---

name: winnow version: "0.2.0" description: "Research what real people say about an app or topic in the last 30 days, ranked by authenticity not raw engagement. Covers App Store + Play Store reviews and Pantip (Thailand) — sources other tools skip." argument-hint: 'winnow Notion | winnow com.whatsapp --sources playstore | winnow ChatGPT for slack' allowed-tools: Bash, Read homepage: https://github.com/SUDARSHANCHAUDHARI/winnow license: MIT user-invocable: true

Winnow

Winnow runs a headless Python engine that pulls last-30-days content from review sources and regional platforms, scores every item for authenticity (burst / duplicate / thin-text / anonymous detectors), and ranks by authenticity-weighted engagement. Your job is to run the engine, read its structured markdown, and synthesize a short honest brief.

The engine does the retrieval, scoring, and ranking. You do the prose. Do not re-rank, re-score, or invent data — the trust labels are computed, trust them.

Run the engine

Resolve SKILL_DIR to the directory of this SKILL.md, then locate the engine. It lives in one of two known spots - bundled inside the skill (standalone installs) or two levels up (repo checkout). Check both, no open-ended search:

ENGINE=""
for cand in "$SKILL_DIR/engine/winnow.py" "$SKILL_DIR/../../engine/winnow.py"; do
  [ -f "$cand" ] && ENGINE="$cand" && break
done
python3 "$ENGINE" "<TOPIC>" --sources <sources> --limit 25

• Default sources: appstore for an app name; appstore,pantip if the user wants regional/Thailand signal; playstore when the topic is a package id (contains dots, e.g. com.whatsapp).
• App names resolve to an App Store id automatically. Package ids go to Play Store.
• One dead source never fails the run; it logs to stderr and continues.

Disambiguate ambiguous brands. When the topic is a word that also means something else (Notion the app vs the band; Apple the company vs the fruit) and you are using a keyword source (reddit, youtube, x, pantip), YOU know which sense is meant - so pass it through:

--context-exclude "song,lyrics,album,music video,band,movie,trailer"
--context-include "app,software,workspace,update,feature"

These filters apply ONLY to keyword sources (review sources are fetched by app id and are already unambiguous). Supply a tight exclude list for any topic whose name collides with an unrelated common sense. Both flags are no-ops if omitted.

Read the output

The engine prints a markdown brief: a header, optional ⚠️ corpus warnings, then ranked items. Each item carries:

• a source tag (appstore / playstore / pantip)
• N★ rating (review sources)
• trust 0.00–1.00 with a label: ✅ trusted · ⚠️ mixed · 🚩 suspicious
• flags: the authenticity signals that fired (e.g. review-burst, near-duplicate)

A corpus warning like "Low corpus trust (48%)" means the engagement is heavily gamed — say so plainly in your brief.

Synthesize the brief

Emit, in this order:

1. One-line headline of what real users are actually saying this month.
2. What's working / what's broken — 2–4 bullet points, each grounded in a cited review. Link every citation as [source](url).
3. Trust note — if any high-ranked items are flagged or corpus trust is low, call it out: "N of the top items show coordinated-push signals."
4. The receipts — pass through the engine's ranked list (or the top 5–8).

Rules, kept short on purpose:

• Never fabricate a review, rating, or number. If the engine returned nothing, say so.
• Every quote comes from the engine output. Every link is a real [text](url).
• Lead with the authenticity story — it is the whole point of this tool.
• No em-dashes in your own prose; use -.

That's it. The engine is the contract; this file is thin by design.