TelemetryDeck Analytics

Answers product-analytics questions against the TelemetryDeck v3 API via the bundled tdq CLI. Stdlib-only, cross-platform, OS-native secret storage, no .env, no repo-local files.

This skill ships as part of the telemetrydeck-analytics Claude plugin. While the plugin is enabled, the tdq wrapper is on PATH, so you can call it directly as tdq <subcommand>. If the wrapper isn't available (e.g. running outside Claude Code), fall back to python3 "$SKILL_DIR/tdq.py" <subcommand>.

User-invokable slash skills

The plugin exposes focused slash commands for common tasks — prefer them when the user's intent maps cleanly to one:

This main skill handles the ambient / conversational case ("is the Pro share growing?", "why aren't signals arriving?") — it's model-invoked when a question needs analytics but doesn't match a slash command exactly.

First-run setup (once per machine)

tdq login

Prompts for email and password, mints a bearer, then lists the apps on your TelemetryDeck account and asks you to pick one. No need to know the app UUID up front. If the listing endpoint isn't exposed for your account, the CLI falls back to a manual UUID prompt.

Secrets (password + bearer) go into the OS-native store:

• macOS — Keychain via security (service telemetrydeck-cli).
• Linux — libsecret via secret-tool if installed (GNOME Keyring, KWallet-libsecret, etc.).
• Windows / Linux without secret-tool — file fallback at <config-dir>/secrets.json mode 0600, with a one-time stderr warning. Install libsecret-tools on Linux (apt install libsecret-tools) for a proper keyring.

Non-secret state (email, registered apps, token expiry) lives in a platform-appropriate config directory:

• macOS: ~/Library/Application Support/TelemetryDeckCLI/config.json
• Linux: $XDG_CONFIG_HOME/TelemetryDeckCLI/config.json (or ~/.config/...)
• Windows: %APPDATA%\TelemetryDeckCLI\config.json

After setup, every subcommand auto-refreshes the bearer on expiry or HTTP 401.

$SKILL_DIR = the absolute directory of this SKILL.md (inside the plugin at skills/analytics/). Use it to read bundled reference files (tql/index.md, reference.md); prefer the tdq PATH wrapper for invocations when it is available, otherwise fall back to python3 "$SKILL_DIR/tdq.py" as noted above.

Verify setup:

tdq doctor

Reports pass/fail for platform (+ which secret backend is active), config, secret store, auth, and a trivial query round-trip with remediation hints.

Managing multiple apps

If your TelemetryDeck account has more than one app, the CLI registers them all at login time. Switch between them, add more, or remove entries with the apps subcommand:

tdq apps                       # list registered apps, * marks current
tdq apps use MyApp             # switch by display name
tdq apps use <uuid>            # switch by UUID
tdq apps use 2                 # switch by 1-based index from `apps`
tdq apps add <uuid> --name "New App" [--set-current]
tdq apps remove <uuid-or-name-or-index>
tdq apps refresh               # re-pull the list from the API

Every query command also accepts a one-off --app-id <UUID> flag that bypasses the current-app setting for a single invocation.

Subcommands

All query-producing commands share --format table|csv|json (default: table — markdown-ready for direct paste into a report), --app-id <UUID>, and --raw.

--raw / TDQ_RAW=1: prints the raw (post-envelope-unwrap) TelemetryDeck response to stderr before the formatted output. Reach for this the moment the formatted table says (no rows) — if the raw blob is non-empty, the parser dropped data (file a bug or re-check _flatten_result); if the raw blob itself is empty or a success-envelope with "rows": [], the query is genuinely empty and you need to widen the window or fix the filter.

Named intervals (--interval) accepted by dau, mau, groupby: last-Nd for any N (e.g. last-7d, last-30d, last-90d), plus calendar-aware windows last-week, this-week, last-month, this-month, last-year, this-year, with aliases wtd, mtd, ytd. --interval wins over --days / --months when both are set.

