agentleFS
TomSzenessyTomSzenessy/AI_TEMPLATE/.agents/skills/review-mining/SKILL.md

review-mining

Skill0 starsChanged 11 days agoMITView on GitHub

Reads what a product's real users say in public reviews — your own product's feedback when there is no reference, or a reference product's reviews when you clone one — and ranks what they hate, what is missing and what is unsolved, in their own words. Every quote is real, verbatim and linked; nothing is invented. Original mode scopes the goal from VISION.md or the accepted design, asking the owner one blocking question when the product or its sources are missing. Turns the evidence into a fix plan and a positioning angle. Use when the user says "what do users hate about X", "read the reviews", "positioning", or "what should we fix".

5 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

3 rules. 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 5 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never fabricate: No invented reviews, quotes, ratings, counts, users or sources.

    line 41
  2. Distinctive:

    Reading, not scraping: Read review pages the way a person does, in the browser, and copy rows into the sheet.

    line 46
  3. Distinctive:

    No fake reviews, ever: Not for your app, not against theirs.

    line 53

“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

review-mining

Two modes, one goal: know what real users hate, in their own words, before you build. Reference mode (cloning a product you studied) reads the scope in reference/recon.md (product-recon), so the mining stays inside the slice being built. Original mode (your own product, no reference) has no reference/recon.md: scope the mining from the owner's goal instead — VISION.md or the accepted design brief, the category, the comparable products — and work under docs/research/. Before launch your product has no reviews yet: the comparables carry the run. Never invent a competitor. If the product or its sources are missing, ask the owner the one blocking question that unblocks them. When run, write the scope record (goal, category, comparables, sources, date) to docs/research/review-scope.md; create it only then, never as a placeholder. Never require product-recon, and never force a full product-kickoff when the user already brought a feedback task.

One variable picks the mode's folder, so the command keeps the same shape:

RESEARCH_DIR=reference   # reference mode; docs/research in original mode
mkdir -p "$RESEARCH_DIR"
python3 .agents/skills/review-mining/reviews.py "$RESEARCH_DIR/reviews.csv" --out "$RESEARCH_DIR/feedback.md"

The rules, which are not negotiable

  • Never fabricate. No invented reviews, quotes, ratings, counts, users or sources. If a source cannot be reached, say so and move on. If there are 14 reviews, say 14.
  • Every quote is verbatim and linked. Copied exactly from the page, with the URL of the review or thread. reviews.py drops any row without a link.
  • Reading, not scraping. Read review pages the way a person does, in the browser, and copy rows into the sheet. No scraping libraries against stores or review sites whose terms forbid it. Official public feeds and APIs are fine within their terms: Apple's customer reviews RSS feed (https://itunes.apple.com/us/rss/customerreviews/id=<APP_ID>/sortBy=mostRecent/json), the Hacker News Algolia API (hn.algolia.com/api/v1/search?query=...), Reddit's official API under its terms.
  • No fake reviews, ever. Not for your app, not against theirs. It is illegal in the US (the FTC's 2024 rule) and in many other places.
  • Reviewers are not your testimonials. Their words are research. Do not put them on your landing page.

Step 1: collect

Aim for 100+ reviews across at least three sources, recent first:

sourcewhere
App Storethe app page, Ratings and Reviews, See All; or the RSS feed above
Google Playthe listing, See all reviews, sort by newest
G2, Capterra, Trustpilotthe product's review pages, filter to 1 to 3 stars too
Redditsearch "X alternative", "switched from X", "X sucks", "X vs"
Hacker Newsthe Algolia API or site search, same queries
its own boardits public roadmap or feature-request board (Canny and similar) and the vote counts
its changelogwhat it shipped, so you do not "fix" what is already fixed

Each row in $RESEARCH_DIR/reviews.csv: source,url,date,rating,text, text copied exactly. Read the 3 and 4 star reviews too. "Love it, but..." is where the best fixes hide.

Step 2: rank

python3 .agents/skills/review-mining/reviews.py "$RESEARCH_DIR/reviews.csv" --out "$RESEARCH_DIR/feedback.md"

A quote cut for length is marked [excerpt], a rating outside 1 to 5 is reported, and the same words under different links count as different users. It sorts reviews into themes (themes.json in this folder, edit it for the product's category), weights low ratings and recent reviews higher, marks themes with fewer than 3 reviews or only one source as thin, lists every request in the users' own words, and surfaces low ratings that matched no theme. Read that last list by hand. It is often the best part.

Step 3: three lists

From $RESEARCH_DIR/feedback.md, write three ranked lists. Each item: the problem in one line, how many reviews, how many sources, one or two linked quotes. In original mode the lists are your own product's users, and they become its plan.

  1. What they hate. Complaints about things the product does.
  2. What is missing. Features people ask for by name.
  3. What is unsolved. Whole jobs or groups the product ignores ("not built for teams", "useless for therapists"). These become positioning.

Thin themes are listed as thin. Do not present three angry Reddit comments as a trend.

Step 4: the fix plan

Pick the top 5 to 8 by evidence times how cheaply you can fix them. For each: what to build or change, size (S, M, L), which step of the build loop takes it, and the evidence. Append each one to docs/product/features.csv with status no and source review-mining (evidence in acceptance) only when that product matrix exists and the fix plan is approved; parity-check lists these as improvements, not parity, and make next queues them. With no appropriate matrix or no approval, propose the issues to the owner for the WAL instead of a tracked live backlog. Pricing and billing complaints feed the angle as much as the fix plan.

Step 5: the angle

Three positioning options, each grounded in a top theme:

For {{who}} who {{hate this about the product, in plain words}},
{{your build}} {{does this instead}}.
Evidence: {{theme}}, {{n}} reviews across {{n}} sources.

Recommend one. It drives brand-sweep's name and voice. Do not put the reference product's name in your app name, ads or store listing. A factual comparison page is a legal question for a lawyer in your country.

Output

$RESEARCH_DIR/reviews.csv, $RESEARCH_DIR/feedback.md, $RESEARCH_DIR/fixes.md (three lists, fix plan, angle), rows in docs/product/features.csv when the matrix exists and the plan is approved, and a summary that states the sample size. brand-sweep only when the angle is used or the product is being rebranded — not every review request.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.