Voice Note to Published Review: The Input, the Pipeline, the Output

What the loop produces

A 60-90 second voice note from the operator produces a complete published review on agenticbotsitter.com in about 20 minutes. The output is a markdown post with first-hand reaction text, an AI-generated featured illustration, an Amazon affiliate link, and a grade chip (A/B/C/D/F). The post renders live in /reviews/<slug>/ and ships with the standard PNG + WebP image pair.

What arrives in the voice note

The operator speaks in their natural cadence. Useful signals to include in the audio:

• The product — what it is, briefly
• The use case — what you used it for, how it performed
• The comparison — vs what you used before, why this is different
• The verdict — would you recommend it; if so, who for
• Any caveats — where it falls short, what to watch out for

These don’t have to be in order. The transcription step uses them as anchors.

What arrives together with the voice note

• The amzn.to affiliate URL (one canonical product link)
• The category for the review (electronics / kitchen / etc.)
• The grade if the operator is sure (A/B/C/D/F); otherwise leave blank and the skill defaults to B

These three live in the same Slack DM / chat message as the voice note transcription. Without them the draft skill does the work with defaults.

What the operator does during the loop

Almost nothing. The typical interaction:

1. Send the voice note and the URL.
2. The transcript comes back from STT. Verify the transcript makes sense (rare errors with brand names). If not, re-send a clarification.
3. The draft comes back as a markdown file with the schema, body, and image prompt. Verify the body reads like the operator’s voice — the skill won’t fix a “I never sound like this” complaint after the fact.
4. The image lands in /srv/abs-site/public/images/. Verify the illustration matches the product (occasionally the AI can pull the wrong object).
5. Build, regression, rsync, CF purge, commit, push — all automated.
6. The post is live. Confirm by opening the URL in a private browser.

That’s the operator’s whole job: send voice note + URL, verify three artifacts, ship. None of the build / tag / pipeline work happens in front of the operator.

What the operator does NOT do

• Edit the body of the draft — the skill’s drafting is final unless there’s a factual error.
• Pick the image style — it’s locked in the botsitter-review skill’s image-style.md.
• Decide on tags or categories — those are inferred.
• Manually rsync or git push — all automated.

The two failure modes

• STT mishears a brand name (e.g. “ZTS Mini MBT” becomes “zits mini MVT”). The operator’s verification step catches this. Re-record the voice note with spelling.
• AI illustration picks the wrong object (a generic box instead of the specific router). Re-issue the image prompt in the bot’s image gen channel; the post picks up the new image on the next deploy.

These are the only two manual interventions in 95% of the reviews that ship. The other 5% are real bugs (skill refactor shipping with a schema mismatch, etc.) and those are caught by the regression suite.

What happens after the post goes live

The published review enters the daily-briefing pipeline that 7-day rolls user traffic. Review-front stats start showing in the post-deploy dashboard within 24-48 hours: page views, unique visitors, affiliate clicks, newsletter signups (for readers who landed on the page and signed up).

What the operator should keep an eye on, sparingly:

• Affiliate click-through at 48 hours — if the operator’s recent review never gets a click, the affiliate tagging may have missed (verify with grep agenticbotsit-20 dist/reviews/<slug>/index.html).
• First-day bounce rate — if a recent review’s bounce rate is dramatically higher than the median, the voice note likely landed but the body didn’t match. Pull the post down and re-record.

Most reviews don’t need any post-ship attention. Trust the system unless the metrics break the band.

If the bot doesn’t reply

Sometimes the assistant side goes silent — STT service hiccup, image gen backend hiccup, an unrelated Hermes updater reboot during the deploy window. The operator’s signal that the bot didn’t reply is the lack of a draft markdown in the queue after 30 minutes.

The recovery:

1. Re-send the voice note (same file, same Slack DM). The skill is idempotent on the input side.
2. If the second send also returns nothing, escalate to manual: open the post yourself with the body already drafted, paste the image once available, ship via the standard deploy chain (G1). The infrastructure-level failure does not block the post from going live.

This is the rare exception, not the rule. The bot replies in 95% of cases within 90 seconds; if you see re-sends becoming routine, the skill refactor needs to land — that’s the bar for a P-class concern rather than an operator concern.

What not to do

• Don’t send a research-only voice note (“this product has 4.5 stars from 1234 reviews”) — the body won’t read like a review. See G4 for the first-hand-vs-research-only distinction.
• Don’t send multiple amzn.to URLs in one voice note — pick the canonical product.
• Don’t invent a grade; if you don’t know, leave it blank. The skill defaults to B but lets the operator adjust.
• Don’t skip the verification step on the transcript — that’s where 80% of the post-publish edits come from.

Verification checklist

#	Command	What it proves
1	The transcript matches the voice note	Body text reflects what you said
2	The draft’s body matches your voice	The post will read like a review
3	The image matches the product	Featured illustration is recognizable
4	npm run build exit 0	Frontmatter parses
5	curl -I https://agenticbotsitter.com/reviews/<slug>/	Live URL returns 200
6	Three-SHA match	Deploy was local + origin + github

Sources

• ABS companion: Botsitter-Review Affiliate Link Convention
• ABS companion: Publishing a Product Review End-to-End on ABS
• ABS companion: First-Hand Versus Research-Only Reviews

Last verified: 2026-07-14 against the live botsitter-review skill and 31 published reviews.