Act 3 · Lesson 11 of 21≈8 minutes○ Saves on this device

Briefing the Machine: Prompting

By the end you'll be able to

Explain why prompts work: they're context that reshapes the probability landscape, not commands to an obedient mind.
Apply a briefing structure: objective, context, constraints, example, output shape.
Diagnose a weak output as usually a weak briefing, not a weak model.

Predict
Brief
Decide
Field note

⚡ 30-second warm-up — From Lesson 4: what actually decides which candidate token gets selected each step?

A weighted candidate list, shaped by everything on the desk, drawn from according to a tunable setting. Today's lesson is about deliberately shaping that list before the draw — the craft of briefing.

Revisit Lesson 4

"Write me something about our product" vs a five-line brief. Same model, same desk. Watch how differently they land.

Start with two requests

Two colleagues ask the same assistant for the same donor email. One types a single vague line. The other spends ninety seconds briefing it properly. Same model, same day, same desk size (Lesson 10) — wildly different results.

Commit to a prediction

The gap between the two outputs is mostly explained by…

This lesson changes one thing: you'll stop hunting for magic wording and start treating a prompt as a briefing — the same discipline you'd use handing a task to a sharp new contractor who knows nothing about your business.

Discover the mechanism

How it works Prompting as context engineering

Everything you've learned lands here. The model continues patterns (Lesson 1); attention weighs what's relevant from the desk (Lessons 8 and 10). So a prompt's job is to load that desk so the best continuation is also the one you want.

In practice that means five things: state the objective (what, and for whom); supply context the model cannot know — your situation, your data, your audience; set constraints (length, tone, format, what to avoid); show an example if format matters (patterns beat descriptions — you're speaking its native language); and add an uncertainty rule ("say unknown rather than guess" measurably helps).

Prompting is genuinely useful craft — and genuinely oversold. It's one layer of effective use, not the whole skill: the rest of this act adds knowledge, tools and workflow, which often matter more than wording.

real usage also involves system-vs-user message roles and techniques like worked-example framing — durable but secondary; this lesson's Go deeper names them. Lesson 20 turns the habit below into a repeatable organisational workflow.

Useful mental model Briefing a talented new contractor

Skilled, fast, zero knowledge of your business, unlikely to ask what you meant — and will confidently produce something whatever you say. Output quality tracks briefing quality almost linearly.

Contractor's general skill = the trained weights (Lesson 5). Your brief = the prompt and context.
Their portfolio-informed guess at what you want = a plausible continuation (Lesson 1). Examples you attach = the "example" layer.
Where it breaks: a real contractor has genuine goals and comprehension, and improves across jobs. The model has neither — "it wants to please you" is shorthand for "trained toward preferred continuations", and it doesn't get better at your account over time (Lesson 5's freeze, again).

Build a brief yourself

Conceptual illustration — canned outputs across three quality tiers, not a live model. Pick a scenario, then predict: which single layer will improve it most? Add it first, then the rest, and watch the output sharpen.

Scenario

Briefing layers Objective — what, and for whom Context — what it can't already know Constraints — length, tone, what to avoid Example — a sample of the voice or shape Uncertainty rule — say unknown, don't invent

Quality: — tick layers to build the brief —

Without JavaScript: a bare "write me something about our charity's work" produces a generic, unusable paragraph. Adding objective, context, constraints, an example and an uncertainty rule — one at a time — sharpens it step by step into a specific, send-ready appeal naming the actual programme and donor history, with no invented statistics. Constraints and context do more work than any "magic phrase".

Where this bites at work

In the field · Luka

Luka turns "write a fundraising appeal" into a five-layer brief with two examples of the charity's voice and a constraint against unsupportable outcome claims. The output goes from generic to send-ready in one pass. When it still needs a second pass, he runs a controlled loop rather than endless re-prompting: evaluate against written criteria, name the specific gaps, request improvement of only those gaps, verify anything load-bearing independently — AI critique of AI output is a drafting aid, never verification — and stop once another round would add less than it costs.

The stop rule is the professional skill: iteration has a real price in tokens, time and attention (Lesson 9) — "good enough for the stakes" is a decision, not a failure.

Where the picture breaks

There are no secret incantations that reliably beat structure and context — phrase-level tricks exist but are marginal, model-specific and perishable. And a contractor escalates when confused; the model's plausibility machinery (Lesson 1) fills gaps with confident invention instead of asking, unless you explicitly instruct it to check first.

Now make the call

Luka's donor appeal draft is flat and generic, and his board meeting is in an hour. What does he do?

Apply it to your work

This week, at your desk

Pick one recurring AI task and write it as a five-line brief: objective, context, constraints, example, uncertainty rule. Save it, and use the controlled improvement loop — with a written stop rule — for anything worth a second pass.

Print the five-line brief & improvement loop

Prove it to yourself

Two quick questions — untimed, retryable, and they count toward ★ Mastered (completed + decision + self-check all correct). This is what defeats the feeling of "that made sense" evaporating by next week.

Explain it in your own words

Two or three sentences for a colleague. Then compare against the rubric — it checks ideas, not wording, and only you see your note.

Evidence & review — how we know what this lesson claims

Claim register for Lesson 11
Claim	Type	Basis	Review risk
Prompts function as context that reshapes the model's continuation, not as commands to an agent	How it works	Architecture / primary research literature	Low
Structured briefs (objective, context, constraints, example, uncertainty rule) reliably improve output quality	Practice-based guidance	Provider and practitioner documentation	Medium
Named "magic phrase" prompting techniques are marginal, model-specific and perishable	Product behaviour	Provider documentation	High — reviewed quarterly
AI critique of AI output is a drafting aid, not independent verification	Practice-based guidance	Series standard, stated explicitly	Low
The briefing-builder demo	Teaching device	Deterministic canned outputs, labelled illustrative	Medium
The freelance-contractor analogy	Useful mental model	Limits stated in the lesson	Low

Concept review due: January 2027. Product-behaviour claims: quarterly.

Where you are in the machine

You're learning to load the desk deliberately. Select any layer for its role, lessons and your live progress.

The machine, bottom to top: training data (2) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).

Next: new senses

You can brief it well — but so far it's text in, text out. What happens when the machine gains eyes and ears?