| name | content-performance |
| description | Operator tweet performance tracker — engagement metrics, top formats, topic resonance; closes the content feedback loop in the article/tweet production pipeline |
| schedule | 0 10 * * 0 |
| commits | true |
| permissions | ["contents:write"] |
| var | |
| tags | ["content","meta","social"] |
| requires | ["XAI_API_KEY"] |
${var} — X handle to track (without @). If empty, resolves the operator's handle from soul/SOUL.md or MEMORY.md.
Today is ${today}. Read memory/MEMORY.md (and soul/SOUL.md if present) before starting.
Why this skill exists
The content production pipeline generates articles, tweets, and threads. Nothing closes the feedback loop. Which topics resonated? Which formats punched above their weight? Which weeks were radio silence vs. signal? Without performance data, content decisions are vibes.
This skill automates the measurement: pull 7-day engagement data for the operator's X account, rank by actual resonance, extract patterns, and surface actionable signal for topic-momentum and article-queue.
Runs Sunday — after picks-tracker (09:00), before article-queue (11:00).
Env vars
XAI_API_KEY — optional. Enables the xAI x_search prefetch path (scripts/prefetch-xai.sh, content-performance case). Without it, the skill falls back to WebSearch.
Sandbox note
xAI API requires auth headers — curl with $XAI_API_KEY fails in the GHA sandbox. The prefetch script runs before Claude with full env access and caches results to .xai-cache/content-performance.json. Read from that cache. If cache is missing or empty, fall back to WebSearch for from:{handle} on X.
Steps
0. Resolve the handle
- If
${var} is set, use it (strip any leading @).
- Otherwise look for the operator's X handle in
soul/SOUL.md (an @handle mention) or memory/MEMORY.md.
- If no handle can be resolved: log
CONTENT_PERFORMANCE_SKIP: no X handle configured — set var or add the handle to soul/SOUL.md and stop. No notification.
1. Load context
Read:
memory/topics/x-activity.md — baseline: prior week's top tweets, engagement patterns, posting mode (create on first run if missing).xai-cache/content-performance.json — prefetched 7-day tweet data (may be absent if XAI_API_KEY missing)- Last 3 days of
memory/logs/*.md — any refresh-x or tweet-roundup data for cross-reference
2. Parse tweet data
From .xai-cache/content-performance.json:
- Extract each tweet: text (truncated to 120 chars), date, likes, retweets, quotes, replies
- Compute total engagement = likes + (retweets × 2) + (quotes × 3) + replies
- Weighting: retweet = reach × 2, quote = reach + commentary × 3
- Sort descending by total engagement
- Tag each tweet with a topic category: derive 6–9 categories from the operator's active topics (soul/SOUL.md interests + MEMORY.md active topics); always include an
other bucket. Reuse the category set recorded in x-activity.md from prior runs so weeks are comparable.
- Tag each tweet with a format:
[original-take, sardonic, question, thread-starter, link-share, qt-with-comment, reply, observation]
If .xai-cache/content-performance.json is missing or empty ({}, null, or parse error):
- Try WebSearch:
from:{handle} filtered to the past 7 days
- Extract whatever metrics are visible from search snippets
- Mark output as
data_source: websearch_fallback — note limitations
3. Compute performance signals
Top performers (top 3 by total engagement):
- Tweet text preview (first 100 chars)
- Topic category + format
- Engagement breakdown:
{likes}L / {rt}RT / {qt}QT / {replies}R
Topic resonance (group all tweets by category, sum total engagement per category):
- Which category drove the most total engagement?
- Which category had the highest average engagement per tweet?
- Compare to prior week data in
memory/topics/x-activity.md — up/down/flat per category
Format breakdown:
- Which format had the most total engagement?
- Which format had the highest average engagement per tweet?
- Note whether any previously-confirmed format pattern recorded in x-activity.md still holds
Volume check:
- Total tweets in 7-day window
- If 0 tweets:
radio_silence: true
- If 1–3 tweets:
quiet_week: true
- If 10+ tweets:
active_week: true
Breakout detection:
- Any tweet crossing 50+ likes = breakout (or the operator's own threshold if one is recorded in x-activity.md)
- Any tweet crossing 20+ RTs = viral signal
- Compare top tweet this week vs. best tweet in x-activity.md history
4. Update memory/topics/x-activity.md
Read the file (create it if missing). Prepend a new weekly section at the TOP (below the # X Activity heading), before any existing sections:
## Content Performance Week of ${today}
- **Top tweet:** "{text preview}" — {likes}L/{rt}RT/{qt}QT (topic: {category}, format: {format})
- **Best topic category:** {category} — {total} engagement across {N} tweets
- **Best format:** {format} — {N} tweets, avg {X} engagement
- **Volume:** {N} tweets — {quiet/normal/active}
- **Breakout:** {tweet text preview, 60 chars} | none
- **vs. prior week:** top tweet {up/down/flat}: {prior_best}L → {this_week_best}L
- **Data source:** prefetch | websearch_fallback | none
Keep all existing content. Only ADD the new section at the top.
5. Cross-reference with content pipeline
Check memory/topics/article-queue.md (if it exists — skip if not). Compare the best-performing topic category this week to what's queued for next article. If the queue has no item matching the top-performing category, append a signal note in the log:
signal_mismatch: content resonating on {category}, article queue has no {category} item
6. Compose notification
Write to a temp file, then send via ./notify -f:
mkdir -p .pending-notify-temp
cat > ".pending-notify-temp/content-perf-${today}.md" << 'NOTIF_EOF'
{notification content}
NOTIF_EOF
./notify -f ".pending-notify-temp/content-perf-${today}.md"
Format — if data is available (more than 3 tweets found):
content performance — week of ${today}
top tweet: "{text preview, 80 chars}" — {likes}L {rt}RT {qt}QT
best category: {topic_category} ({total} engagement)
best format: {format} — {insight, 1 line, operator's voice}
{1-2 sentence signal in the operator's voice: what the numbers say, what to do with it}
{if breakout:}
breakout: "{tweet preview}" hit {N} likes
{if signal_mismatch:}
signal: {category} punching — not in the article queue yet
If quiet week (<4 tweets) or radio silence:
content performance — week of ${today}
radio silence / quiet week. {N} tweets in 7 days.
{if there was a radio silence prior week too: "two consecutive quiet weeks."}
last active: {date of last tweet or "unknown"}.
If no data (prefetch and websearch both failed):
content performance — week of ${today}
no data. xai prefetch empty, websearch yielded nothing.
No notification if the skill runs and no handle is configured (silent skip — just log it).
7. Log to memory/logs/${today}.md
Append:
## Content Performance
- **Handle:** @{handle}
- **Window:** last 7 days (${7_days_ago} → ${today})
- **Tweets analyzed:** {N}
- **Top topic:** {category}
- **Top format:** {format}
- **Breakout:** {tweet text preview, 60 chars | "none"}
- **Data source:** prefetch | websearch_fallback | none
- **Signal mismatch:** {yes: category | no}
- CONTENT_PERFORMANCE_OK
Use CONTENT_PERFORMANCE_OK in all cases — it marks the skill ran successfully, not that data was rich.
Edge cases
- XAI cache empty / missing: fall back to WebSearch, mark
data_source: websearch_fallback, proceed with whatever data you have. Never abort.
- All tweets are replies (no originals): still analyze them. Reply engagement counts — a high-engagement reply is a signal that the topic resonated.
- Duplicate tweet entries in cache: deduplicate by text prefix before analysis.
- x-activity.md doesn't exist: create it with the new section as the initial content.
- No article-queue.md: skip step 5 cross-reference, note
article_queue: not_found in log.