Period comparison (--compare prior-period) on dau, mau, groupby: runs the same query against the immediately prior period, joins on the dimension (for topN / groupBy) or sums metrics (for timeseries), and emits one row per key with current, prior, delta, pct columns. For last-Nd, prior = the preceding N days. For last-week/last-month/last-year, prior = the completed period before that. For this-X to-date intervals, prior = the full last-X calendar period (note: this compares partial-to-full, which is usually what you want for "trending up?" questions but may understate growth early in a period — call this out in reports).

The query subcommand auto-injects the mandatory appID + isTestMode=false filter when the TQL has no filter key or sets it to the sentinel { "__auto_app_and_test_mode_filter__": true }. If you write your own filter, include both selectors yourself — otherwise the result mixes other apps in the tenant and test-mode traffic.

Worked example — "Is license mix shifting?"

User: "Has our Pro share grown month-over-month?"

Claude runs:

tdq groupby license --event App_launched \
  --interval last-30d --compare prior-period

Output:

| license | count_current | count_prior | delta | pct     |
|---------|---------------|-------------|-------|---------|
| Free    | 1823          | 1520        | 303   | +19.9%  |
| Pro     | 412           | 380         | 32    | +8.4%   |
| Trial   | 0             | 27          | -27   | —       |

Claude reports inline:

License mix, last 30d vs. prior 30d (opt-in sample): Pro event share moved from 19.7% → 18.4% despite absolute Pro events rising +8.4% (412 vs 380). Free grew faster (+19.9%), diluting the Pro share. The Trial segment disappeared (27 → 0) — worth confirming with engineering whether the Trial license value is still emitted. Counts are App_launched events, not unique users; for user-share, re-run with --metric users.

Worked example — "What's our license mix?"

User: "What's our Free/Pro split over the last 30 days?"

Claude runs:

tdq groupby license --event App_launched --days 30

Output:

| timestamp            | license | count |
|----------------------|---------|-------|
| 2026-03-19T00:00:00Z | Free    | 1823  |
| 2026-03-19T00:00:00Z | Pro     | 412   |
| 2026-03-19T00:00:00Z |         | 98    |

Claude reports inline:

License mix, last 30d (opt-in sample): Free 78.2% (1823), Pro 17.7% (412), unknown/null 4.2% (98). Counts are App_launched events, not unique users — use --metric users for unique-user share. Numbers reflect the opted-in cohort only; note any known opt-in bias when interpreting.

Report structure (only if user asks for a saved report)

Default is inline in chat. If the user wants a file, ask where to save it — write to their absolute path. Suggest ~/Documents/TelemetryDeck/YYYY-MM-DD-<slug>.md or ~/Downloads/. Never create analytics/, reports/, or any new top-level directory in the current working directory.

Structure:

• Question — one sentence.
• Top-line answer — one paragraph, concrete numbers and date range.
• Method — CLI invocation or resolved TQL so the run is reproducible.
• Caveats — opt-in sample bias, test-mode exclusion, channel skew (App Store vs. direct), cohort window.
• Raw result (optional appendix).

Triage heuristics

• "DAU / MAU trend" → dau --interval last-30d, mau --interval last-6d (or mau --months 6).
• "Week-over-week growth?" → dau --interval last-7d --compare prior-period.
• "Is this month beating last month?" → dau --interval this-month --compare prior-period. Remember: this-* compares partial-to-full; note the caveat in the report.
• "License or version split" → groupby license --event App_launched --interval last-30d, groupby majorSystemVersion --metric users --interval last-7d.
• "License mix trend" → groupby license --event App_launched --interval last-30d --compare prior-period.
• "Which events exist?" or "is event X firing?" → events (merged 7d + 30d table). If X is absent from 7d but present in 30d, the event stopped firing — likely gated by a user setting (classic failure: @AppStorage default not mirrored via UserDefaults.register(defaults:), so UserDefaults.bool(forKey:) silently returns false).
• "Where do users drop off?" → write a funnel TQL query and pipe to query -. The CLI doesn't ship a funnel recipe yet; funnel steps vary enough that a bad default is worse than raw JSON.
• "Crashes" → not this skill. Sentry or equivalent.

