Retreat Recap Video
The program plan behind the recap of the Melbourne Infinite Leverage retreat, written on the 5D framework after we made it. Phone photos, a handful of clips and ten testimonials filmed on the Kooyong balcony became a 2:45 video that walks the two days block by block, with each block proven by a founder in their own words. Claude did the transcribing, cutting, timeline and recording in Claude Code; the host made four decisions.
Show what they learned, then let them prove it.
The first cut was a montage with testimonials on top, and the host sent it back: light on content, and day two barely there. A recap is a promise to the next cohort about what they will learn. So the video follows the curriculum, block by block, and every block is followed by a founder saying, in their own words, what that block did for them. See the result on the Melbourne recap page.
The Melbourne recap, 2:45
Define the problem
Two build days end with a camera roll: photos at odd angles, short clips of the room, and a row of testimonials filmed between sessions, each a minute or more of someone gathering their thoughts with an interviewer off camera. Nobody on the team is a video editor, and the recap is worth most in the week after the event, while the cohort is still talking about it and the next cohort is deciding.
Who needs what
- The cohort needs the two days back in one place, with their own work and their own words in it
- The next cohort needs to see what is taught and what people walk out with, not just that it was fun
- The host partner (here, EO Melbourne) needs something they are proud to share with their members
The goal
- A two to three minute video that covers every block of both days with equal weight
- Every block backed by at least one founder in their own words, captioned word for word
- Every testimonial trimmed, named (role and company) and usable on its own
- Published in the portal for the cohort, and on the public page once the client agrees
Discover the material
Most of what the video needs already exists before anyone opens a timeline. The work is finding it and joining it by person.
| Source | What it gives you | Watch out for |
|---|---|---|
| The camera roll | Photos and short clips of the room, the venue and the people. | Sideways and upside down photos (the rotation lives in metadata), HEIC files, portrait clips. |
| Balcony testimonials | Ten founders, thirty seconds to two minutes each, the best raw material there is. | Interviewer questions at the start and in the middle, wind on the mic, false starts. |
| The curriculum | The eight blocks of the two days, from the teaching deck: the spine of the video. | The deck has fifty slides. The video needs one line per block. |
| Kickoff survey | Company and role for every attendee, and a confidence score on arrival. | Self-reported spelling. Use theirs, not yours. |
| Final feedback | NPS, confidence and capability ratings, and who agreed to be quoted. | Different wording before and after. Never present two questions as one. |
| Day one reflections | The live address of what each person built. | Admin links behind a login. Only public home pages go in a video. |
| The cohort chat | Names as people write them, and the thank-yous written the evening it ended. | Phone numbers and emails. Read it for names and quotes only. |
The transcript is the index
- Every clip is transcribed locally with a time for every word, before anyone watches anything
- Picking a soundbite becomes reading, and cutting becomes choosing two words
- A short vocabulary list (the venue, the product names, the host) fixes most mishearings up front
Check every name against two sources
- File names, chat names and the kickoff survey disagreed more than once
- One company came from a chat display name, one from an email domain and a live site
- Two people were unnamed until the host matched faces to names from a photo
Design the workflow
The video has one repeating beat: a card that says what the block taught, then the founder who proves it. The card shows the block's title, one line of what was taught, and a strip of all eight blocks with the current one lit, so a viewer always knows where they are in the two days.
| Segment | On screen | Sound |
|---|---|---|
| Open | The two logos and the event name, then the venue | The beat builds and drops on the venue |
| Day card | “Day one” or “Day two”, and what that day was for | Full beat |
| Block card | Title, what we taught, the strip of eight with this block lit, over a photo or clip of the room | Full beat, landing on the bar |
| Soundbite | The founder in a portrait frame; the block, their name, role and company; every word lighting up as it is said | The beat drops to a quiet bed about 18 dB under the voice, and builds back into the next card |
| Close | The group photo and both logos | Full beat, then out |
The soundbite map
Read every transcript once, then give each block the line that proves it. In Melbourne the strongest lines landed on day two, which is what fixed the first cut: harden, nurture and agents each had a founder saying exactly what that block did.
