Overview

Use this skill as Kelly's voice-of-customer desk. It aggregates raw user feedback from every channel (support email, Discord, Slack, X replies, app-store reviews, in-app surveys, interviews) into one Busabase App-in-Skill dashboard, where the agent's dedupe/clustering work becomes feature requests with frequency and user weight, and agent-proposed roadmap changes wait in a decision queue for Kelly's verdict.

Division of labor: the skill (scripts + agent) collects, clusters, and drafts; the app is where Kelly triages feedback and decides on proposals; approved roadmap changes are exported/executed by the agent outside the app (updating a roadmap doc or changelog, handing decline replies to kelly-messenger or kelly-email). Kelly Feedback sits downstream of kelly-email, kelly-messenger, and kelly-social: those skills' agents hand feedback payloads to scripts/ingest_feedback.mjs.

Default behavior is AirApp-first. Unless the user explicitly asks only for explanation, ingest whatever payloads are available and give the user the clickable AirApp URL (or the local preview URL when local preview is explicitly requested). Use chat-only mode only when the user says "纯聊天", "chat only", "不要打开 UI", or similar; in that mode present proposals as numbered items (Proposal #1, #2, ...) and record verdicts by asking the agent to run the same writes busabase-sdk-side that the app would.

This skill is an implementation of the App-in-Skill pattern — a Codex/agent skill paired with a small companion UI for review and approval. See the spec paper: https://mr-kelly.github.io/research/app-in-skill-specification-for-pairing-agent-skills-with-a-local-companion-ui.pdf.

Mandatory Dependencies

Read and follow $kelly-app-skill-creator for product behavior, visual quality, responsive layout, and the complete canonical content/kelly-feedback-app/ artifact.
Read and follow $busabase for connection, target Space, node discovery, ChangeRequests, review, and merge behavior.
Read and follow $busabase-app-creator for resource modeling, AirApp runtime limits, security, validation, and deployment.

If a dependency is unavailable, preserve this skill's local artifact and product contracts, stop before the unavailable Busabase operation, and report the exact missing dependency. Do not invent a second data backend.

Boundary

The AirApp reads Busabase records and writes only feedback triage, request effort estimates, and roadmap-proposal verdicts. It must not call feedback platforms, post replies, publish changelogs, or mutate remote systems.
Any outbound side effect — replying to a user, publishing a changelog note, editing a public roadmap — is approval-required and executed by the agent via other skills (kelly-messenger, kelly-email, docs edits) only after the matching proposal is approved and scripts/execute_decisions.mjs --apply marks it handoff_ready. scripts/execute_decisions.mjs never sends anything itself.
Own-community data only: ingest feedback addressed to Kelly's own products from Kelly's own channels and accounts via scripts/ingest_feedback.mjs. Do not scrape third-party communities or collect data about other companies' users.
Treat feedback as user PII-adjacent. Never commit real tokens, raw platform exports, or Busabase credentials.

Busabase Resources

Eight Bases under one application Folder (kelly-feedback), declared in content/kelly-feedback-app/app/js/config.js and the generated template sidecars under content/:

products: products feedback can be about (id, display name, tagline).
sources: feedback channels — email, Discord, Slack, X, app-store, survey, interview — with collection method, env-var names for tokens (never values), freshness, and item count.
feedback: raw feedback items normalized from every channel, one row per item (user context, text, sentiment, triage, linked request-id).
requests: clustered feature requests (title, product, status, trend, effort estimate, problem statement, spec summary, representative feedback ids, decision history). frequency/weighted-score are not stored — they are always recomputed client-side from feedback by recomputeDerived().
roadmap: roadmap lane items (now/next/later), read-only in the app; changed only via approved proposals executed by scripts/execute_decisions.mjs.
proposals: the decision queue — agent-proposed promote/decline/merge/publish-changelog changes with reason, evidence, an editable draft, the human verdict (status/review-note/decided-at), written directly onto the proposal record.
sync-log: append-only history of ingest/cluster/execute runs.
settings: one row (record-id: "config") with plan-weight scoring and roadmap lane names.

