Investigate Metric

Investigating a metric change

For “why did X change?” questions about a saved insight, dashboard tile, or pasted query. Don’t load this skill for plain “what is X?” questions — only when there’s an observed change to explain.

Tools

Targets PostHog MCP v2. Typed query tools accept the query body directly — pass kind, series, dateRange as top-level fields, do not wrap in InsightVizNode.

Plus the standard PostHog tools the playbooks reference by name (feature-flag-get-all, experiment-get-all, annotations-list, query-error-tracking-issues-list, query-logs, query-session-recordings-list, cohorts-list/-create, annotation-create, insight-create).

Helper scripts

• compare_to_prior_periods.py (opens in a new tab) — auto-detects interval and compares recent values to the natural cycle (day-of-week, hour-of-week, or sequential). Use to resolve step 2.2 cheaply.
• breakdown_attribution.py (opens in a new tab) — ranks breakdown segments by absolute delta and flags offsetting moves.

python3 scripts/compare_to_prior_periods.py < query_result.json
WINDOW=7 python3 scripts/breakdown_attribution.py < breakdown_result.json

Step 1 — Classify the metric

Read query.kind from the source the user pointed at:

• Saved insight (URL, short_id): posthog:insight-get → query.kind. Use posthog:insight-query if you also need the numbers.
• A query you already ran or the user pasted: read kind directly.
• Nothing pointed at: ask for the URL or short_id. Don’t guess.

If kind === "TrendsQuery" and trendsFilter.display === "BoxPlot", use box-plot-playbook.md (opens in a new tab) — distribution metric, no breakdowns.

For HogQLQuery insights, classify by the SQL’s shape: count over time → trend playbook, multi-step conversion → funnel playbook, cohort return → retention playbook. Run the SQL through posthog:execute-sql to get the data, then follow the closest playbook’s steps. See HogQL insights in shared-patterns.md.

If the user’s question spans multiple kinds, run the playbooks in sequence.

Step 2 — Common opening moves

2.1 Confirm the anomaly

Run the primary tool. Record baseline, current, delta (absolute and %), and the start of the anomaly window.

2.2 Variance check

Widen to 3–4× the user’s interval (or use compareFilter: {"compare": true} on TrendsQuery / StickinessQuery; for other kinds run two date ranges). Pipe the widened result through compare_to_prior_periods.py (opens in a new tab) — it flags seasonality, partial right-edge buckets, and real anomalies. If the movement is normal variance, report that and stop.

2.3 Known changes in the window

In rough order of signal:

• posthog:feature-flag-get-all → flags with updated_at near the anomaly start.
• posthog:experiment-get-all → start_date / end_date near the start.
• posthog:annotations-list → date_marker near the start.
• git log for the window if the repo is reachable (highest signal when available).

Any match is a hypothesis to confirm in the playbook (usually via breakdown on $feature/<flag_key>, app_version, or utm_source).

Step 3 — Run the playbook

Open the playbook for the kind from Step 1 and follow its numbered steps. Carry the record from 2.1 and any candidates from 2.3 into it.

Step 4 — Cross-check

Pick a segment the suspected cause should not have affected and rerun there. Stable in the control = strong hypothesis; moved too = expand the investigation. Skip when 2.2 already explained the movement.

Step 5 — Write findings

Use the format below. Offer to save key charts via posthog:insight-create. If a cause is found and no annotation marks it, offer posthog:annotation-create. See common-causes.md (opens in a new tab) for the cause taxonomy.

# Investigation: <metric>

**Anomaly**: <baseline> → <current> (<delta>) starting <date>

## Likely cause

<one sentence>

**Confidence**: low | medium | high — <one-line reason>

**Evidence**

- <query result>
- <flag / experiment / annotation / commit if applicable>

## Possible causes (ruled out)

- <hypothesis>: <why>

## Affected segment

Confidence rule of thumb:

• high — multiple independent signals corroborate (e.g. a segment isolates the delta and a flag/version aligns and an error or annotation matches).
• medium — one corroborating signal, or strong pattern-match without a cross-check.
• low — pattern matches a known cause but no corroboration, or the data only rules things out.

Link insights and dashboards inline: [Name](/insights/short_id).

Reference files

• Playbooks: trend (opens in a new tab), box-plot (opens in a new tab), funnel (opens in a new tab), retention (opens in a new tab), stickiness, ,