| Block | Day | The line that proves it |
|---|---|---|
| Learn and plan | Day one | Fiona: “embarrassingly awful with AI”, then Peter: “weaponized with some tools” |
| Build small | Day one | Ivan: the setup “gives me the confidence”, then Ashton: “follow the bouncing ball” |
| Full build | Day one | Four sites built that day, then Fiona: “about 80% there with my CRM” |
| Data and dashboards | Day one | Lana: “I can have all of my sales transcripts be automatically read, be ranked and populated into a CRM” |
| Harden the code | Day two | Ashton: “uncovered some pretty serious security flaws” |
| Nurture campaigns | Day two | Fiona: the system “building automatic emails” to new referrers |
| Agents in the cloud | Day two | Ashton: “an agent in the cloud” for his daily dashboard, then Lana on agents |
| Reflect and repeat | Every day | Nadi: “a plan to roll out” more apps, Sue, and Andrew to close |
Cutting rules
Every rule below came from a cut that went wrong first. They are cheap to follow because the transcript already knows where every word starts.
Where to cut
- Start a tenth of a second before the first word, end in the pause after the last one
- Never fade over a word. A short fade inside the silence is fine; a fade on the last syllable reads as a cut-off
- If the next word starts straight after (“following them up, like”), end at the phrase before
- Questions from the interviewer come out at the start and end; inside a clip they are captioned as a question
What the captions may change
- Spelling and names: the venue, the product, the person
- A word the transcriber clearly misheard (“floors” for “flaws”)
- A swear word in the caption, never in the audio
- Nothing else. If a line needs smoothing to work, pick a different line
The flow
Four human decisions, the rest is Claude and two scripts. The host approves the structure, confirms the names, signs off the soundbite per block, and says when it goes public.
The flow in detail
The rail above is the happy path. The real run has three checks that send work back instead of shipping a weak cut: the structure (Melbourne's first cut failed here), the last word of every soundbite, and the level of the voices over the music. And one gate before anything is public.
How each step works
Copy the camera roll, never touch the originals. Turn HEIC into JPEG, apply the rotation hidden in each photo's metadata (four of Melbourne's eleven came out sideways or upside down otherwise), shrink everything to web size, and make one contact sheet of every photo and one frame from every clip, so the whole camera roll can be reviewed in a single image.
Whisper runs locally on the laptop and returns a time for every word. A short vocabulary prompt (the venue, the products, the host's name) fixes most mishearings before they happen. The transcripts are what the rest of the plan reads; nobody scrubs through video looking for a quote.
Claude matches each clip to a person from file names, the cohort chat and the kickoff survey, and lists the gaps and the conflicts. The host fills them: in Melbourne, two testimonials had no name until the host sent a photo of each with the name underneath. Company and role come from the person's own kickoff answer.
Read every transcript against the eight blocks and pick the line that proves each one: five to ten seconds, a complete thought, ideally with one word worth lighting up in the caption. Day two needs as many lines as day one. Lines about the venue, the host and the overall experience go to the open and the close.
The host sees the block cards and the soundbite per block as stills before the full render. This is where Melbourne's first cut was caught: a montage plus testimonials, light on what was taught, with day two reduced to a few seconds. Fixing it here costs minutes; fixing it after the render costs another render.
One file per soundbite, cut on word boundaries from the transcript, with loudness evened out across speakers. Then the proof: transcribe the last two seconds of every cut and check the last word is whole. In Melbourne this caught five testimonials where a fade had swallowed the final syllable, and one where the next word ran straight on, so the line had to end a phrase earlier.
