Day 2: ACR and DRA Thinking Frameworks, Connectors and Scheduled Tasks, the Image-to-Video Ad Pipeline, MoSCoW + 5 Whys, and Building LearnVault Back-End-First in Lovable

Three moves, in order. ASK: 'before you answer, ask me every question you need to do this well' - the model interviews you and the answers become the context it would otherwise invent. CHECK: 'list the assumptions you are making and flag the ones you are least sure of' - surface the guesses before they harden into output. RECOMMEND: only now 'give me your recommendation, with the trade-offs'. Demoed on a marketing-plan request: the interview surfaced budget, audience and channel constraints the one-line prompt never had, and the plan went from generic to specific.

DESIGNER frames the problem and the ideal outcome; RESEARCHER gathers evidence for and against, with sources; ANALYST weighs the evidence and decides. Run as three separate chats (or three projects) so each role's context stays clean, passing only the artefacts forward. The 'raise the stakes' trick: telling the model 'a CEO will read this' or 'my job depends on the accuracy' produced visibly tighter, more careful output in the live comparison - the trainer's explanation is that stakes language shifts the model toward its more careful training examples. The McKinsey prompt: 'You are a McKinsey partner. Interview me one question at a time until you can write the brief' - ACR and DRA in one line.

Connectors (Claude, ChatGPT) are vendor-built integrations you switch on; MCP is the open protocol underneath; Composio is a marketplace of hundreds of pre-built MCP servers for when the tool you need has no native connector. Scheduled tasks (Claude) and routines (ChatGPT) run a prompt on a timer with the connectors it needs. PROCEDURE - the Gmail triage build: connect Gmail -> write the prompt ('every morning at 7, read unread mail from the last 24 hours, label each as Action / FYI / Ignore, summarise the Action items in one table, draft a reply for each Action item and leave it in Drafts') -> set the schedule -> test once manually -> confirm the drafts appear. Guardrail stated: draft, never send.

Jordan's template, field by field: SUBJECT (what, doing what), ENVIRONMENT (where, time of day), STYLE (photoreal, editorial, 3D, illustration - with a reference artist or era), LIGHTING (golden hour, softbox, rim), CAMERA (lens, angle, depth of field), MOOD (one or two adjectives), NEGATIVE (what must not appear - text, extra fingers, watermarks). Write it as one paragraph, not a list, for most models. Demoed by taking a bland prompt through each field and watching the output converge on the brief. Aggregators - Krea and FAL - give one interface and one bill across Flux, Nano Banana, Kling, Veo and the rest, so you pick the model per job instead of per subscription.

PROCEDURE. Phase 1, stills: product photo + the seven-field prompt into Nano Banana 2 for four to six hero frames (consistent product, varied scene). Phase 2, motion: each still into Kling with a motion prompt ('slow push-in, steam rising') for 5-second clips; regenerate the ones that warp. Phase 3, voice: script into ElevenLabs, pick a voice, export. Assembly in Premiere (or CapCut): clips on the timeline, voice-over, music bed, captions. Cost math on air: a few dollars of aggregator credits per finished ad. Caveats: hands, text and logos still fail; keep the product static and let the camera move.

MoSCoW: Must have (the product doesn't exist without it), Should have (important, not launch-blocking), Could have (nice), Won't have (this version - written down so it stops being argued). 5 Whys: ask 'why' five times from the surface complaint to the root cause - the restaurant example goes from 'we need an app' to 'the phone rings during the dinner rush and orders get lost' in four whys, and the Must-have becomes an order-capture form, not an app. Sukhin's rule: run the 5 Whys before MoSCoW, or you will scope the wrong product perfectly.

PROCEDURE for LearnVault (a personal learning-resource tracker). Step 1: write the PRD in a chat (ACR-style interview) - entities: users, resources, tags, progress. Step 2: in Lovable, connect Supabase and ask for the schema first: 'create tables for resources, tags and progress with row-level security so users see only their own rows'. Step 3: enable email auth. Step 4: only now 'build the dashboard screen listing my resources with filters by tag and status'. Step 5: iterate one screen at a time; when a build breaks, paste the error back rather than re-describing the feature. Publish, then share the URL. Stated reasons for back-end-first: the schema is the contract, the screens are cheap, and retro-fitting auth onto a finished UI is where most beginner builds die.

PROCEDURE: new notebook -> add sources (PDFs, Drive docs, YouTube links, pasted text) -> before asking anything, open Configure and set the persona and goal ('you are a patient tutor; I am preparing for X; quiz me after each explanation') -> use the chat with citations back to the source passages -> generate the audio overview for a commute-length recap. Everything it says is grounded in your sources, so it is the safe place for private documents; it will say 'not in your sources' rather than invent. Used on stage with the boot camp's own session notes.

Happenstance connects to your own accounts (LinkedIn connections, Gmail contacts, X follows) and answers natural-language questions about the people in them, with the reason each match qualifies. Demoed for a hiring question and an intro request. Positioned as the people-layer that ChatGPT and Claude do not have.

Sukhin's closing frame for the whole day: tools change monthly, so copying a workflow (cooking) has a shelf life of weeks; understanding the pattern behind it (the framework, the build order, the reason a prompt is structured that way) is what transfers. The homework is framed this way: rebuild LearnVault for your own domain, not the tutorial's.