Resources provision lazily through an idempotent Busabase ChangeRequest the first time the app runs in a Space; see references/feedback-schema.md for exact field shapes. Request frequency/weighted_score and every snapshot metric are recomputed client-side from feedback/requests/proposals on every read via recomputeDerived() — never stored, so numbers always agree after any merge.

First Run And Onboarding

On invocation, check the sources Base. If empty, guide setup before ingesting real feedback.

Onboarding asks, turn by turn (non-secret setup details only):

Products: id, display name, one-line tagline for each product feedback can be about.
Sources: which channels exist (email/discord/slack/x/appstore/survey/interview), how each is collected (sibling-skill handoff, export, manual notes), and which env var names hold any tokens.
Scoring weights: plan weights (e.g. free=1, pro=3, team=5), default weight, recency half-life.
Roadmap lanes (default Now / Next / Later).

Never ask the user to paste secret values into chat. Secrets belong only in local env files. Register products/sources by including them in the first scripts/ingest_feedback.mjs payload's optional products[] field and required source field (see references/feedback-schema.md); scoring weights/roadmap lanes are written directly into the settings Base's single record-id: "config" row via busabase-sdk.

Local App

Default behavior is AirApp-first — give the user the clickable AirApp URL. Start pnpm --dir content/kelly-feedback-app dev only when local preview/debugging is explicitly requested.

Required app views (hash routes):

#/overview: VoC command desk — human-attention panel (roadmap decisions waiting, new uncategorized feedback, requests needing info), feedback-this-week inflow by channel with platform badges, sentiment split bars (inline SVG), top clusters by momentum, and freshness per source.
#/inbox and #/inbox/<feedback_id>: raw feedback stream. Rows show channel badge, user handle, product, one-line preview, sentiment, cluster link, and triage state (new/clustered/ignored/insight). Detail shows the full text, user context (plan, tenure, revenue weight), source permalink, linked request, agent note, and triage buttons (assign to request / ignore / mark insight) that write directly onto the feedback record through busabase-sdk.
#/requests and #/requests/<request_id>: clustered feature requests with title, product, frequency, weighted score (frequency × user revenue weight), trend arrow, status (candidate/roadmap/declined/needs_info), and linked feedback count. Detail shows the agent-drafted problem statement and proposed spec summary, representative quotes, all linked feedback, an editable effort-estimate field, and decision history.
#/roadmap: decision queue of agent-proposed roadmap changes (promote to Now/Next/Later, decline with drafted reply, merge duplicates) with reason, evidence, editable draft, a Review note textarea, decision buttons (Approve / Request changes / Block), stable refs (Proposal #1), and standard workflow states (needs_review / changes_requested / approved / done / blocked) written directly onto the proposal record. The current roadmap columns (Now / Next / Later) render read-only below the queue.
#/settings: sanitized config summary — products, sources (channel + collection method), scoring weights, env readiness booleans, onboarding state, and recent sync log. Never expose secret values.

Demo mode:

?demo=overview, ?demo=inbox, ?demo=requests, and ?demo=roadmap select named mock scenes.
?demo=detail opens a request detail with representative quotes.
lang=en or lang=zh forces UI chrome language for screenshots.
Demo mode never reads or writes Busabase; demo decisions are in-memory only.

UI language: support English and Chinese chrome with Auto default (persisted override in the sidebar). Keep user quotes, handles, product names, and imported data in their original language.

File Contract

Read references/feedback-schema.md before editing the app, scripts, or any Busabase field shape.

Ingestion Workflow

scripts/ingest_feedback.mjs is the single write path for raw feedback. Anyone with a payload — sibling skills' agents, platform exports, or manual notes — hands feedback to Kelly Feedback the same way:

Write a payload JSON file (shape in references/feedback-schema.md): a source block (source_id, channel, name, collection) plus items[] with stable external_ids, text, user context, and timestamps. Optionally include products[] to register/update product catalog entries.
Run node skills/kelly-feedback/scripts/ingest_feedback.mjs <payload.json> [more.json ...] --apply.
The script validates, dedupes by fb-<source_id>-<external_id> (idempotent re-ingest), upserts the source/products, and appends a sync-log entry — all directly against Busabase. Omit --apply first to see a dry-run summary.