Raw TQL workflow

For queries the recipes don't cover, pipe TQL to query -:

cat <<'EOF' | tdq query - --format table
{
  "queryType": "topN",
  "dataSource": "telemetry-signals",
  "granularity": "all",
  "aggregations": [{"type": "eventCount", "name": "count"}],
  "metric": {"type": "numeric", "metric": "count"},
  "dimension": {"type": "default", "dimension": "countryCode", "outputName": "country"},
  "filter": { "__auto_app_and_test_mode_filter__": true },
  "relativeIntervals": [{"beginningDate":{"component":"day","offset":-30,"position":"beginning"},
                          "endDate":{"component":"day","offset":0,"position":"end"}}],
  "threshold": 20
}
EOF

For TQL syntax, start at tql/index.md and follow the routing table to the specific topic (query types, filters, aggregators, dimensions, intervals, granularity, funnel, retention, recipes). Each file is focused — load only what you need. reference.md covers only the HTTP / auth / async mechanics; it does not document TQL.

Saved-insight TQL shape (what the server emits)

When the insight subcommand resolves a saved insight to TQL, the server returns a slightly different shape than the recipes above: baseFilters: "thisApp" instead of an explicit appID selector, and aggregators like thetaSketch for user counts. Both styles are valid — the server accepts either. The explicit-filter style in the recipes is more portable (works across apps, machine-verifiable), the baseFilters style is what the dashboard emits. When porting a saved insight to a raw recipe, feel free to rewrite baseFilters: "thisApp" into the explicit appID + isTestMode selector pair without changing semantics.

Debugging empty / weird results

When the formatted output shows (no rows) or doesn't match what you expect:

1. Re-run with --raw (or TDQ_RAW=1). The raw JSON goes to stderr. Compare it to the formatted output — if raw has data but formatted doesn't, the parser dropped rows.
2. tdq.py test — runs three known-good query shapes and dumps raw + parsed row counts. If all three show zero rows, the API may have changed its response envelope. If only one shape is broken, _flatten_result in tdq.py needs an update for that shape.
3. tdq.py doctor — asserts the 7-day round-trip has rows. A doctor failure with "0 rows" means either the app is genuinely empty or the parser regressed; --raw on any other command will tell you which.
4. Widen the window — --days 30 or --interval last-30d. A 1-day window can legitimately be empty on a quiet app.
5. Check the event name — tdq.py events to see what's actually firing. Typos return zero silently.

Guardrails

• Never echo secrets. Don't cat .env, don't print Keychain contents, don't paste the bearer into chat, reports, or commits.
• Never write to CWD without permission. Default is stdout; saved reports go to a user-chosen absolute path.
• Abort runaway queries. The CLI caps at 30 polls / 120s wallclock. If a query hits those limits, tighten --days, narrow --event, or drop the row cap.
• Respect opt-in bias. Analytics samples are self-selected by users who left the setting on. Every report drawn from opt-in-gated events must state that caveat.

Project-specific context

App-specific analytics knowledge (custom event names, known pipeline quirks, license-parameter semantics) belongs in the host project's CLAUDE.md, not in this skill. Keep the skill portable.

Known limitations

• security add-generic-password -w <value> (macOS) briefly exposes the password to ps on the local machine. Same-user visibility only, but a real limitation; fix requires a pty workaround.
• File-backend fallback (Windows, Linux without libsecret) stores secrets at <config-dir>/secrets.json with mode 0600. Disk-readable by the same user, which is strictly weaker than a keyring. Install libsecret-tools on Linux for a proper keyring; on Windows, prefer env vars (TELEMETRYDECK_TOKEN, TELEMETRYDECK_PASSWORD) until Credential Manager support lands.
• No funnel/retention recipes yet. Raw TQL via query - works.
• --compare prior-period for this-X intervals compares partial-to-full (e.g. this-month-to-date vs. all of last month). Call this out in reports where it matters; reach for last-Nd --compare prior-period if you want like-for-like windows.