AI content creation
Practise product images, video, original audio and repeatable content workflows with concrete quality checks.
37 lessons in 12 modules
What you'll practise
- Write an image or video brief and check product identity across versions.
- Develop original audio with an iteration budget and a release checklist.
- Plan, review and organise a useful set of content for a defined audience.
Before you start
Use product facts and reference assets you have permission to use, or the clearly labelled fictional course examples. Generation tools may require separate credits or subscriptions.
Course syllabus
Module and lesson summaries follow the club curriculum. Open any module to read its outline.
Image models compared, with prompts you can reuse
Four models, one brief: the honest scorecard
Use a shared product reference and brief to compare available image tools. Log exact mode, all attempts, reference fidelity, accepted images and actual cost; separate text-only from reference-conditioned runs. Complete a scorecard and a provisional choice, or report that nothing passed. The Library contains a fictional worked method, not an attributed store benchmark or a verified model winner.
One prompt structure that survives a model switch
Build a prompt in six slots: subject, setting, light, lens, mood, and what must not appear. Translate it for your chosen models, checking supported controls and keeping the same creative direction. Save the master and translations in the prompt bank. The ChatGPT Images 2.5 exercise adds a product-preservation instruction and one checked edit; compare the actual result before reusing it for a second product.
Edit the image you have instead of rolling again
Practise three isolated fixes with instruction editing and masks, then compare what held against the original. Try a short ChatGPT Images 2.5 edit sequence, accepting each version only after checking the product and surrounding frame. Save the reviewed hero image, rejected versions and a drift log. Use that evidence to decide when to continue editing, return to an accepted version or revise the whole brief.
ComfyUI without the fear: one workflow, saved
You open ComfyUI, load the Flux workflow from the starter repo and learn the five nodes that matter: loader, prompt, sampler, decode, save. You change two things only, the resolution and the seed, then run a batch of eight variations from one prompt with the seed fixed. ComfyUI earns its setup time when you need fifty product images on the same background, not for a one-off. You leave with a saved workflow file, eight consistent variations, and a clear line between the jobs that go to a hosted model and the jobs that go here.
Product and ad creative in an afternoon
Product shots with Nano Banana from one phone photo
Use an authorised product photo to practise five scene briefs and an edit pass. Check shape, material, label and composition against the original at full size. Save the actual prompts, attempts and acceptance decisions. The Library's fictional arithmetic teaches cost per usable image; listing-ready outputs must be generated and reviewed by you before they can be claimed.
Twelve static ads from one angle, sized for Meta
Build a grid with three hooks across four static visual formats, then export at 1:1, 4:5 and 9:16. Keep factual copy and editable text layers for repeated variations. Try one ChatGPT Images 2.5 layout alongside the layered design. Verify exact lettering, hierarchy, spacing, product geometry and each placement's safe zones before accepting an export.
A UGC-style video ad without hiring a creator
Write a factual six-shot demonstration with a synthetic presenter, product inserts and a timed voice script. Use the ChatGPT Images 2.5 exercise for UGC-style stills and storyboard continuity checks. Motion, voice and video assembly remain separate steps. Work within a recorded allowance and review product fidelity, captions and disclosure; a finished twenty-second ad requires an assembled, inspected export.
The handoff to Meta: naming, sizes and a first test
You select two or three variations that test one question, reserving the video for a separate format comparison. Define the outcome, tracking checks, affordable spending cap and review date before launch. The three-day candle budget is illustrative; a short run does not guarantee enough evidence to choose a winner. Distinguish exploratory ad-set delivery from a controlled split, and reported clicks from attributed purchases. The recorded Cirka CPC comparison uses the last-30-days snapshot collected on 2 September 2026, not the 90-day report, and cannot establish profit. You leave with a measurement plan that records uncertainty and allows an inconclusive result without automatically extending spend.
Video models: what each one is good at
Veo, Kling, Runway, Higgsfield: one shot, four ways
Compare one motion prompt in currently available video modes, recording reference inputs, settings, attempts, actual usage and failure timestamps. Keep sound, camera and product fidelity as separate checks. The Library supplies a written method and fictional cost example. No summer render log, universal model ranking or finished footage is claimed; retain actual outputs before reporting your own comparison.
Image-to-video: keeping the product real in motion
Use the same product still in available image-to-video modes and write motion-only instructions. Check the product at the start, middle and end, then review every frame through playback. A reference or supported final frame can guide a reveal but cannot guarantee identity between frames. Save accepted and rejected takes and a provisional comparison based on your actual outputs.