The video is a single web page where every frame is a pure function of time: block cards, the founder frames, the captions lighting word by word, and music synthesised in the browser. A headless browser renders it thirty frames a second, waits for each clip to land on the exact frame, and pipes the frames to ffmpeg; the soundbites are mixed in at their cue times. Stills of every card and a loudness reading per bar are checked before anyone watches the whole thing.
The full recap goes into the cohort portal first, behind the access code. The public recap page gets the video, the eight blocks, the live projects people shipped and the ratings, only once the client has agreed to a public recap. The host gives the go for each, and checks the live page, not a preview.
The seven elements
Every workflow we document has the same anatomy: seven elements, each assigned to a human, a machine, or both. This is the Centaur Map from our workflow design method.
The host drops the camera roll into a folder in the days after the retreat and asks for a recap. Nothing runs on a schedule.
Photos, clips and testimonials from the camera roll; the curriculum from the teaching deck; names, companies and roles from the kickoff survey; ratings from the final feedback; live sites from the reflections; venue and partner imagery.
Claude proposes which line proves which block, from the transcripts. The host decides the structure, confirms every name, and signs off the soundbites before anything is rendered at full length.
Each soundbite goes to the block it proves. Interviewer questions are cut or captioned as questions. Login-only projects and anything private are routed out before they reach a frame.
One recap video, block by block, with a card per block and a founder after it. Ten trimmed testimonials with name cards and captions. A web copy and a poster frame for the public page.
The full recap is published in the cohort portal behind its access code. Once the client agrees, the video, the blocks, the live projects and the ratings go on the public recap page.
The scripts prove every cut ends on its last word and the voices sit clear of the music. The host watches it end to end. The real measure is whether the cohort shares it and the next cohort asks about the blocks.
Determine the ROI
What it costs
- A few working sessions across the days after the event, most of it waiting on renders
- Four decisions from the host: the structure, the names, the soundbites per block, and the go to publish
- No editing software, no stock music licence, no editor: the music is synthesised in the page
- Tokens, and a laptop that can record a browser for a few minutes
What it returns
- A 2:45 recap that sells the curriculum, not just the atmosphere
- Ten trimmed, captioned testimonials with name cards, usable one at a time on any channel
- A public recap page with the video, the blocks, thirteen live projects and the ratings
- A repeatable run: the next cohort is new footage through the same eight steps
Deploy, and the definition of done
The video is a web page. The timeline, captions, cards and music are HTML and JavaScript, so the same page plays live in the portal with sound and records frame by frame to an MP4 for everywhere else.
| Piece | Role in the system |
|---|---|
| Footage folder | The camera roll, as dropped by the host. Never edited in place; every output is a new file. |
| Local transcriber | Whisper, run on the laptop: a transcript with a time for every word, for every clip. |
| Bite cutter | A short script over ffmpeg: one soundbite per file, cut on word boundaries, loudness evened out. |
| Timeline page | One HTML page where every frame is a pure function of time: block cards, word-timed captions, the music. |
| Recorder | A headless browser that renders each frame, waits for every video to land on it, and pipes the frames to ffmpeg with the voices mixed in. |
| The retreat site | The full recap in the portal behind the cohort's access code; the public recap page once the client agrees. |
| Claude Code | Reads the transcripts, maps soundbites to blocks, writes the page and the scripts, checks every cut. |
Guardrails, always on
- Nothing goes public without the client's recorded agreement. Until then the recap lives in the portal
- Captions are what they said. Spelling and names are fixed; words are never added or smoothed
- Only public pages are shown or linked. A login screen, or an internal tool's address, stays out
- Every name, role and company is checked against two sources before it is burned into a frame
- House style survives the render: no em dashes, and a brand name never set in all caps by a button style
Definition of done
- Every block of both days has a card and at least one founder backing it, and day two gets as much time as day one
- Every soundbite ends on its last word, proven by transcribing the last two seconds of each cut
- Voices sit well clear of the music under them, checked bar by bar
- Every testimonial carries a name card with role and company, and captions throughout
- The video plays in the portal and on the public recap page, and the host has watched it end to end