Prompting and direction
Write clear briefs, define acceptance checks and improve AI outputs through focused feedback and comparison.
15 lessons in 5 modules
What you'll practise
- Define inputs, constraints and deliverables in a reusable prompt.
- Direct image, video and text work with references and specific corrections.
- Compare outputs against evidence and keep a record of what improved.
Before you start
Bring a task, source material and a text model you can access. You can write and review the brief in a document without buying a new subscription.
Course syllabus
Module and lesson summaries follow the club curriculum. Open any module to read its outline.
Briefing a model like a contractor
The five-part brief: role, job, inputs, output, limits
You take a vague ask, 'write ad copy for this mug', and rewrite it as a brief with five parts: who the model is, the job, the material it gets, the exact shape of the answer, and what it must not do. You run both versions in Claude and ChatGPT and put the outputs side by side. The usual mistake is skipping the output part, so you get an essay when you wanted a table with three columns. At the end you have a brief template in a text file and one filled-in brief for your own project, which every later lesson builds on.
Feeding the model real inputs, not descriptions
Build an inputs pack from authorised source material: the product page, price list and approved customer wording where available. Cirka's documented brand-scan workflow starts from a business URL; inspect the extracted facts before using them. The written exercise supplies a fictional unbranded mug page at an example.com placeholder address, with no customer endorsements. Keep facts, audience, voice and banned words in one file, leave absent information unstated, and reference that pack from the brief you wrote in lesson 1.
Output contracts: shape, length and a self-check
You define the answer before you ask for it: a table with named columns, a JSON object with fixed keys, or three paragraphs under 80 words. Then you add a self-check line that makes the model list the constraints and confirm each one before it answers. You run the same brief three times and score each run against the contract. Cirka's ad-copy generator returns a fixed shape, headline under 40 characters, primary text under 125, so the interface never breaks on a long answer. At the end you have a brief with an output contract and a scoring sheet that tells you when a prompt is done.
Reviewing the first draft like a client, not a fan
The second message matters more than the first. You take the draft from lesson 1 and write three review notes the way a client would: what is wrong, what to keep, and one change per note. 'Make it better' is the mistake; it gives the model nothing to act on, so you replace it with a note it can follow. You run two revision rounds in ChatGPT and log which notes moved the output and which did nothing. At the end you have a revision loop you can time, brief to usable draft in under 15 minutes, and a list of notes that work for your project.
Vibe directing: storyboards, shot lists, continuity
From one idea to a six-panel storyboard
You take a 20-second ad idea for a fictional unbranded mug and turn it into six panels with ChatGPT or Claude: what is in frame, where the camera sits, and what changes from one panel to the next. The model drafts, then you cut, because it writes moods when the video model only understands what a camera can see. You rewrite each panel until it is under 30 words and describes something filmable. At the end you have a storyboard document with six panels that becomes the shot list in the next lesson, and a rule you keep: if you cannot picture the frame, the model cannot render it.
Shot prompts that Kling and Veo can actually follow
Turn each panel into one filmable action with framing, camera, light and duration. Check the current tool's supported inputs and set an affordable attempt allowance. Prepare six shot prompts and a review sheet. Record actual accepted clips if you generate them; neither three-to-five retries nor a minimum number of keepers is promised.
Keeping the same product and person across shots
Write a continuity block for the product, place and person, then paste it unchanged into each shot prompt. Practise two ChatGPT Images 2.5 storyboard stills and check geometry, identity and layout against the original references. Motion uses a separate video workflow. Inspect adjacent frames and complete clips, log mismatches and version only supported changes; the block guides review without guaranteeing six matching shots.
Cutting six shots into a 20-second ad, then scoring it
Assemble available shots to the written timings, add factual spoken lines, captions and disclosure, then inspect product visibility, message clarity, continuity and the call to action. The score sheet identifies which shot needs revision. It is a practice method, not a measured improvement to Cirka's output; an exported ad is complete only after actual assembly and review.
Workflow prompts you run every week
The Monday prompt: a week of posts from one input
You build one prompt that stays fixed while the input changes: a release, product or approved customer message becomes five draft posts with distinct jobs. Supply voice examples and a list of phrases to avoid. Complete a written run, check every fact and time the editing as well as generation. Save the prompt and reviewed posts; the lesson promises neither a founder growth result nor a fixed runtime.
The research prompt: three competitors in 15 minutes
You write a prompt that takes three competitor pages, pasted or fetched in Gemini or ChatGPT where browsing is allowed, and returns one fixed table: offer, price, main hook, proof they show, and what they avoid saying. The columns never change, so week three compares cleanly with week one. The mistake is letting the model fill gaps; you require the words 'not stated' wherever a price or number is missing, because an invented price will end up in your own pitch. Cirka's brand scan does a version of this from a URL. At the end you have a table you regenerate weekly and a list of what changed.
The Friday prompt: numbers in, one decision out
Review comparable weekly inputs with sources and definitions. Ask for measured changes, possible explanations clearly labelled as hypotheses, options and missing evidence. Save one decision and its reason. This is a practice workflow, not a founder account history; four weeks of data does not guarantee an interpretable pattern.
Saving, versioning and sharing prompts
One file per prompt: a folder you will keep
You put every prompt you own in one folder, one plain text file each, with the same header at the top: name, version, date, what it is for, the inputs it needs and the model it was tested on. A GitHub repo or a Notion page works; the format matters more than the tool. This is how the club library is built, every item carries a version and a date, so an old prompt can still be trusted. You move five prompts from your chat history into the folder. At the end you have a prompt folder with a consistent header and nothing left in scrollback.
Versioning a prompt: what changed, and did it help
Version one prompt while preserving its prior text. Compare both versions against the same three inputs and rubric, record the actual scores and decide whether to adopt the change. Keep dates and results unfilled until the run happens. The worked changelog is illustrative; Library versions are not evidence of member failures or completed future tests.
Sharing a prompt so someone else gets your result
A prompt that works for you fails for someone else because they lack your inputs, your model settings and your review habit. You package it: the prompt, one example input, the output it produced, the model and date, and the three lines to change for a new project. You post the pack in the community and ask one member to run it cold. If their output matches yours it goes to the library; if not, the gap tells you what the pack was missing. At the end you have one shared prompt with a proof pair and a result from a run you did not touch.
Welcome and your first completed mission
Find your route, choose your setup and finish one useful thing
Navigate the classroom, pick a task and aspect ratio, complete a supplied first mission, and prepare a specific feedback request. No account signup, recording or separate worksheet is required.
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.