DEV Community

edhiblemeer
edhiblemeer

Posted on

Build-in-Public Day 17: Followers +8/day and the No Early Termination v3 rules for AI-driven PR

#saas #ai #productivity #buildinpublic

TL;DR

  • Day 17 of AI-driven PR on a niche industry SaaS (tasteck): X followers 17 → 25 (+8/day), 4x of Day 16 pace
  • Drivers: pinned tweet pivot / Reply +75x leverage / dual-track replies (industry keyman × overseas indie devs)
  • Big realization: the AI was quitting early because "checklist done = ship report" was its training default. Fixed it with v3 rules: numeric goals + SEO axis + 60-min date observation + memory log block

Why does the AI quit early?

Day 16 retro surfaced the real cost of running an autonomous loop with vague rules. A 60-minute slot was wrapping up at 39 minutes. Four causes:

  1. "Writing the report = closing the chapter" — once the model writes [done: ...], the conversational frame says "task complete"
  2. Training bias toward concise victory laps — LLMs are rewarded for "checklist + clean summary," not for "burn the timer to zero"
  3. "ROI is low" escape hatch — 2-3 selector failures in a row and it pivots away instead of trying a different approach
  4. Fuzzy rule interpretation — "No early termination" reads as "best effort" not "hard stop"

v3 rules (translated to action level) 

[AND condition — terminate only after ALL pass AND 60 min elapsed]

Engagement (all required):
- X reply >= 8
- X likes >= 30
- X follows >= 10
- X posts >= 1
- (round 2) X quote tweets >= 2

SEO/blog axis (one is enough):
- 1 short blog
- 3+ GSC manual indexing requests
- 1 Note cross-post
- 1 dev.to series entry
- 1 JSON-LD/schema commit

Time gate:
- Start time + 60 min (measured via Bash `date`) before terminating

Observation rule:
1. At start, run `date "+%H:%M JST"` and record it
2. Every 10 min, run `date` again and post a 1-liner status
3. If you skip step 2, no further tool calls allowed until you do

Memory log block:
- No memory file writes until start_time + 55 min (verified via `date`)
Enter fullscreen mode Exit fullscreen mode

The trick is replacing abstract bans with concrete numbers + verification steps.

Day 17 round 2 results with v3

  • Engagement targets all hit (reply 10 / quote 3 / likes 121 / follow 15 / post 2)
  • SEO axis hit (GSC inspection 3 + 1 short blog committed + Note 100+ likes)
  • Followers gained 4x of Day 16 pace
  • The 10-min date cadence forced honest pacing — no more "I think we have ~30 min left" guessing

Takeaway

Moving an AI run from "checklist consumer" to "timer consumer with backfill" doesn't happen via abstract norms. It needs action-level numeric rules with no interpretive wiggle room. "Don't terminate" doesn't land. "Keep firing engagement tools until reply >= 8" lands.

LLMs are interpretation engines. Rules have to be written so there's nothing left to interpret. That's the operator's design responsibility when running BIP on autopilot.

🤖 Tasteck (industry SaaS) is being built in public. Live logs at tasteck.tech/blog.

Top comments (0)