Signal Skills: research-suite
Router and map for the podcast/review-scout/ideation/signal-report skill family — explains what each skill does, when to reach for which one, and how the shared infra (render_report.py, template.html, analytics-check.md, per-skill ledgers) fits together. User-invoked only; never fires automatically.
Four skills, one shared rendering/analytics spine, each with its own state. This skill doesn't do research itself — it's the map, so you (or a teammate new to this setup) don't have to read four SKILL.md files to figure out which one to reach for.
Naming heads-up
signal-report here is a local render/report-generation skill. It
owns the canonical render_report.py/template.html that the other
three skills ship generated copies of. It was called signal-scout
before plugin version 2.0.0 and is distinct from Oren's
separate signal-scout customer-discovery product
(https://github.com/OrenSegal/signal-scout), which is a different tool.
Which skill for which ask
| You're asking | Reach for | Why not the others |
|---|---|---|
| "What's this podcast/episode actually saying" | podcast | Extracts insights from audio/transcript, not reviews or app-store data |
| "What are users saying / app reviews / store feedback" | review-scout | Same extract-don't-summarize discipline, applied to reviews across mobile/extension/web-SaaS, not audio |
| "What should we build/fix/test next" | ideation | The only one that reads across the other two's findings and produces a prioritized, cited proposal — not another findings dump |
| "I need a report rendered from a findings payload" | (shared render_report.py, lives in signal-report) | Not a research skill itself — the mechanical renderer all three call |
If a request doesn't match any row, it's not this family's job — answer directly or say so.
Shared infra map
render_report.py/template.html— canonical copies live insignal-report;podcast,review-scout,ideationship generated copies (seescripts/sync-shared-files.sh). Mechanically enforces no-em-dash, required confidence tags, and theme-token safety — not manual-review rules anymore.analytics-check.md— canonical copy lives inpodcast;review-scoutsymlinks it. Backend-agnostic cross-check contract (PostHog, Amplitude, Mixpanel, GA4), detected at runtime.ideationdoesn't need its own copy — it reads already-cross-checked findings from the other two rather than querying analytics directly.ledger.py— deliberately NOT shared. Each ofpodcast,review-scout,ideationhas its own real copy, same shape, different keys (show/episode, app/batch, idea/status). A shared symlinked copy was tried once and reverted:Path(__file__).resolve()follows the symlink back to the other skill's directory and silently writes state into the wrong skill'sconfig.json/state/. If you're tempted to dedupe this file again, don't — read the docstring in any of the three copies for the full story first.AGENTS.md— every skill in this family has one, pointing any non-Claude harness (OpenCode, Codex CLI, etc.) back atSKILL.mdas the single source of truth, and documenting the two Claude-specific fallbacks every skill needs (subagent fan-out → sequential fallback;Artifactpublish →render_report.py --open).
On proactive routing
This router is disable-model-invocation: true — it only runs when you
explicitly invoke it, and it never auto-invokes the other three skills
either. That's a deliberate choice, not an oversight: Garry Tan's
gstack suite (github.com/garrytan/gstack) takes the opposite stance —
its router skill defaults to proactively invoking downstream skills
whenever a request pattern-matches, on the reasoning that a false
positive (invoking a skill that wasn't needed) is cheaper than a false
negative (answering ad-hoc when a structured workflow existed), with a
persisted per-user toggle to opt out. If that tradeoff appeals more than
this family's current "only when asked" default, the pattern to copy is:
give podcast/review-scout/ideation normal (non-disabled)
description frontmatter, so their own descriptions do the matching, and
skip a central proactive router entirely — Claude Code already invokes
model-invoked skills off their descriptions without one.
Completion criteria for this skill
Answering "which skill / why / how do they connect" from this file alone,
with no need to open any of the four SKILL.md files first, unless the
user wants the actual step-by-step procedure — at which point, name the
specific skill and stop; don't reproduce its steps here.
- Downloads
- Security audit
- Pass
- Last updated
- 1w ago
- Current version
- v1.0.0
- License
- MIT-0
