Update to More Good Reviews

{
  "description": "Produces a short, ranked to-do list of what to do next to improve reputation in a More Good Reviews project — unreplied negatives, reply gaps, review-volume shortfalls, outreach problems, and location-specific fixes. This is a to-do list, NOT a metrics report (use reputation-report for the dashboard) and it does NOT send anything (use review-pipeline to schedule asks). Use when the user asks what to do, how to improve, what to fix, what needs attention, or what their priorities are.",
  "included_files": [],
  "name": "reputation-actions",
  "skill_md_contents": "---\nname: reputation-actions\ndescription: Produces a short, ranked to-do list of what to do next to improve reputation in a More Good Reviews project — unreplied negatives, reply gaps, review-volume shortfalls, outreach problems, and location-specific fixes. This is a to-do list, NOT a metrics report (use reputation-report for the dashboard) and it does NOT send anything (use review-pipeline to schedule asks). Use when the user asks what to do, how to improve, what to fix, what needs attention, or what their priorities are.\n---\n\n# Reputation Actions\n\nTell the user the highest-impact things to do next for one More Good Reviews project. Output is a **ranked to-do list** — not a report, not a data dump.\n\n**Prerequisite:** Requires a More Good Reviews Project API MCP server connected. This skill calls its tools and has no other data source. If no such server is connected, say so and stop.\n\n## Core principle: actions, not analytics\n\n- Every line is something the user can **do** — reply, ask, investigate, fix, brief a team.\n- Lead each action with a verb. Include only the context needed to act (count, location, one review reference). No charts, leaderboards, or period-over-period tables.\n- Ignore reviews flagged as duplicates (`is_duplicate`); they never count toward a metric or an action.\n- Rank by urgency first, then impact. Cap the list at `max_actions` (default 7).\n- If nothing needs work, say so plainly and offer one optional maintenance item.\n\n## Read-only contract\n\nOnly call read tools: `listReviews`, `listLocations`, `listMessages`, `listCustomers`. Never mutate anything. This skill recommends and drafts; it never sends, replies, or schedules.\n\n## Step 1: Select the project\n\nEach More Good Reviews Project API MCP connection is scoped to one project. A user may have several connected at once.\n\n- If exactly one is connected, use it without prompting.\n- If more than one is connected, name them and ask which to use.\n- Use one project's tools consistently for the whole run.\n\n## Step 2: Resolve the window\n\nDefault to the **last 30 days** when the user gives nothing. Also compute an equal-length **prior** window — used internally to judge what is normal for this project, not printed.\n\n- `days` — default `30`.\n- `date_from` / `date_to` — explicit `YYYY-MM-DD`, inclusive; overrides `days`.\n- Reviews use `date_key: \"created_at\"`. Outreach uses `date_key: \"sent_at\"`.\n\n## Step 3: Gather signals (minimum needed)\n\nList endpoints are paginated; use `pagination.total` for counts and only fetch review text when naming specific reviews or diagnosing a theme.\n\n**Exclude duplicates.** Reviews flagged as duplicates must not count toward any metric or appear in any action. The `filter` parameter holds only one value (and has no \"not-duplicate\" option), so it cannot be combined with `without-replies`. Instead, read each returned review's `is_duplicate` field and drop every review where it is `true` before counting, computing coverage, or naming reviews. When relying on `pagination.total` for a count, fetch the rows and subtract flagged duplicates rather than trusting the raw total.\n\n- **Unreplied negatives** — `listReviews` with `filter: \"without-replies\"`, `score: 1`, then `score: 2`, current window, `sort_dir: \"desc\"`. Pull enough rows to name up to three.\n- **Reply gap** — `listReviews` `with-replies` vs `without-replies` over the current window; compute coverage.\n- **Volume trend** — `listReviews` totals, current vs prior window.\n- **Location regressions** — `listLocations`, then per-location totals and average score for both windows (only when multiple locations exist).\n- **Outreach health** — `listMessages` `channel: \"email\"`, current window; watch for high failure rate or unusually low volume.\n- **Ask pipeline** — `listCustomers` `filter: \"not-asked\"`, `limit: 1`; read `pagination.total` only.\n\n## Step 4: Judge against the project's own normal, then rank\n\nPrefer **learned baselines** over fixed thresholds — what matters is deviation from this project's trajectory:\n\n- **Rating** — flag a location or the project when its current-window average drops noticeably below its prior window (and below its usual level). Use `rating_alert_drop` (default `0.2` stars) only as a fallback when prior data is thin.\n- **Volume** — compare current vs prior review count; flag a clear drop. Use `review_target` only if the user supplied one.\n- **Reply coverage** — compare to the project's recent norm; if it has historically been near-complete, even a small slip matters. Use `reply_coverage_target` (default `90%`) as a fallback floor.\n- **Outreach** — flag failure rate above `failure_rate_alert` (default `5%`) or a sharp drop in sends.\n\nState briefly when you are judging against the project's own history versus a fallback default, so the reasoning is transparent.\n\nMap each surviving signal to one action; skip healthy signals and do not pad.\n\n## Step 5: Output\n\n```markdown\n# What to improve — [Project] — [date_from] to [date_to]\n\n## Do these first\n1. **[Verb] [specific target]** — [why it matters now: counts, location, or a review reference]\n2. ...\n\n## Do next\n1. **[Verb] [specific target]** — [one sentence of context]\n\n## Looking good\n- [1–3 things that need no action right now]\n\n_Snapshot: [avg rating] · [n] reviews · [reply coverage]% replied · [n] never asked — [date_from] to [date_to]_\n```\n\nRules:\n\n- **Do these first**: urgent items only (unreplied negatives, sharp rating drops, delivery failures). Max 5.\n- **Do next**: important, not urgent (volume, coverage gaps, pipeline). Max 2.\n- **Looking good**: brief; omit if everything needs work.\n- The closing **Snapshot** line is one paste-able row the user can drop into the next run for continuity (this skill keeps no memory between runs).\n- When an action is \"reply to negative reviews,\" offer a handoff: ask whether to draft those replies now. Drafting and publishing happen in a separate reply workflow, not here.\n\n## Configurable inputs\n\nAll optional; use defaults when the user gives nothing.\n\n- `days` — lookback window. Default `30`.\n- `date_from` / `date_to` — explicit window; overrides `days`.\n- `max_actions` — cap for Do-these-first plus Do-next combined. Default `7`.\n- `reply_coverage_target` — fallback floor. Default `90%`.\n- `rating_alert_drop` — fallback star drop. Default `0.2`.\n- `failure_rate_alert` — outreach failure flag. Default `5%`.\n"
}