Capstone: your AI-augmented project
Forty lessons of theory collapse into one real exercise. Take a brief, run it through the whole AI-augmented spine, and document exactly where AI led, where you verified, and where you stayed human.
Everything you've learned, on one real brief, in one sitting. This is where it becomes yours.
You've spent this course learning that AI is a plausibility machine: brilliant for diverging, dangerous when mistaken for truth. You've learned to diverge with AI and converge with your judgement, to treat it as the brilliant intern and never the architect of record, to keep a human in the loop. Now you prove it. Pick one real brief — a 3BHK fit-out, a small site, a courtyard house — and run it through the entire AI-augmented spine: concept imagery, then options, then a render, then a spec, then the checks, then the ethics pass. At every step you'll mark where AI led, where you verified, and where you stayed human. That document is your graduation. Let's build it.
One brief, the whole spine, fully documented
Concept imagery to options to render to spec to checks to ethics
The capstone is the whole course assembled into one pipeline. You'll move a single brief through six linked stages, each one drawing on a module you've completed.
Concept imagery (the mood and direction). Options (layout and parti, diverging wide). Render (your chosen direction, made vivid). Spec (the language-AI draft of scope and finishes). Checks (the red-list verification — dimensions, code, structure, products, prices). Ethics (copyright, data, liability and disclosure — Module 9, applied).
Notice the shape: it's the value curve made literal. AI leads hard at the front (concept, options, render), assists in the middle (spec drafting), and stays firmly in the passenger seat at the checks. The ethics pass wraps the whole thing. Your job isn't to let AI do the project — it's to conduct AI through it, deciding at every bar where it leads and where you take the pencil back.
You're not the AI's assistant. You're the conductor. The model plays; you decide the music.
For every stage: where AI led, where I verified, where I stayed human
The capstone's real product isn't the render — it's the decision log. As you run each stage, you'll keep three columns.
WHERE AI LED — the divergence, the speed, the first drafts: twenty facade directions, a dozen palettes, a fluent scope draft. This is the green list, where plausible-and-fast earns its keep. WHERE I VERIFIED — the red-list catches: the dimension you re-derived, the NBC clause you confirmed against the real bye-law, the product you checked was real and available, the price you sourced. This is the human staying in the loop. WHERE I STAYED HUMAN — the judgement no model holds: which option is right for this client, the taste call, the decision you'd stake your registration on.
This three-column log is the whole course in evidence. It proves you can use AI fearlessly because you know exactly where it leads and exactly where you don't let it. That's the difference between a designer who's augmented and one who's exposed.
The spine you carry out of this course
When your decision log is done, you've internalised the one idea this course exists to teach. Say it back in your own words and it's yours for good:
AI is a plausibility machine — it makes the most plausible next thing, which is not the same as true. So you diverge with AI and converge with your judgement. You treat it as the brilliant intern, never the architect of record. You keep a human in the loop wherever being wrong is costly. And because tools date fast, you learn the category and the judgement and let the tool names come and go.
That spine doesn't expire when the next model launches. It's the thing that makes you faster everywhere and safer where it counts — across every project, for the rest of your career. You came in wondering whether AI would replace you. You leave knowing how to conduct it. Now go run a real one.
Let the machine draft. You design. That was always the whole course.
Pick a brief with real compliance teeth — a small residential plot with setback and FAR constraints, or a courtyard house. Run the front of the pipeline AI-forward (concept imagery, layout options, a render off your own massing), then make the checks column do real work: re-derive every dimension, confirm every setback/FAR/NBC reference against the actual bye-law, and verify the structure. Your decision log should read like a QA trail you'd be comfortable showing in an inquiry. The capstone proves you can move fast at concept and stay watertight where you sign.
