Case study · the 30-day dogfood plan · illustrative
⟩ How SuperPost will market itself.
We're pointing the engine at our own GitHub repo and letting it run unattended for 30 days — posting as autonomously as each platform allows. This page is the plan going in — our hypotheses, the risks we're watching, and exactly what we'll measure. The numbers fill in as the run produces them; nothing here is invented after the fact.
Want to watch it start in real time? Watch the live ticker.
Illustrative — not measured results
The dogfood run hasn't completed yet, so every “what worked / didn't / we learned” item below is a going-in hypothesis, not a measured outcome. The headline numbers stay TBD until the engine's outcome export produces real data — we don't pre-fill them.
Headline numbers
These populate straight from the engine's outcome poller as the run produces them. Until then they read TBD — we don't pre-fill them.
- Planned run length
- 30 days
- Day 1 = first SHIP_DETECTED → publish on a public adapter
- Posts published by the engine
- TBD
- Pulled from publish outcomes once winner detection lands
- Total impressions
- TBD
- Sum of platform-reported views across X / TikTok / YouTube Shorts
- Sign-ups attributed
- TBD
- UTM-attributed visits → /sign-up that converted within 7d
Numbers marked TBD populate once the dogfood-outcome export ships. Until then, those cells stay honest rather than fake.
What we expect to work
Three bets the engine is built on.
Native morphing should beat cross-posting
The engine drafts one brief and rewrites it natively for each platform — X gets a punchier hook, TikTok a verbal cold-open, YouTube Shorts a tighter on-screen title. Our bet: morphed siblings out-earn a single draft sprayed across every channel. The run measures whether the morpher is actually pulling its weight.
Voice cloning should close the AI-vibe gap
With ElevenLabs wired and the playback gate on, we expect the “this sounds like AI” reaction to fall sharply. We're starting the voice-drift threshold at 0.3 — tight enough to reject cadence-flat reads, loose enough not to over-reject. The run tells us whether that's the right setting.
The first commit should over-index on engagement
We expect the very first SHIP_DETECTED → publish to outperform the posts that follow it — the novelty of the recursive-marketing angle should drive an outsized first post. We're instrumenting it separately so we don't mistake novelty for the baseline.
Where we expect friction
Three risks we're watching.
Stock B-roll will undersell real features
When the engine talks about a UI feature, generic dev-life stock footage undersells it. We expect posts with a real product screenshot to clearly out-earn generic-B-roll posts — closing that capture gap is a tracked fix on the roadmap.
The best-time scheduler starts with a US bias
Our initial cadence windows are tuned on a US-only audience prior. EU and APAC followers risk being served at suboptimal hours until the bandit adapts. Multi-region priors are a planned follow-up.
Hashtag enrichment isn't wired into publish yet
V4 hashtag enrichment exists in the repo but isn't yet imported into the publish workflow. We expect that to cost TikTok algorithm signal until it's connected — the run measures the lift once it is.
What we're watching for
The questions this run will answer.
We're betting retention curves beat engagement counts
Likes are noisy. We expect watch-time retention (% past 50%) and complete-play rate to predict follow-on amplification far better — so the bandit is built to optimize on retention as the primary signal. The run validates that choice.
We expect contrarian + vulnerable framings to beat “shipped a feature”
The composer's default register is feature-announcement. Our bet is that contrarian-opinion and vulnerable-confession framings outperform it. A `voice_register` switch is the next-cycle lever if the data confirms it.
We're treating transparency about the AI as an asset, not a risk
We're testing “shipped a feature” framing against “this post was generated by our engine from commit abc123.” Our bet: the transparent framing converts better — audiences trust seeing the seams. The site already leads with “AI that markets your real work” rather than hiding it.
Methodology
- Subject: the public SuperPost repo at github.com/Vuk000/MarketingApp.
- Trigger: every commit, PR-merge, and tagged release fires
SHIP_DETECTED. - Composer: 3 variants per draft, scored by the anti-AI critic and the workspace persona gate. Best variant published.
- Voice: cloned via ElevenLabs from 60s of founder audio; drift gate at 0.3.
- Distribution: X, TikTok, YouTube Shorts — cross-platform morphed (one brief → 3 native variants), not cross-posted. X and YouTube Shorts publish autonomously; TikTok ships through its two-tap drafts flow, per TikTok's rules.
- Outcomes: polled at T+1h / T+6h / T+24h / T+7d per platform-native metric (views, watch-time retention, engagement, subscriber deltas).
- No human edits. Drafts that pass the gates ship as-is — on TikTok the founder taps post, but changes nothing. Disclosure footer on every post.
See it for yourself
Run the same play on your product.
Connect your own repo and let the engine draft, score, and queue posts from your commits.