The 8-part brief anatomy; constraints as direction; references and the consent line; critique loops.
All tracks · 4 hours · drop-in friendly
Prompting Like an Art Director: Briefs, References, and Critique Loops
You leave with a reusable briefing system — a creative brief that works on AI (and on humans), plus a critique rubric tuned to your track.
01This Week at a Glance
You will make: one AI-ready creative brief for a real project + your track critique rubric.
One brief adapted live for six disciplines — and turned into a rubric.
Prompt surgery in teams → your real brief → your rubric.
Before/after pairs, authorship lines, rubrics that bite.
Why this matters in real work
The skill gap this term isn't "who knows the magic words" — it's who can give creative direction. Briefing is an art director's craft, and everything here also makes you better at briefing printers, photographers, and collaborators. The AI is just the cheapest, most patient practice partner you'll ever get.
Kill this myth now
There are no magic words. "Prompt engineering" for creative work is ~90% classic creative-brief discipline — something agencies have done for a century. The other 10% is machine quirks, covered below.
02Key Concepts
1 · Prompting is briefing, not spellcasting
| Part | What it does | Without it |
|---|---|---|
| Role | "Act as a practical creative director" — sets voice and standards | Generic assistant tone |
| Context | Project, audience, where the work lives | The model invents a (wrong) default |
| Constraints | Format, palette, duration, non-negotiables | The average of the internet |
| References | Qualities you want — shown or described | "Style" becomes cliché |
| Avoid list | What you explicitly don't want | The clichés walk right in |
| Output format | "5 directions, each with a risk" | One long confident essay |
| Criteria | How drafts will be judged | Nothing to critique against |
| The decision line | "Don't pick a winner — that's my job" | The model decides; you rubber-stamp |
2 · Constraints are creative direction
Counterintuitive but reliable: more constraints → more interesting output. "Poster, any style" returns the global average. Tight constraints force the model into corners — and corners are where interesting work lives.
Watch out
Models confidently ignore constraints they can't satisfy (exact word counts, arithmetic, precise layouts). Constrain direction; verify compliance yourself. Tip: "list my constraints back before answering."
3 · References: show, don't just tell
- Telling: "1970s print feeling" — okay.
- Telling precisely: "flat colour, heavy grain, slight misregistration, like a 1970s risograph gig poster" — much better. Name qualities: materials, light, colour behaviour, texture, type feeling.
- Showing (few-shot): paste 1–3 examples of what good looks like. The fastest taste transfer.
The consent line — course rule
We don't prompt with living artists' names. Describe qualities; use historical movements as qualities. Beyond the ethics: naming qualities forces you to understand why you like a reference — and that understanding is the transferable skill.
4 · The critique loop: your hostile reviewer on demand
The highest-value prompting skill — it works even if you never generate anything. Work + your criteria + a demand for specifics + your decision about what to accept. A model with no criteria gives compliments (the word is sycophancy). A model holding your rubric gives you a second pair of eyes at 2am.
5 · Iteration etiquette
- React with specifics — "No. 2 is closest but the tone is corporate; push it stranger."
- One variable at a time when close; start over when not.
- Fresh chats for genuinely different directions (context anchoring is real).
- Three failed revisions in a row = the brief has a hole.
- Briefs live in docs, never only inside chats.
03This Week's Words
PromptEverything you give the model: instruction, context, constraints, examples.
System promptStanding instructions framing every reply — a persistent job description for your tool.
Creative briefStructured statement of project, audience, goal, constraints, references, criteria.
ConstraintsThe limits you set. Where creative direction lives.
Style referenceThe qualities you want, named precisely — not a living artist's name.
Few-shot / examplesShowing 1–3 examples of "good" instead of describing it.
Negative prompt / avoid listExplicitly stating what you don't want. Cheap and powerful.
Critique loopFeeding work back with criteria and demanding structured criticism.
IterationGenerate → reject → refine. Selection is the skill.
SycophancyThe model's habit of flattering you. Counter with criteria that bite.
04Demo: One Brief, Six Tracks
I made this, so can you
One base brief (Week 1's degree-show poster) adapted live for motion, UX, and 3D — with illustration, graphic design, and animation versions prepared. Then the brief becomes a critique rubric.
Prepared demo artifact
Instructor backup / finished example: one-brief-six-tracks.html. Use it if the live tool run fails, or to reveal the finished structure after critique.
- The spine never changes: audience, goal, tone survive every translation.
- The moving parts: constraints, deliverable specs, references, critique criteria.
- The planned failure: one adaptation keeps an irrelevant constraint (riso colours in a website brief). Catching that is your job — every time, with every tool.
- The closing move: a 6-line rubric where every line cites a real constraint — and generic lines get rejected and regenerated.
MotionSpecs and timing for a 15s story; type-in-motion constraints; sound note.
UXWayfinding, accessibility, mobile-first, performance constraints.
IllustrationMedium, silhouette readability, print process constraints.
3DMaterials, sightlines, fabrication and budget constraints.
05Step-by-Step Workflows
Prompt surgery (the diagnostic loop)
- Run once; read closely. What kind of bad? Generic = missing constraints/references. Wrong direction = missing context. Essay wall = missing output format. Bossy = missing decision line.
- Name the missing organ using the anatomy table.
- Transplant it in. Add the one missing part. Re-run.
- Compare side by side. Note what changed — your evidence trail.