Take a 3BHK or villa fit-out brief. Lead hard with AI on mood boards, palette and styling directions — your home turf — and a staged render to align taste. Then let the verified column bite: turn the AI's invented furniture and finishes into a real, sourced, costed FF&E schedule, and confirm clearances and ergonomics. Your decision log should show exactly where concept imagery ended and specified, buyable reality began. That's the discipline that stops the 'but the render showed marble' dispute and proves you direct AI rather than being dazzled by it.
You can do this entire capstone on free and cheap tools, and it's the single best portfolio piece you'll make. Pick any real brief — even a friend's flat — and run the full spine, because as a solo you ARE every column: the AI-led divergence, the verification, and the human judgement. Keep the decision log honest, including where you caught the model being confidently wrong. A clean, documented AI-augmented project says more to a future client or studio than any certificate: it proves you use these tools like a professional, not a tourist.
Midjourney / FLUX + Studio Matrx Moodboards
Stage: concept imagery + options
Diffusion engines for mood and direction; Moodboards (Style Explorer) as a worked example for fast style exploration. AI's highest-value zone — diverge wide. The limitation: these invent products, finishes and structure that look real and aren't, so everything here is a direction to verify later, not a spec.
Veras / a render tool + Maket for layout options
Stage: render + layout optioneering
Render your own geometry (Veras, inside your CAD) and generate rapid layout options (Maket). Keeps your control high. The limitation: an AI plan's dimensions and areas are not a buildable drawing, and the render is not a contract document — both feed the checks column, not the issue set.
Claude / ChatGPT for the spec + BOQ draft
Stage: spec drafting
Claude holds long documents (full scope, spec); ChatGPT is strong at tables (BOQ, schedules). A fast first draft of the boring-but-necessary language. The limitation: every code reference, product, dimension and price it produces is red-list — confident and unverified — so it goes straight to the checks column.
Your before-my-stamp checklist + clearance card (Lessons 9.1-9.3)
Stage: checks + ethics
The verification gate and the copyright/data clearance you built earlier in this module, applied to the capstone deliverables. This is the human-in-the-loop made concrete. The limitation: it only protects you if you actually run it on every AI-touched output before that output carries weight.
“If I do the capstone well, the AI should be able to take my brief and produce a near-finished, buildable project that I just tidy up.”
That's the exact mindset the whole course warns against. The capstone isn't AI doing the project with you tidying — it's you conducting AI through a project, leading hard where it pays and taking the pencil back where it bites. A 'near-finished' AI output is a confident illusion: the dimensions, code, structure, products and prices are unverified plausibility. The point of the three-column log is to make your judgement visible at every bar. The value is in the conducting, not the output.
Workshop — the capstone: run a real brief through the full AI-augmented spine
This is it — the project that pulls the whole course together. Take one real brief and conduct AI through all six stages, keeping the three-column decision log and scoring yourself against the rubric at the end. Set aside a focused half-day.
One real brief (a 3BHK fit-out, a small plot, a courtyard house). Free: any chat + image AI. Better: Midjourney/FLUX + a render tool + Claude/ChatGPT. Plus your checklists from 9.1-9.3.
CAPSTONE DECISION LOG - Brief: ____________
| WHERE AI LED | WHERE I VERIFIED | WHERE I STAYED HUMAN
--------------+-------------------+--------------------+----------------------
1 CONCEPT IMG | mood / directions | (n/a yet) | which feel is right
2 OPTIONS | layout variants | areas plausible? | which parti to take
3 RENDER | the vivid image | not a contract doc | the design intent
4 SPEC | scope/BOQ draft | codes/products/$$ | what we actually spec
5 CHECKS | (AI steps back) | dims/NBC/structure | the professional call
6 ETHICS | (AI steps back) | licence + data ok | disclosure + the stamp
SELF-ASSESSMENT RUBRIC (score each 0-3):
[ ] Diverged wide with AI at the front (concept + options)
[ ] Caught the model being confidently wrong at least once
[ ] Re-derived every red-list item (dims, code, products, prices)
[ ] Render never treated as a contract document
[ ] Cleared copyright + data + licence (Lessons 9.1-9.2)
[ ] Verified everything before it carried my stamp (9.3)
[ ] Decision log honestly shows led / verified / human
[ ] I could defend every stage to a client or an inquiry
Score /24 -> 18+ augmented professional . 12-17 close, tighten checks . <12 re-run the spine- 1Choose your brief and write it at the top of the log. Real is better than perfect — a friend's flat counts.
