brand-auto-tuner
Analyze winner vs loser articles to find features that distinguish them, then suggest brand-guideline.yaml updates. Use after ≥10 articles have outcome tags. Closes the data-loop: production → measurement → learning → next-cycle adjustment.
XuanRanL · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add XuanRanL/loamwright-SEO-Skill --skill brand-auto-tuner技能说明
name: brand-auto-tuner description: Analyze winner vs loser articles to find features that distinguish them, then suggest brand-guideline.yaml updates. Use after ≥10 articles have outcome tags. Closes the data-loop: production → measurement → learning → next-cycle adjustment. allowed-tools: [Bash, Read] disable-model-invocation: false user-invocable: true
Brand Auto-Tuner
The third leg of the closed-loop. After outcome_tagger labels articles as winner / mid / loser based on real GSC + AI citation data, this finds the patterns that distinguish winners and suggests brand-guideline tweaks.
When to invoke
- After 10+ articles have outcome tags (run
outcome_tagger.pyfirst) - Monthly or quarterly, as the dataset grows
- After a major content campaign to see what worked
- Before starting a new content campaign (lock in learnings)
/tuneror/auto-tune(user invokes)
How to invoke
# Single site
python -m scripts._core.brand_auto_tuner --site my-site
# All sites in portfolio
python -m scripts._core.brand_auto_tuner --all-sites
# Higher threshold for noisier datasets
python -m scripts._core.brand_auto_tuner --site my-site --min-per-group 10
Output
Two files per site:
projects/{slug}/brand-tuner-report-{date}.json— full findingsprojects/{slug}/brand-guideline.diff.{date}.yaml— suggested overlay
The diff is NOT auto-applied. User reviews, then manually merges desired
findings into brand-guideline.yaml.
Sample output
━━ Brand tuner: my-fishing-site ━━━━━━━━━━━━━━━━━━━━━━━
Articles: 47 (winners 12, mid 22, losers 13)
Confidence: medium
Findings: 6
Output: projects/my-site/brand-tuner-report-2026-05-19.json
Diff: projects/my-site/brand-guideline.diff.2026-05-19.yaml
Significant patterns:
🔴 [large ] first_person_pct
winners have higher first_person_pct (mean 14.3 vs losers 6.2).
Consider targeting around 14.3 (band ±15%).
🔴 [large ] info_gain_marker_count
winners have higher info_gain_marker_count (mean 3.1 vs losers 1.2).
Consider targeting around 3.1 (band ±15%).
🟡 [medium] tier1_citation_count
winners have higher tier1_citation_count (mean 4.2 vs losers 2.1).
Consider targeting around 4.2 (band ±15%).
🟡 [medium] format=listicle
format='listicle' wins more often (rate 67% vs portfolio avg 26%).
Bias towards this format.
🟡 [medium] ai_slop_score
winners have lower ai_slop_score (mean 18.2 vs losers 31.5).
Consider targeting around 18.2 (band ±15%).
🟡 [medium] format=how-to-guide
format='how-to-guide' wins less often (rate 12% vs portfolio avg 26%).
Avoid this format.
Method
Numeric features (Cohen's d effect size)
For each feature in [word_count, h2_count, citation_count, tier1_citation_count, image_count, info_gain_marker_count, first_person_pct, ai_slop_score, core_eeat_score, cite_score, reviewer_score, ...]:
- Pool winners' values + losers' values
- Cohen's d = (mean_winner - mean_loser) / pooled_SD
- |d| ≥ 0.8 → large (worth acting on)
- 0.5 ≤ |d| < 0.8 → medium (directional hint)
- < 0.5 → ignored (too noisy)
Categorical features (win-rate delta)
For each category of format, voice, purpose:
- Compute win rate per category vs portfolio-wide average
- Delta ≥ 0.10 → flag direction
- Delta ≥ 0.20 → strong signal
Confidence tiers
| Winners n | Confidence | What to do |
|---|---|---|
| <5 | very low | Wait — publish more |
| 5-9 | low | Directional only; don't tune |
| 10-19 | medium | Apply medium+ findings cautiously |
| 20+ | high | Apply large findings; review medium |
How to apply findings
- Open
projects/{slug}/brand-guideline.diff.{date}.yaml - Review each suggestion against your own judgment
- Manually copy approved items into
projects/{slug}/brand-guideline.yaml - Move applied finding from
pending→appliedin the report file - Run brand-tuner again after next batch to verify the change helped
This is intentionally NOT auto-apply. Brand voice is too important to algorithm out. The tuner finds signals; the human decides.
Limitations
- Statistical method is intentionally simple (Cohen's d, no regression). With small N, false positives are common. Manual review essential.
- Confounding: longer articles may correlate with everything; doesn't prove causation. The tuner reports correlations; reader interprets.
- Doesn't account for time-decay: an article that became a winner after 6 months has the same weight as one that won in 2 weeks.
What this skill does NOT do
- ❌ Auto-edit brand-guideline.yaml
- ❌ Run regressions / ML models (intentionally simple)
- ❌ Determine causation (correlation only)
- ❌ Combine learnings across clients (per-site only — different niches)
Composition
Closed-loop, quarterly:
1. gsc_api_ingest --all-sites --mode last-28
2. ai_citation_tracker --all-sites --top-priority 50
3. outcome_tagger --site X → labels winners/mid/losers
4. brand_auto_tuner --site X → suggests guideline updates
5. Human review → apply approved findings
6. Next batch: /batch-article picks up tuned guideline