Social media marketing automation tools that read their own report card before every draft

Those tools are calendar schedulers. The product is a grid of time slots you drag content into, plus an optional AI caption generator that uses a shared tone preset. S4L has no calendar. The core primitive is a Postgres posts table and a Python script (scripts/top_performers.py) that re-ranks its own past output against a composite SCORE_SQL on every engage.sh cycle. The output of that script becomes a verbatim block of the next drafting prompt. In other words, the automation the listicles describe is about scheduling; the automation here is about learning.

Because comments are the only signal that proves a person stopped scrolling and wrote something in response. Impressions and views measure algorithmic distribution, not what the content taught anyone. The comment at line 26 of top_performers.py is explicit: 'Comments are the strongest imitation signal (real discussion), upvotes are second, views deliberately excluded (viral-by-algorithm != a pattern worth imitating).' The 3x weighting is there because comments are rarer than upvotes on every platform tested.

Because the OP's own self-upvote inflates the score returned by the Reddit API. Without compensation, every post starts at 1 upvote instead of 0, which pollutes the ranking when a post has no genuine engagement. GREATEST(0, COALESCE(upvotes,0) - 1) on lines 32-33 removes that 1-vote floor on Reddit only. Twitter and LinkedIn do not have the same self-upvote problem, so their score is just COALESCE(upvotes, 0).

Because a winning post that happened to mention 'fazm' or 'cyrano' would teach the next draft that mentioning those names is a winning pattern. It is not; those wins come from the surrounding context, not the brand mention. PRODUCT_NAMES on lines 62-65 lists the 9 tokens currently filtered: fazm, assrt, pieline, cyrano, terminator, mk0r, s4l, vipassana.cool, vipassana-cool. has_anti_pattern() on lines 83-93 runs a lower() substring check for each, plus a URL check for http://, https://, and www. The function returns True for any hit, and top posts that return True are dropped from the report.

get_top_posts on lines 205-247 fetches limit * 3 rows from Postgres, then runs the anti-pattern filter, then slices to limit. With limit=15 that is 45 fetched to get roughly 15 clean. If fewer than 15 survive (a very self-promotional account), the function returns whatever is left. The point is that the LLM never sees the self-promoting winners, even if that means the report is shorter.

Lines 96-124: 'mentions <product name>', 'contains URL/link', 'product-adjacent pitch (phone/call capture)', 'product-adjacent (mentions own project)' for mcp-server/desktop-agent/accessibility-api matches, 'too many questions (reads as interrogation)' when content.count('?') >= 3, 'curious_probe style (negative avg on Reddit)' when 'curious' appears with a question mark, 'too short without being punchy' when len(content) < 100, and a fallback 'likely wrong subreddit or off-topic' when nothing else matched. Each bottom post gets its reasons joined with ' | '.

Because the scale of reactions differs by platform. On Reddit, a post with 10 upvotes and a few comments is noteworthy. On LinkedIn, a post with 3 reactions and 1 comment is also noteworthy. PLATFORM_MIN_SCORE on lines 39-46 sets reddit at 10, twitter and x at 5, linkedin at 3, moltbook at 3, github at 3. A default of 5 covers anything unknown. If no platform is passed, min_score_for(None) returns 5.

skill/engage.sh line 75 captures TOP_REPORT=$(python3 $REPO_DIR/scripts/top_performers.py --platform reddit 2>/dev/null || echo '(top performers report unavailable)'). Line 150 of the same file, inside the big heredoc that becomes the Claude prompt, literally includes '## FEEDBACK FROM PAST PERFORMANCE (use this to write better replies): $TOP_REPORT'. The whole report lands as a top-level block of the drafting context, ahead of the engagement styles block.

It changes two things. First, the style ranking table shifts which archetypes Claude reaches for first. Storyteller with avg_cm=1.7 on Reddit will be picked more often than snarky_oneliner with avg_cm=0.4 because the prompt now shows that gap. Second, the bottom-post reasons teach Claude not to repeat specific failure patterns; after a few cycles where 'curious_probe style (negative avg on Reddit)' appears in bottom-10, the model stops drafting curious_probe on Reddit at all, which is why PLATFORM_POLICY also bans it at the hard level.

Yes. PLATFORM_POLICY in scripts/engagement_styles.py is tone policy, not data policy. It bans curious_probe on Reddit and snarky_oneliner on LinkedIn regardless of what the numbers say. top_performers.py is performance data, not policy. The two layers stack. A style has to clear both: not banned by PLATFORM_POLICY, and surfaced (or demoted) by the live score ranking. This separation is deliberate: policy is about what we want to be seen doing, performance is about what works.

Lines 55-58 of top_performers.py: 'DO NOT REMOVE OR SIMPLIFY THE FUNCTIONS BELOW. These are data-driven improvements based on analysis of 3,000+ posts. They have been reverted by other agents twice already. Protected by pre-commit hook. See CLAUDE.md.' The block got refactored out by AI coding agents who saw 'unused PRODUCT_NAMES' in a static check and removed it. The warning plus the pre-commit hook exist so the block survives future refactors.

The pattern is small. A SQL query with a composite score, a per-platform threshold dict, a substring-based anti-pattern filter, and a rule-based failure annotator. If you already log your own posts to a database with an upvotes or reactions field and you run your drafts through an LLM, you can copy this shape verbatim, change the column names, and inject the output into your prompt. The only S4L-specific piece is which engagement styles are tracked; you would pick your own.