AGENT SKILL

data-digest

Use when asked for a recurring data digest, a "what changed since last time" update, a weekly metrics readout, or a trend/anomaly check on a CSV, spreadsheet, SQL result, or connected data source — compare against the prior run, lead with the deltas, and call out anomalies.

License 许可
Apache-2.0
所属仓库
matangilabs/kortny
最近更新
2026-10-02
Stars
★ 6
Forks
3

01 前言 Front Matter

name
data-digest
metadata
{ "tags": "data, digest, recurring, metrics, trends, anomalies, deltas", "version": "1.0.0", "display_name": "Data Digest" }
description
Use when asked for a recurring data digest, a "what changed since last time" update, a weekly metrics readout, or a trend/anomaly check on a CSV, spreadsheet, SQL result, or connected data source — compare against the prior run, lead with the deltas, and call out anomalies.

02 资源文件 1

03 SKILL.md 正文

Goal

Produce a recurring data readout whose value is comparison: what moved since last time, what's anomalous, what's worth acting on — across whatever source the data lives in.

This is the recurring, longitudinal cousin of data-brief. data-brief turns one posted file into a one-shot story for a busy reader. data-digest runs again and again on the same metrics and earns its keep by comparing this run to the last — deltas, trend direction, anomalies. If there's nothing to compare against (a single dropped file, first-ever look), use data-brief instead.

Steps

  1. Identify the source and the cadence. Source can be a posted CSV/spreadsheet, a SQL query result, or a connected data tool (a Composio integration like a warehouse, analytics, or BI connector when one is available). Note which it is and the period each run covers.
  2. Find the prior baseline. This is the whole point. Look, in order, for:
    • the previous digest in prior thread context or episodes (Kortny's own last run is the cleanest baseline — reuse its numbers and wording);
    • a prior period inside the same dataset (last week's rows vs this week's);
    • workspace facts recording targets or expected ranges. If no baseline exists, say "first digest — no comparison yet, establishing the baseline" and report levels only.
  3. Compute the deltas for each tracked metric: absolute and percent change vs the baseline, and the direction. See references/comparison.md for the delta and anomaly conventions (what counts as material, how to handle small denominators, week-over-week vs same-period-last-year).
  4. Flag anomalies — a metric outside its usual range, a sign flip, a sudden spike or drop, a metric that stopped updating. An anomaly gets a one-line "why this might be" only if the data supports it; otherwise just flag it.
  5. Lead with what changed, not with a full table. The standout movements go first; steady metrics get a single "holding steady" line.
  6. Name the action if the deltas support one ("retention dropped two weeks running — worth a look before it compounds").

Output shape (Slack mrkdwn)

*Data digest — week of Jun 9* (vs week of Jun 2)

*Moved*
• Activation 41% → 58% (+17pts) — best week this quarter.
• Weekly active 12.1k → 11.4k (−6%) — second week down.

*Anomalies*
• Signup→trial conversion jumped to 22% (usual ~12%) — check the source; possible double-count.

*Holding steady*
• Churn ~3%, NPS 41 — no material change.

*Worth acting on:* the two-week active-user slide. Want the cohort cut?

Offer the table or a chart as a follow-up artifact (chart-maker / spreadsheet-builder) rather than dumping it inline.

Rules

  • Every number is from the data. Never round a guess into a figure, and never fabricate a baseline to manufacture a delta.
  • Percent changes need a real denominator — on tiny numbers report the absolute change and say the base is small, don't post a misleading "+300%".
  • If the source didn't refresh (same rows as last run), say so plainly: "source hasn't updated since last digest" — a stale digest pretending to be fresh is worse than none.
  • Anomaly callouts are observations, not diagnoses. Speculate on cause only when the data points at it; otherwise flag and ask.
  • Keep prior-run wording stable so readers can track a metric across digests.
  • This is built to run on a schedule — pair with Kortny's scheduler for "post this every Monday"; the body is identical by hand or on schedule, and the scheduled run is what makes the comparison automatic.

Worked example

Scheduled Monday run over a metrics sheet, prior digest in episodes:

Data digest — week of Jun 9 (vs week of Jun 2)

Moved • Trial→paid 18% → 24% (+6pts) — strongest since March. • Weekly active 12.1k → 11.4k (−6%) — down two weeks now.

Anomalies • Support tickets 90 → 210 (+133%) — spike concentrated on Thursday; possible incident, not a trend.

Holding steady • MRR growth ~4%/wk, churn ~3%.

Worth acting on: the active-user decline and Thursday's ticket spike — they may be the same story. Want me to pull the Thursday cohort?

04 同仓库更多 Skills