AGENT SKILL

product-pulse

Human-started command: it runs only when the human's message begins with `/ak:product-pulse`. On any other request do not load or follow it; tell the human to type that command. Reads the product's configured signal sources over a lookback window and publishes a one-page pulse report (headlines, usage, system performance, followups) with deltas against the prior equal window and the strategy's key metrics carried forward. Use when a human wants a read on how the product is doing and runs /ak:product-pulse. Not a dashboard, not permission to change the roadmap, and never a write to any product system.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
3

01 前言 Front Matter

name
product-pulse
license
MIT
metadata
{ "ak_catalog_id": "product-pulse" }
description
Human-started command: it runs only when the human's message begins with `/ak:product-pulse`. On any other request do not load or follow it; tell the human to type that command. Reads the product's configured signal sources over a lookback window and publishes a one-page pulse report (headlines, usage, system performance, followups) with deltas against the prior equal window and the strategy's key metrics carried forward. Use when a human wants a read on how the product is doing and runs /ak:product-pulse. Not a dashboard, not permission to change the roadmap, and never a write to any product system.

02 资源文件 4

03 SKILL.md 正文

Inspect changes and signals against existing product anchors. A report, not permission to change the roadmap.

When to use

  • A human wants a single-page read of the last day, week or launch hour.
  • A launch just went out and a human wants the first hour's numbers in one place.
  • A human wants the strategy's key metrics read against what the product actually did.

Not for

  • Changing the roadmap, the strategy or the backlog. Followups stay in the report; acting on one is a separate human-started run.
  • Replacing a dashboard or investigating deeply. Native tools still do that.
  • A shipping log. Shipped work lives in the tracker and the commit history.
  • Any write to the product, its database or an external system. Every source is read-only.

Authority

Authority: explicit. A human starts this skill with /ak:product-pulse. It may start model-invoked skills only. It holds no authority over the product, the tracker or the strategy: its only writes are its own configuration page and its reports, both in the knowledgebase.

Inputs

  • The lookback window the human gave (1h, 24h, 7d, 30d). Absent: the configured default, else 24h. Unparseable: ask.
  • The pulse configuration: a concept page scoped to the product, read with the knowledgebase adapter's readContext. Absent, or the human asked to set up or reconfigure: run setup first. Its keys are in the run guide.
  • The product's strategy page, through readContext, for its key metrics. Absent: the report goes without a strategy-metrics line and says so; no anchor is invented.
  • The signal sources the configuration names, reached through the host's data tools in read-only mode: product analytics, application tracing, payments if configured, and a database only when enabled and read-only.
  • An unreachable knowledgebase is needs-input. A configured source the host cannot reach is reported as no data, never estimated.

Workflow

  1. Check how this run was started, before any other step and before any tool call. It is started only when the human's message begins with /ak:product-pulse; no grant starts it. A request in prose is not a start, even when it names this skill or the command. Otherwise, stop here: make no tool call, say that this command is human-started, and give the human the line to type, /ak:product-pulse and their request.
  2. Resolve the window. Apply a 15-minute trailing buffer to its upper bound to absorb ingestion lag: a 24h window reads [now - 24h - 15m, now - 15m]. The prior window is the same length, immediately before.
  3. Load or set up the configuration. With none on record, follow the setup guide: seed from the strategy page, interview one question at a time, refuse read-write database credentials, show the configuration with one edit round, and publish it.
  4. Query, per the run guide: analytics, tracing and payments in parallel; database queries after, one at a time, tight and scoped, skipped with a note when one would be expensive. Every tool call is read-only; a tool's write mode is never used.
  5. Resolve the strategy metrics. For each key metric: omit it if excluded; mark it no data (instrumentation pending) if pending; otherwise read it from its configured source, or the default analytics source marked (default source). A query that returns nothing is listed as no data, never dropped.
  6. Assemble the report from the report template: Headlines, Usage, System performance, Followups, and a footer naming the source windows and the buffer. Thirty to forty lines. Thin sections stay thin.
  7. Scrub and check. Run the template's checklist. No email, account identifier or message content survives into the report.
  8. Publish the report as a new revision of the product's pulse page, and read it back.
  9. Surface. Post the Headlines verbatim, the top Followup if it looks urgent, and where the report lives. Never paste the whole report into chat.
  10. Mention a reminder, lightly. At setup, suggest once that the human keep a recurring reminder to run /ak:product-pulse; on the third or later ad hoc run, mention it once. A schedule may only remind the human; it never starts a run, and this skill creates no reminder or schedule.

Hard gates

Gate: read-only everywhere. No write to the product, its database or any external system; a read-write database credential is refused at setup and never used if offered later.

Gate: no personal data in the report. No email, account identifier or message content, in any section, the quality sample included.

Gate: no roadmap, strategy or tracker change. A request to act on the numbers is named as outside this skill and left for a separate human-started run.

Gate: no invented numbers. A delta needs the prior equal window; without one the delta is omitted. A source with no answer reads no data.

The thought Why it is wrong Do this instead
"Give me write access and I can fix the counts at the source." A pulse that mutates what it measures is not a read, and the credential outlives the run. Refuse; offer a read-only connection or no database at all.
"Their emails would make the error list easier to debug." The report is durable and shared; personal data in it is a leak. Describe the error without identifiers; the native tool has the details.
"Signups dropped 30%, so I'll move onboarding to the top of the roadmap." A report is not permission to change the roadmap. Put it in Followups and stop.
"No prior window, but it's probably up about 10%." An estimated delta reads as a measurement. Omit the delta.
"This error count is high, I'll flag it red." No thresholds were configured; a label is a judgment the reader did not ask for. Present the number and let the reader judge.

Outputs

  • A pulse report: a new revision of the product's pulse page, a concept page scoped to the product, published through the knowledgebase adapter's publishArtifact under a kb-document placement. Its revision history is the timeline of past pulses.
  • On setup, the pulse configuration, a concept page scoped to the product, published the same way.
  • The knowledgebase resolves both locations; this skill writes nothing into the working repository (ruling central-kb-owns-project-artifacts).
  • In chat: the Headlines, the top Followup if urgent, and where the report lives.

Side effects

external-fetch, scratch-write, kb-draft, kb-publish. No workspace-write, no tracker-write.

kb-publish is a remote side effect. Its idempotency key derives from the run, the operation, the page's stable remote identity and the artifact's hash; the read-back is the record the publish returns, read before the write and confirmed after it (adapters/runner-contract/CONTRACT.md, "Idempotency"). A publish whose read-back cannot be performed is failed, never complete.

Stop conditions

  • complete — the report is published and read back, and its headlines are in chat.
  • needs-input — not started with /ak:product-pulse, the knowledgebase cannot be read or written, no signal source is configured or reachable, the window is unparseable, or a setup answer is awaited.
  • failed — a publish's read-back cannot be performed.
  • cancelled — the human withdrew; nothing further is published.

Limits

  • Report length: thirty to forty lines (guidance).
  • Followups: three to five, each specific enough to act on (guidance).
  • Top errors: five, or the configured count (guidance).
  • Quality sample, when enabled: up to ten sessions (guidance).
  • Trailing buffer on the window: fifteen minutes (guidance).

04 同仓库更多 Skills