Hinge Auto-Date Bot, ROAS.to Launch, Agentic Loops — AI Daily Jun 07

397 messages · 67 active members

397
messages
67
active members
@mb29266, @Kieran, @GuruTime
top contributors

Overview

Saturday opened with @Vaibhavcste showcasing a Hinge automation bot that filters profiles by personal taste (slim, Slavic, low overweight probability), auto-likes, drafts replies via autopilot, and logs numbers/dates to a Google Sheet. The thread spiraled into half-joking investment theses about retatrutide upside plus tactical advice on like/pass ratios to protect Hinge ELO. Meanwhile @GuruTime soft-launched roas.to — a fully Meta-approved FB ads platform with VWO-style page optimization, voluum-style splitters, sub-minute spend tracking, auto-cloning on ROAS recovery, and a near full-stack MCP layer, offering 2-3 month free trials to a few serious testers. Kieran wrestled with Meta rejecting loan ads, with @offharoun recommending the Special Ad Category route and @LionOnX dropping a Russian placements bypass guide. On the coding side, builders dug into agentic loops — self-prompting systems that triage GitHub issues, plan, code, and review autonomously. The core failure mode: agents pick band-aid fixes (adjusting error messages) over root-cause solutions, so the fix is feeding them a vision.md describing the target user. Design chatter zeroed in on Claude's tells (gold/beige palette, serifs, dot patterns, eyebrow lines) and the rebuild workflow needed to escape the obviously-AI look. Infra threads covered isolated Hermes/OpenClaw setups on separate Mac minis with scoped tokens and Tailscale bridges, plus an Apify + n8n + Airtable stack for FB/TikTok ad library intelligence. The Codex vs Claude debate continued — Hermes only supports Codex OAuth — and @ecomSon flagged TurboVec compressing 31GB of vector memory down to 4GB.

Topics

@Vaibhavcste built a Hinge bot that filters profiles by personal taste, auto-likes, drafts/sends messages, and logs phone numbers and dates to a Google Sheet. Autopilot learns from prior conversations; he maintains an 80% like cap and ~30% match rate. Community suggested wiring it into Google Flights and Airbnb for full geo-arbitrage dating ops.

@GuruTime is launching roas.to after 8 months internal use and a from-scratch SaaS rebuild. Full Meta tech provider approval, custom domains per user, sub-minute spend updates at $500/day, VWO-style DOM manipulation, voluum-style splitters, auto-cloning on ROAS breakback, and a nearly full-stack MCP. Offering 2-3 month free trials to 3-5 serious users.

Kieran's loan ads keep getting flagged on one account despite running fine elsewhere. @offharoun recommends declaring as Special Ad Category to avoid profile-level bans with minimal performance hit. @LionOnX shared a cpa.rip placements-moderation bypass tactic, and @swh800 noted creative-level targeting now matters more than ad-set targeting anyway.

Builders discussed loops that pull GitHub issues, triage, propose solutions, design tests, and write code autonomously. Core challenge: agents pick band-aid fixes (like adjusting error messages) over root-cause solutions. Fix: vision.md docs describing the target user steer optimization toward real customer experience. Isolated Hermes setups on separate VMs with scoped tokens and Tailscale bridges recommended.

Members called out Claude's recurring design fingerprints — gold/beige colors, serif fonts, dot patterns, and eyebrow lines above headlines. Recommended workflow: first-pass with Claude + Stitch + 21st.dev references, then rebuild against a hand-picked real lander once the AI-look becomes obvious.

Key Takeaways

  • Hinge automation works: 30% match rate at 80% like cap, with autopilot drafting replies — keep like/pass ratio sane to protect ELO
  • ROAS.to is offering 2-3 months free to 3-5 testers with full Meta API approval, sub-minute spend tracking, and an MCP layer for AI-driven ad ops
  • For Meta financial/political ads, declare as Special Ad Category to avoid profile bans; performance hit is often negligible
  • Agentic loops fail when they optimize for graceful failure instead of root-cause fixes — feed agents a vision.md anchoring decisions to a real user
  • Separate RAG (documents in Elasticsearch) from agent memory (facts in Postgres); TurboVec's 87% vector compression could unlock laptop-scale RAG

Hot Threads

@Vaibhavcstestarted

Hinge auto-date bot with taste analyzer and autopilot messaging

35 replies8 participants
@GuruTimestarted

ROAS.to launch — Meta-approved ads platform with MCP, looking for beta testers

22 replies7 participants
Kieranstarted

Meta keeps rejecting loan ads on one account — Special Ad Category workaround debate

18 replies5 participants

Linked Items