- One more pass. Two surgical passes beat ten random tweaks.
Build your track rubric (10 minutes)
- Take your brief's constraints and goal.
- Convert each into a pass/fail question. ("Must work as a phone story" → "Does type survive at story-size?")
- Add the course's five standing critique questions (rubric template).
- Test on any AI output. If a mediocre draft passes everything, sharpen until it bites.
06Prompt Recipes
The full brief skeletonROLE: Act as a practical [creative director / design director / story editor] for my discipline: [track].
PROJECT: [what's being made, for what occasion]
AUDIENCE: [who actually encounters it — be narrow]
GOAL: [what they should think/feel/do — one sentence]
CONSTRAINTS: [format, dimensions, colours, duration, budget, deadline, non-negotiables]
REFERENCES: [2–3 named QUALITIES you want, not living artists]
AVOID: [clichés, competitor looks, anything off-limits]
OUTPUT: [exact shape: e.g., 5 directions, each with one-line concept, layout in words, one risk]
Don't pick a winner — that's my job. Ask up to 3 clarifying questions first if anything is unclear.
The rubric builderHere is my creative brief: [paste]. Build me a 6-line critique rubric. Each line: one pass/fail question I should ask of any draft, tied to a SPECIFIC constraint or goal in this brief. No generic design-school questions — if a line could apply to any project, replace it.
The track adapterHere is my base brief: [paste]. Rewrite it for a [motion designer / illustrator / UX designer / 3D artist] producing [specific deliverable]. Change only what must change: deliverable specs, constraints, references, critique criteria. Highlight what you changed and why.
The prompt diagnosticianHere is a prompt I wrote and the output it produced: [paste both]. Diagnose the prompt like an art director reviewing a bad brief: which of these is missing or weak — role, context, constraints, references, output format, criteria, decision ownership? Rewrite the prompt fixing only the weakest two parts.
The anti-sycophancy clauseDo not compliment me. Do not soften. If the work is generic, say where and why. I am asking for the review a tired, skeptical creative director would give at 6pm on a Friday.
07Lab: Surgery → Brief → Rubric (70 min)
Cold start friendly — Week 1 attendance not required.
Part A — Prompt surgery, in teams (25 min)
Teams of 2–3. Three genuinely bad prompts. For each: predict what's wrong before running → run → fix the two weakest organs only → re-run → save before/after.
The patients
- "Design me a logo for my brand. Make it modern and clean."
- "Write something for my portfolio about my project."
- "Give me ideas for my animation."
Part B — Your AI-ready brief (25 min, solo)
Full skeleton for a real current project. Run it; one surgery pass. Then mark ✏️ every line only you could have written — your audience knowledge, your taste, your constraints. That's the authorship evidence.
Part C — Your track rubric (20 min)
Six lines from your brief. Test on Part B's output: at least two lines must fail. If everything passes, sharpen the rubric, not your feelings.
Zero-generation track
Part A as paper diagnosis (predict, don't run). Parts B and C are writing tasks — the actual deliverable — with the model as optional critic only. Equal credit.
08Integrity Check
This week's four
- Whose taste is in the brief? If the model wrote your references and you kept them unread — whose direction is this?
- Are your references qualities, or a living artist's name wearing a trench coat?
- If this brief produces portfolio work, your disclosure note starts now: "Brief and direction mine; AI used for [x]."
- Briefing with a classmate's project or client-confidential material? Permission first.
09Troubleshooting
| Problem | Fix |
|---|---|
| Output ignores a constraint | Repeat it at the END of the prompt; ask it to "list my constraints back before answering" |
| Everything comes back as bullet lists | Specify format: "flowing prose, 120 words, no lists" |
| It agrees with everything | Sycophancy — append the anti-sycophancy clause; give criteria with teeth |
| Endless clarifying questions | Cap it: "ask up to 3 questions, then proceed with stated assumptions" |
| Five outputs, all weirdly similar | Context anchoring — fresh chats for genuinely different directions |
| My rewrite made it worse | Too many organs at once. Revert, change one part, compare |
10Resources
- VideoPrompting 101 — Anthropic (Code w/ Claude)Embedded above. The official version of everything this page teaches.
- VideoVerify before teachingAI prompt engineering: A deep dive — AnthropicHour-long roundtable for students who want depth.
- TutorialLinkedAnthropic's Interactive Prompt Engineering TutorialFree, hands-on, 9 chapters. Techniques transfer to every chat tool.
- ArticleLinkedThe 2026 Guide to Prompt Engineering — IBMSober corporate reference — proof that prompting is documentation skill, not wizardry.
- DocsLinkedOpenAI prompting guideCompare with Anthropic's and notice they agree: context, constraints, examples.
- ArticleLinkedWrite a Design Brief That Keeps Projects on Track — The FuturBrief discipline from the client-services world: goal in one sentence, parameters neither stifling nor wide open. Same anatomy, aimed at humans.
- VideoVerify before teachingLive Design Critique with Chris Do — The FuturA human creative director running a critique loop — then notice this page's rubric prompts are the same moves.
Optional extension
- Set up your tool's custom instructions / Project system prompt: role line, track, standing standards.
- Anthropic's interactive tutorial, chapters 1–3.
- Hard mode: write the worst possible brief that still looks professional; trade with a neighbour and diagnose each other's sabotage.
Next week: Research, Moodboards, and Concept Development Without Copying →