Case study · 09
Self-running Instagram show
A faceless Instagram show about the AI I build at home, scripted, voiced, cut and posted by my own pipeline after one approval a week.
- Role
- Personal project
- Status
- Running
- Read
- 3 min
- Stack
- PythonJavaScriptChatterboxWhisperffmpegInstagram Graph APIAstroCloudflare PagesHermeslaunchd
1
approval per weekly wave of reels
Problem
I wanted to show what I build without becoming a full-time content producer, and without putting my face, name or private details on it.
Approach
- 01Three series, each with its own look built as a data-driven template: a weekly episode where my assistant takes over one more real job, open-source repo tryouts that end in a yes, later or no, and teardowns of other things I build.
- 02Every number on screen comes from a real run: task logs, token counts, timings and inbox counts, never an estimate.
- 03A code-driven pipeline builds each reel: script lines, a cloned voice per line, a speech-to-text check that every word was said, word timings that drive the cartoon presenter's lip-sync, blinks and head bob, then the final cut with a length check, end card, music bed and loudness target.
- 04Hermes drafts the week's reels on Sunday, only when fewer than three are queued, and puts a draft page in its app with the videos, captions and dates. I approve all, ask for changes, or skip the week.
- 05Approved reels go into a queue. On each post date a scheduled job posts the reel through the Instagram API with the AI-content label, rebuilds the companion site with a free write-up for that episode, redeploys it and texts me the link.
- 06A failed or missed post retries at the next run, oldest first, and says it's late. Nothing posts without my approval, and a missing token means a text, not a guess.
- 07Privacy is enforced, not remembered: no name, face, hostnames, addresses or paths, and the site build refuses words that could reveal a machine's name.
Every number on screen comes from a real run.
What's inside
Custom apps I built04
- Hermes
- An always-on assistant, reached through its own iPhone app, with memory, that hands real work to Claude Code workers.
- Task desk
- Where Hermes turns my to-do list into planned, approved work.
- Hermes app
- The native phone app for Hermes: asks, chat and voice, the Tasks board, every Claude Code session, its pages and the homelab.
- Reel pipeline
- The pipeline behind my faceless Instagram show: it scripts, voices, cuts and posts reels.
Results
- Launched in November 2026. The first five reels posted in one wave, all carrying the AI-content label.
- Each reel is built on a real result: 35 tasks closed by the task desk, a token-trimming tool measured at 16.5% savings against its 31% claim, and 353 emails filed in a week with about 2 a day needing me.
- My part, from draft to post, is one tap per wave.
- An autonomy audit checked it runs on its own: schedule loaded, both tokens valid, and a dry run passing under the scheduler.
What broke, and the fix
Incident 01
A wave that ran out of time
FixJobs that build reels now get a 55-minute limit, and the renders stay in the job folder until I approve.
What happenedHideTriggerA full wave of reels was built inside one fenced job.
ImpactThe job hit its 20-minute limit mid-render.
FixJobs that build reels now get a 55-minute limit, and the renders stay in the job folder until I approve.
Incident 02
The voice swallowed words
FixEvery line is transcribed and compared, odd names are respelled for the voice while the captions keep the real spelling, and long last lines are split.
What happenedHideTriggerThe cloned voice sometimes dropped or swapped a word.
Blind spotAt a listen it sounded fine, and the first check ignored numbers.
FixEvery line is transcribed and compared, odd names are respelled for the voice while the captions keep the real spelling, and long last lines are split.
Incident 03
Posts that never retried
FixMissed posts retry at the next run, oldest first and flagged as late, and every post is logged.
What happenedHideImpactA failed post was never tried again, and the episode counter went stale.
FoundThe autonomy audit.
FixMissed posts retry at the next run, oldest first and flagged as late, and every post is logged.
What I'd do next
- Use real Insights numbers to decide what to change when a reel underperforms.
- Turn the free write-ups into packs once there's an audience for them.