompt, 200 lines, hope it works |
| Simple + why | One ADR line: "chose on-the-fly averages - dataset is tiny" | Three new abstractions "for later" |

> Printable version lives in [`assets/`](assets/). The cheatsheet and both session cards pull their wording from this file.