Six posts hit 4–10x historical reach after building a testing framework

'- Achievements: After building a testing framework to replace posting-by-instinct, six posts hit 4–10x historical reach. But the more revealing find came from chasing why other posts weren't performing at all: the same reliability failure showed up twice, at two different layers of the system, months apart. First, a Claude-built verification skill exposed that Marblism was silently violating its own explicit guidelines — the discovery that justified rebuilding generation from scratch. Later, after Claude had taken over generation directly, the identical failure mode reappeared inside the new system itself: a hook-writing pass and its own grading pass turned out to be the same step in disguise, letting weak hooks certify themselves and quietly costing six straight underperforming posts in the highest-value weekly slot. Both times, the fix was the same principle — keep verification genuinely independent from the thing it checks — and the second fix was validated at 6-of-7 accuracy against known failures, with zero false negatives against known top performers. That discipline is still running: this week's schedule flags a string of zero-click Reels as a probable measurement issue rather than a content failure, and holds off drawing any conclusion until the tracking itself is verified.

- Tools: Claude (five custom skills: content-schedule-planner, daily-data-sync, fb-posts-chain-linker, guidelines-manager, performance-analysis), Airtable (content + performance + follower tracking), Marblism Social Media Manager (weekly still image only), GoHighLevel (scheduling), a custom GPT (script humanization), ElevenLabs (voice cloning from my own recorded sample), HeyGen (avatar video), Nanobanana and OpenAI Image (avatar generation — originally 7 personal photos at different angles via Nanobanana, expanded to ~30 avatar variants via OpenAI Image).

- Methodology: I used Claude to write the original guidelines document and fed it to Marblism, which generated the scripts. To check whether it was actually following those guidelines, I built a separate Claude skill whose only job was verification — and that's what exposed the problem: Marblism was violating its own guidelines on a regular basis, despite them being explicit. That finding is what justified moving generation to Claude entirely rather than patching Marblism. Claude now handles script generation end to end, under a strict four-part architecture: a hook that must pass a blind classification test (written and graded in separate passes, so the system can't grade its own homework), a personal bridge selected from a tiered library of my own real experiences and matched to the hook by specificity, a named framework pulled from an approved list of leadership authorities, and a closing action tied to the hook type. Production then hands off to a chain of specialist tools — ElevenLabs for voice, HeyGen for the video itself, avatars sourced from my own photos and generated at scale — with GoHighLevel handling the schedule and Marblism narrowed down to just the one still image a week it's actually good at. The loop that keeps all of this honest: every post's real performance (reach, click rate, watch time) is exported nightly from Meta Business Suite and synced into Airtable by a dedicated skill; a second skill checks that data against the guidelines' own published targets; and when the data says a rule isn't holding, a third skill revises the guidelines document itself. Every change to the rules is traceable to specific data, not intuition — that loop is what's taken the document from v9 to v13.

- Learnings: Separating generation from validation is what actually surfaces failure, and that lesson showed up twice. First at the platform level — Marblism generating scripts, a Claude skill checking them against the rules, which is what caught the regular violations. Then again, later, inside Claude's own process: a hook-writing pass and its grading pass turned out to be the same step in disguise, letting weak hooks certify themselves and quietly costing six straight underperforming posts in the highest-value weekly slot. The fix both times was the same principle — keep the check genuinely independent of the thing it's checking, not just procedurally separate.

- Relevant URLs: The page itself: https://www.facebook.com/RamonValadez67 — the automation infrastructure behind it (Airtable base, skills) is internal and not public-facing.