Camera language: prompts that direct, not describe
You learn the twenty words a video model reads: shot size, the move (push-in, orbit, handheld, static), lens, speed and duration. Then you write a six-shot sequence for a thirty-second club video, one line per shot, and phrase each line the way its model wants it: a preset in Higgsfield, a sentence in Veo. You also learn the four moves that fail most often, fast orbits, whip pans, walking through doors and anything with a mirror, and you design around them. You leave with a shot list you can hand to any of the four models and a cheat sheet next to the screen.
Cutting generated clips into a thirty-second piece
You bring six generated clips into CapCut, keep the best two to four seconds of each, and cut on movement rather than at the end of the clip. You lay the voice track first, then the clips, then captions, then a bed from your own Suno track. Export at 9:16 and 16:9 from one timeline. The mistake is using all eight seconds of every clip, which is how AI video gets its slow, floaty feel. You finish with a thirty-second piece, a saved export preset, and the ffmpeg script from the repo for burning captions into a batch without opening an editor.
Consistent characters across a series
A character bible you can regenerate on demand
Write a short character bible with fixed visible traits, then plan front, three-quarter and profile references. Use an authorised or clearly fictional adult and inspect each generated view against the anchor. Keep the text, settings and reviewed images together. The exercise does not establish a founder cover-production history or guarantee that a model preserves identity.
Same face, new scene: reference tools compared
Compare five character scenes using reference workflows in available tools; ChatGPT Images 2.5 can replace one comparison row. Count face and signature-item matches, mark unobservable details and record all attempts. Check one local edit separately so it does not distort the scene comparison. Save a provisional still-image choice and reviewed references for the separate motion exercise in lesson 4.3.
Moving the character without losing the face
Plan a walk, a head turn and a spoken line using the same character reference and voice settings. Confirm that your current tools support the required inputs before paid generation. Review face, mouth, hands and product for the entire clip. Short shots are a useful practice constraint, not a universal five-second identity threshold. Record observed failures; finished clips require actual generation and review.
Six episodes, one look: thumbnails and a naming system
Prepare a repeated opening, closing and thumbnail layout for a fictional six-episode series. Save prompts, references, model versions and seeds where supported. Generate and compare the thumbnails as an optional execution step, recording drift and rejects. This is a written series method, not verified Spotify cover history; a seed alone does not guarantee repeatable output.
Faceless channels that survive without editing skills
Picking a faceless format you can ship every week
You score five faceless formats on three things: minutes to produce, how long a video stays useful, and how it earns. The five are narrated lists, one-idea explainers, ambient visualisers like the ones behind the Spotify artist, product-in-use loops using a fictional unbranded candle, and a running log of an AI-run business. You pick one, write ten titles for it, and put the first three in a calendar. The mistake this lesson prevents is choosing the format you like watching over the one you can produce when you are tired on a Tuesday. You leave with a chosen format and ten titles.
Script to voice: narration you can make in 15 minutes
You write a ninety-second script to a fixed shape, hook, three beats, close, at about 150 words a minute, and write it for the ear before you generate anything. Then you produce the voice in ElevenLabs with one voice and one settings block that never changes between videos. Pauses come from punctuation in the script, not from cutting silence in afterwards. The common mistake is rewriting for the ear after hearing the take, which doubles the time. You leave with a script template and a ninety-second voice file that sounds like last week's.
B-roll from the script: visuals without a camera
Split the script into ten beats and use one stable style prefix for stills. Animate only the beats where motion communicates something the still cannot; a three-clip version is a planning exercise. Log attempts, cost and production time instead of assuming a fourfold saving. Finish with a beat sheet, reviewed assets and a rough assembly you can watch end to end.
Captions, thumbnails and the publish checklist
You add auto captions in CapCut and fix the five words it always gets wrong, build a thumbnail template with three words at most and one object, and write the title and description from a formula. Then you publish one video and schedule two more in the same sitting. The Spotify artist did not lose its listeners to worse songs; it lost them when the posting stopped. So you set a schedule you can hold on a bad week, not a good one. You leave with one video live, two scheduled, and the checklist printed.
AI music and audio
A song brief that gives Suno something to hold
Choose v6, v6-wild or v6-mini for the job and your access. Write an original scene, musical brief and lyrics, set an attempt cap and compare complete candidates. Save the model, settings, actual credits and listening notes with the chosen song link. The dated v6 guide explains how to revisit older songs. A written brief is a useful outcome when generation is unavailable.
