| name | posting-time-trends |
| description | Weekly analysis of WHEN good-matching job posts appear — day-of-week and hour-of-day distributions of high-scoring posts in the scraped corpus — so you can track shifts in posting patterns and tune when to check the digest / run the alert poller. Use when asked about best time/day to look for jobs, to refresh the timing recommendation, or to run the recurring trend report. |
Posting-time trends
Answers "when do the good posts get posted?" from the real corpus
(data/posts.json) and tracks how that drifts week to week.
The output drives two decisions: when to manually check the digest, and the
cron window for the morning alert poller (see linkedin-job-digest). This is
descriptive analytics only — it does not scrape, score, or change config.
Run it
Self-contained (stdlib only), reads data/posts.json, writes
data/trends-YYYY-MM-DD.md and prints the same to stdout:
uv run python .claude/skills/posting-time-trends/analyze_trends.py --since-days 30
--since-days N — only posts from the last N days (use this for trend
tracking; 0/omit = all history, which dilutes recent shifts).--tz OFFSET — report timezone, default -5 (Bogotá). All clock times are in
this zone.--min-score S [S ...] — score cohorts, default 15 20 (the strong and
near-ideal tiers; mirror the digest/alert floors here).--out PATH, --posts PATH — overrides.
Weekly cadence: run every Monday over --since-days 30 (a rolling month is
enough signal without being swamped by stale history). Schedule via the
schedule skill (a cron routine) or a /loop. Compare each week's report to the
prior data/trends-*.md to spot movement.
What it reports
- Per score cohort (>=15, >=20):
- Day of week — count + bar, peak flagged. Where the volume of good
matches concentrates.
- Top hours — busiest hours in your zone, each annotated with the
regional business clock (US-ET / CEST / IST) so you can attribute a peak
to a region's working hours (see caveats).
- Top-3 hours per day — whether timing diverges by weekday (it does — e.g.
a midday Tuesday peak vs. morning Wed–Fri).
- Hourly quality curve (score >= 0 only) — count and average score per
hour, to find the slot that is both busy and high-quality (mean is dragged
down by deep-negative flagged posts if you don't filter, hence positives-only).
- Company pages by volume — how many posts each company page authored in the
window (
/company/<slug> of author_profile_url; personal-profile posts are
excluded from this count). High counts flag likely job indexers/aggregators;
pages already muted in
filtered_accounts.json are tagged
[filtered]. Use this to decide which new pages to add to that filter (they're
then dropped from the digest/alerts — see linkedin-job-digest).
How to read it (methodology derived from the corpus)
- Translate hours to the poster's region, not yours. The corpus is
India/Europe/US-heavy; each region posts at its own ~9am, so a peak in your
zone is really "9am somewhere." The regional clock annotations exist for this.
Historically the strong cohort peaked at 08–09:00 Bogotá ≈ 09–10:00 US-ET
(the US-morning wave, which carries the most profile-relevant/remote posts),
with a high-volume overnight band = Europe+India morning.
- Day pattern: Tue–Thu has been the reliable core; weekends are too sparse
to trust (single-digit n — treat "peaks" there as noise). Watch whether the
weekday peak moves (it has shifted between Thursday and Monday as the corpus
grew — that drift is exactly what weekly runs catch).
- Quality vs volume are different questions. A busy hour isn't always the
highest-scoring; use the hourly quality curve to confirm a slot is worth it.
Caveats (state these when reporting)
- Attribution is inferred from timestamps, not poster location — regional
mapping is a heuristic. For ground truth you'd capture author region at scrape
time (a known follow-up in
linkedin-job-digest).
- DST: regional annotations use standard-time offsets; US/EU clocks shift ±1h
in summer. Bogotá (the default report zone) has no DST.
- Small cohorts wobble. With
--since-days 30 the >=20 cohort is ~80 posts;
per-day-per-hour cells are tiny (counts of 1–6), so read tendencies, not exact
hours. Don't over-fit the schedule to a single week.
- Posts dedupe by id but near-duplicate reposts aren't collapsed, so a heavily
reposted job can nudge a given hour/day — corroborate big shifts across weeks.
