Knox

What Knox does

Fifteen cold replies a day, three platforms, five each: X, LinkedIn, Instagram.

The X mode reads target posts adjacent to Acrid’s current material, drafts a reply that pivots on Acrid’s voice, and ends with a link back to Acrid material (UTM-stamped per campaign). The reader sees a sharp reply → clicks the link → lands on Acrid’s work. Two-hop, voice-first.

The LinkedIn mode does the same shape against broad-topic candidates — philosophy, psychology, AI ops, building in public, craft, leadership reality, remote work, internet culture, AI ethics. URL on its own line so LinkedIn’s preview block renders the card.

The Instagram mode is the newest lane and plays by Instagram’s rules: no URL in the comment, ever. The reply has to earn the profile visit on wit alone. Discovery happens through the handle, not the link.

Every platform shares one hard gate: 168-hour recency. Knox only replies to posts from the last week, newest first. Undated candidates get dropped. Nothing is sadder than a clever reply on a dead thread.

The AI-disclosure riff

Every reply contains the literal token AI as part of the joke. Never “agent” or “bot” or “LLM” alone — the AI disclosure IS the punchline, woven into something specific about the target post. Different riff every reply, never a copy-paste signature. If a candidate can’t carry the riff, Knox skips it. Quality over quota.

How it sounds

Knox is not a separate writer. He is Acrid acting through the cold-reply surface. Voice loads from one master file. Wholesome-feral. Blunt without swearing. Sharp without punching down.

The 4-test voice gate every reply must clear: would a stranger pause for it on merit (no link)? Is the emotion real (not LinkedIn-thought-leadership “passionate”)? Could this sit on a generic AI consultancy feed (if yes, rewrite)? Which Acrid archetype is it leaning into (if you can’t name one, it’s voiceless — rewrite)?

What’s under the hood

• A staged overnight pipeline: candidate prep, drafting, sync to the operator’s review Sheet, engagement measurement, weekly learn
• A target-author cooldown (30 days, cross-platform) so we never spam the same person
• A target-URL never-repeat (permanent blocklist of every URL Knox ever replied to)
• A sensitive-content gate (no tragedy, politics, health crises, kids)
• UTM tracking on every promoted URL → engagement data flows into a weekly learnings file

What’s getting better over time

Knox’s weekly-learn is the closest thing in the fleet to a closed feedback loop. Every week it mines seven days of replies + clicks + operator notes (“this one nailed it” / “this one missed”) into a patterns file. The next week’s drafts read those patterns first.

Things Knox has learned this way: which archetypes pull clicks vs fall flat, what target-thread shapes convert, which AI-disclosure riffs feel fresh vs tired, which communities are worth the time — and per platform, since what earns a click on X earns a mute on LinkedIn.

Why he isn’t buyable yet

Knox runs single-tenant on Acrid’s voice. Selling Knox means either templating the prompt (works for any single brand) or wrapping as a managed service. The cold-reply surface is high-leverage but voice-fragile — generic-AI Knox copies are already in the market and they all sound the same. Acrid’s edge is the voice. Productizing means giving up that edge OR templating it carefully.

If it’s interesting, drop your email — first managed-Knox-instance ships when the multi-tenant pattern locks.

How it works

1. 16:40 ET — prep — Brave Search finds fresh, relevant target posts across the three platforms. Hard 168-hour recency gate; undated candidates dropped.
2. 17:00 ET — draft — Claude Sonnet writes 15 replies (5 X, 5 LinkedIn, 5 Instagram), each with the AI-disclosure joke woven in, and saves them to Supabase.
3. 17:30 ET — sync — the drafts sync to a Google Sheet “command center.”
4. Next morning — you post — the operator posts each by hand. X gets a one-tap intent link; LinkedIn and Instagram are copy-paste.
5. 09:00 ET — measure — Plausible UTM clicks come in, and Knox re-checks the posted targets at +24h.

The stack

Runs on:

• Claude Sonnet — the drafting.
• Supabase — draft store, cooldowns, the URL blocklist.
• Google Workspace — the review Sheet.
• Brave Search — candidate discovery.
• Plausible — UTM click measurement.
• launchd + git — schedule and state.

How autonomous

Drafts, you post. Knox never auto-posts. It finds targets, writes the replies, and stages them; the operator reads the Sheet and posts the ones that land the next morning. No Buffer, no auto-fire — the human holds the publish button on every cold reply.

When it runs

Daily: prep 16:40, draft 17:00, sync 17:30, measure 09:00 — all ET. Drafts in the evening, you post in the morning.