From a take to a release folder in one sitting
Match the edit to the audible defect, compare against the original and assemble a versioned release folder. Review the full audio, measure the exported master and prepare artwork in a suitable image tool, optionally Cirka, our founder's app. Check the exact output's permitted download, rights evidence and destination requirements. Save accurate metadata and mark unresolved checks pending before choosing a release date.
A release calendar you can sustain and measure
Build a fictional twelve-week calendar with six feasible targets, preparation deadlines and a buffer or reschedule decision. Reserve generation spending and final-download capacity separately, including possible corrections. Check distributor lead times, rights and asset readiness before submission. Keep planned and actual dates distinct; review production costs and audience observations without attributing growth to cadence.
An audio kit for your own videos: voice, bed, stings
Build a reusable voice, bed, sting and outro kit with an original instrumental brief and targeted listening checks. Set speech first, distinguish fader gain from measured loudness, then review the full export on headphones and a phone speaker. Trace each music file to its source and permitted use, check voice permissions separately and save processing settings with actual meter readings. A plan remains labelled as a plan until audio is made.
The weekly cadence: a one-hour production line
The one-hour line: five steps, five tools, timed
You run the production line behind the club's weekly videos with a timer running: ten minutes for the script, five for the voice in ElevenLabs, twenty for visuals in Nano Banana and Kling, fifteen for the cut in CapCut, ten to post. Sixty in total, and you write down where you overran. Most people overrun on visuals, because they generate before the script is locked. You leave with your own timed run, the one step you fix first, and the printed line from the library taped next to the screen.
Batching: four weeks of posts in one sitting
You batch by step, not by video: four scripts in a row, then four voices, then all the visuals, then four cuts. Each tool stays open once. Then you load the four finished videos into a sheet with dates and captions, and the n8n flow from the repo posts them on schedule. Making four videos one after another is the mistake; switching tools sixteen times is where the afternoon goes. You leave with four weeks scheduled, the sheet that drives them, and a flow you can hand to the automation track to extend.
Keeping the schedule: three numbers, one weekly review
You set up a fifteen-minute weekly review with three numbers: posts shipped against posts planned, one reach number, and one result number such as listeners, subscribers or orders. You fill it in for this week and put the slot in your calendar as a standing event. Two rules decide the rest: a format that sits below its baseline for four weeks gets cut, and one outlier gets a second video on the same idea. The stall on Spotify came from skipping this review, not from the music. You leave with the review filled and the slot booked.
References, light and delivery resolution
Source references and direct a controlled lighting study
Build a reference ledger and compare three lighting treatments while keeping subject and camera fixed. Includes the supplied mug reference, exact direction and a paper-based fallback.
Upscale, inspect and deliver an honest 4K master
Calculate a delivery crop, compare resampling with optional generative enhancement, and inspect edges, lettering and texture before export. Record what was enlarged and what was invented.
Finish the picture and build specialist ads
Repair an AI cut, build foley and compare video edits
Run a defect-led finishing pass, construct a three-layer sound mix, and use one controlled brief to compare available video editors. Gemini analysis and video transformation have separate jobs.
Build podcast, 2D, clay and game-style ad variants
Translate one factual offer into four complete visual treatments using awareness, proof and a single action. Includes scripts, shot timing, animation instructions and claim checks.
An original narrative short
Write a small story with a visible dramatic turn
Develop a wordless short from desire, obstacle and choice, then turn the script into a timed shot plan. The complete original story and continuity ledger are included.
Animate, edit and break down your narrative short
Build a thirty-second animatic, direct motion shot by shot, assemble sound and captions, and deliver a production breakdown. Still-image animation provides a complete local route.
Channel identity and retention decisions
Research a niche and build a coherent channel kit
Compare repeatable formats across three niches and produce names, branding, a channel description and a six-episode launch plan. Synthetic research examples are separated from real observations.
Diagnose retention and run a bounded creative test
Read a supplied synthetic retention table, calculate the largest losses and write a one-variable edit test. Follow the documented route to your own analytics when evidence is available.
A recoverable creative production pipeline
Wire brief, image, motion, audio and assembly contracts
Define every stage of a creative pipeline with copyable inputs, outputs and approval gates. Trace a complete synthetic job and use owned stills and recordings for the local media path.
Replay failures, control batches and hand off a production run
Use an executable offline state-machine exercise to handle duplicate jobs, unknown provider outcomes and rejected creative. Build a run ledger and a practical handoff with explicit production limits.
Keep learning in the club
Course outlines and these guides are free to read without an account. Full lesson bodies and member features require a paid membership. The preview uses sample data; external tools may charge separately.