Typical handoffs: kelly-email exports support threads mentioning features; kelly-messenger exports Discord/Slack community posts; kelly-social exports X replies. Those skills' agents build the payload from their own data; this skill never reads their private files directly. New items land with triage: "new" and appear in the Inbox and the human-attention panel.

Clustering Workflow

Clustering is LLM work done by the agent; scripts/apply_clusters.mjs is the deterministic write path.

Read feedback from Busabase and pick out triage: "new" items (and any requests needing re-scoring).
As the agent, dedupe and cluster: group items expressing the same underlying need, draft or update request records (title, problem statement, proposed spec summary, representative feedback ids, trend), and decide non-request items (ignored for spam, insight for bugs/patterns worth routing elsewhere).
Write a cluster-assignment payload (shape in references/feedback-schema.md) and run node skills/kelly-feedback/scripts/apply_clusters.mjs <assignments.json> --apply. The script validates ids, upserts request drafts, links feedback, and logs the run. Request frequency/weighted_score recompute automatically client-side — the script never writes them.
When clusters warrant action, draft proposals directly into the proposals Base (promote/decline/merge/changelog) with reason, evidence, and an editable public draft, then send Kelly to #/roadmap.

Roadmap Decision Workflow

Kelly reviews #/roadmap: edits drafts, writes review notes, and clicks Approve / Request changes / Block. The app writes the verdict directly onto the proposal record (status/review-note/draft/decided-at) through busabase-sdk — never a separate decisions file. From a standalone local preview the write merges immediately (trusted operator); from the deployed AirApp it creates a pending ChangeRequest for the trusted process to merge.
Before executing, run node skills/kelly-feedback/scripts/execute_decisions.mjs (dry-run) to see the plan for every approved proposal: update_roadmap (target lane) and merge_requests are LOCAL operations; publish_changelog_note (draft id) and send_decline_reply (handoff to kelly-messenger/kelly-email) are always handoff_ready — this script never publishes a changelog, edits a roadmap doc, or sends a reply itself.
Show Kelly the dry-run summary. After confirmation, run with --apply: LOCAL operations (roadmap lanes, merges) are applied directly to Busabase and the proposal's status is set done; outbound operations remain handoff_ready only.
Execute handoff_ready operations as the agent via the appropriate skill (send the decline reply through kelly-messenger/kelly-email, update the changelog/roadmap document), then log the outcome as a sync-log entry.
For a proposal moved to changes_requested, revise the draft per the review-note and write it back to needs_review directly via busabase-sdk.

Safety Defaults

Treat every outbound message (decline replies, changelog posts, roadmap publications) as approval-required, one proposal at a time.
Never invent feedback or inflate counts; frequency and weighted score must derive from real linked items via recomputeDerived(), never be written directly.
Keep stored content minimal: trimmed feedback text and safe permalinks, not raw platform API responses or attachments.
Use stable ids everywhere (external_id dedupe keys, request ids, proposal ids) so re-ingest and re-execution are idempotent.
scripts/execute_decisions.mjs performs no external side effect ever — it never sends a reply or publishes a changelog itself, even with --apply.

Useful Commands

node skills/kelly-feedback/scripts/ingest_feedback.mjs payload.json --apply
node skills/kelly-feedback/scripts/apply_clusters.mjs assignments.json --apply
node skills/kelly-feedback/scripts/execute_decisions.mjs
node skills/kelly-feedback/scripts/execute_decisions.mjs --apply
pnpm --dir skills/kelly-feedback/content/kelly-feedback-app dev

In normal use, invoke /kelly-feedback, let the skill ingest/cluster whatever is available, and open the AirApp.

kelly-feedback

Kelly Feedback

App UI Screenshots