- 2Stage 1-3, diverge wide: generate concept imagery, then layout options, then a render of your chosen direction. Fill the 'AI LED' and 'STAYED HUMAN' columns as you go — which feel is right is yours, not the model's.
- 3Stage 4, draft the spec: have an LLM draft the scope and a BOQ. Drop every code, product, dimension and price straight into the 'VERIFIED' column as a to-do — none of it is trusted yet.
- 4Stage 5, run the checks: work your before-my-stamp checklist from Lesson 9.3 — re-derive dimensions, confirm NBC/bye-law references, verify products and prices. Record at least one place the model was confidently wrong.
- 5Stage 6, the ethics pass: run the copyright + licence clearance (9.1), confirm no confidential data went to a public model (9.2), and write your client-disclosure line (9.3).
- 6Score yourself on the rubric, total out of 24, and write three sentences: where AI led, the best catch you made, and the one judgement you'd stake your name on.
- 7Keep the finished log as a portfolio piece — it's the single clearest proof that you conduct AI like a professional.
You’ll walk away with
A complete, documented AI-augmented project: a three-column decision log across all six stages, a rubric score out of 24, and a portfolio-ready artefact that proves you can diverge with AI, converge with your judgement, and stay accountable end to end.
If a full half-day isn't possible yet, start here.
- 01Run just Stage 1 and Stage 5 on a real brief: generate concept imagery, then immediately run the red-list checks on anything it implied. The gap between the two is the whole course in five minutes.
- 02Take your finished log and circle the single best 'confidently wrong' catch you made. That instinct — spotting plausible-but-untrue — is what you actually graduated with.
The capstone runs one real brief through the whole AI-augmented spine — concept, options, render, spec, checks, ethics — with a three-column log of where AI led, where you verified, and where you stayed human. The value isn't the output; it's the conducting. Do it once, honestly, and the course's spine is yours for every project that follows.
Run a real brief through six stages: concept imagery, options, render, spec, checks, ethics. Keep three columns — AI led, I verified, I stayed human. Diverge with AI, converge with your judgement; brilliant intern, never the architect of record; human in the loop; learn the category, not the tool. Let the machine draft. You design.
What does an AI-augmented design workflow actually look like end to end?
Six linked stages: concept imagery for mood and direction, layout options to diverge wide, a render of the chosen direction, an AI-drafted spec and BOQ, a human verification pass on every red-list item (dimensions, code, structure, products, prices), and an ethics pass (copyright, data, liability, disclosure). AI leads at the front, assists in the middle, and steps back for the checks — with a human conducting throughout.
How do I document where AI helped versus where I made the decisions?
Keep a three-column decision log for each stage: 'where AI led' (the divergence and first drafts), 'where I verified' (the red-list items you re-derived and confirmed against real sources), and 'where I stayed human' (the judgement calls only you can make). This log is your proof of professional process — useful for client disclosure, your own learning, and a portfolio piece that shows you direct AI rather than being directed by it.
What's the single idea to take away from this whole course?
AI is a plausibility machine — it makes the most plausible next thing, which is not the same as true. So diverge with AI and converge with your judgement; treat it as the brilliant intern, never the architect of record; keep a human in the loop wherever being wrong is costly; and learn the tool category and the judgement, not this month's interface. That spine outlasts every model. Let the machine draft. You design.
_That's the course. The tools will change; the spine won't. Carry it into every brief — and let the machine draft while you do the one thing it never can: design._