Skip to content

Gets data

Message mining

Reads your reviews, call notes and tickets line by line and counts the phrases buyers repeat, tagged by pain, outcome, objection and trigger.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick message-mining.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Message mining
Needs
Buyer text (reviews, call notes or transcripts, tickets, survey answers)
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedCreative
SkillCreative and conversion

Message mining

Example output

Mine what buyers repeat
Buyer phraseTagBuyers
security review stalls the dealTrigger9
questionnaires eat our weekPain12
answers go out of dateObjection5

Reads your reviews, call notes and tickets line by line and counts the phrases buyers repeat, tagged by pain, outcome, objection and trigger.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

message-mining/SKILL.md

In the folder

  • message-mining/SKILL.md
  • message-mining/scripts/phrase_counts.py
---name: message-miningdescription: >-  Reads reviews, sales call notes, support tickets and survey answers line by line, pulls the phrases buyers  repeat, tags each as a pain, outcome, objection or trigger, and counts how many different buyers said it. Returns  a phrase bank in the buyers' own words with anonymous quotes, the alternatives they name, and the gap between  their words and the company's messaging. Use it before building a messaging house, rewriting a landing page or  briefing ads. Not for scoring a finished page (use landing-page-critic) or testing a message with buyers (use  message-test).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Uses a connected review site, CRM,  call recorder or help desk if the agent has one, otherwise exports or pasted text. Python 3 is optional, for  counting phrases.metadata:  author: SHAKEUP  version: "1.0.0"  stage: get  plugin: creative-conversion  homepage: https://www.shakeuphq.com/skills/message-mining--- # Message mining Finds the words buyers already use for their problem, counted per buyer, so pages and ads can say it back to them. ## Purpose - One job: turn raw buyer text into a counted, tagged phrase bank.- Stage: get. It reads data in and structures it, and ends in a coverage verdict, not a recommendation.- It never writes headlines or messaging (that is `messaging-house` and the skills after it), never invents or paraphrases a quote, and never names a reviewer or their company unless the profile records the quote as cleared. ## Use when / not | Situation | Use || --- | --- || Before building or rewriting the messaging house | this skill, then `messaging-house` || A landing page or ad set needs the buyer's words | this skill, then `landing-page-critic` or `creative-brief` || A pile of reviews, call notes or tickets and a question about what buyers say | this skill || Checking whether a finished message lands with real buyers | `message-test` || Finding the situations that start a search | `buying-moments` (it reads this skill's trigger phrases) || Scoring a published page | `landing-page-critic` | ## Inputs An item is one review, one call, one ticket, one survey answer or one interview. Every count in this skill is items, never mentions. | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || Buyer text: reviews, call notes or transcripts, tickets, survey or interview answers | yes | a connected tool (review site, CRM notes, call recorder, help desk); an export the reader uploads (CSV or text); text pasted into the chat | ask once; with none, `BLOCKED` || Per item: source, and if known the role and deal stage (won, lost, open, customer) | no | columns in the export; a header line per item; the reader's note | treat all items as one source; coverage `partial` || Competitor reviews | no | the review site; pasted text | mine own sources only; say so in open items || The company's current messaging | no | `## Messaging house` and `### Words to use` in the profile; the homepage text | skip the language gap (step 8); coverage `partial` | ## Profile read From `company-profile.md`: `## Company` (brand terms, to keep brand mentions out of the phrases), `### Buying group` (roles to tag items with), `## Messaging house`, `### Words to use`, `## Claims and proof` (a customer quote is cleared only when a `Claim:` line holds it with a source), `## Competitors` (names to catch as alternatives), and every `## Learnings` entry for `message-mining` with status `noted`, `fact` or `accepted`. No profile: ask for the brand name and the product category in one message, so brand mentions can be removed and off-topic items spotted. Say that `company-profile` saves this step next time. Treat every quote as not cleared. ## Steps 1. **Apply learnings.** A learning that changes a rule (a source to down-weight, a threshold) overrides the default below. Name it in the output when it fires. 2. **Strip personal data first.** Before reading for meaning, remove names, emails, phone numbers, handles, signatures and the reviewer's company name from every item. Keep the role. Drop role details that identify one person together (role plus company plus city). Count what you removed. 3. **Check that it is buyer language.**    | Material | Do this |   | --- | --- |   | Reviews, call transcripts, buyer emails, tickets, survey or interview answers | mine it |   | A sales rep's summary of a call ("prospect worried about price") | mine it, source `rep summary`; the words are the rep's, so never quote them as the buyer's |   | Competitor reviews | mine as their own group: they show pains with the alternatives |   | The company's own site, deck, ads, emails or an AI summary of buyers | don't mine: it is the company's words. If nothing else was given, `BLOCKED` | 4. **List the items.** One row per item: id, source, role, stage group, date. Stage groups: `pre-sale` (reviews, sales calls, lost-deal notes, prospect surveys) and `post-sale` (tickets, churn notes, customer interviews). Drop items under 5 words (too short to hold a phrase and its context) and exact duplicates (a review posted twice is 1 buyer). Put items older than 24 months in an `older` group and keep them out of the counts (the product and the market have moved; an old pain may be solved). 5. **Count repeated phrases.** Minimum items for a phrase to count as repeated: 2 when there are under 50 items; 3% of items, at least 3, from 50 up (a large pile throws up incidental pairs, so the bar scales with it). Run:    ```   python3 scripts/phrase_counts.py --file items.txt --brand "<brand term>" [--brand "<product name>"]   ```    Items in the file are separated by a line with only `---`; an item can open with `[source: review; role: CMO; stage: won]`. Without Python, do it by hand: in each item, mark phrases of 2 to 5 words inside one sentence that carry a problem, a result, a doubt or an event; tally 1 per item, not per mention; drop phrases with the brand name; keep phrases at or above the minimum. Say which way you counted. 6. **Read line by line for what counting misses.** Buyers say the same thing in different words. Read every item without summarizing, group paraphrases into themes, label each theme with its most common verbatim phrase, and count items per theme. Keep up to 5 `standout` lines: said once, but vivid and specific (a picture, a number, a moment). A standout is never presented as a pattern. 7. **Tag each phrase or theme.** First match wins:    | The phrase says | Tag |   | --- | --- |   | what happened that started the search (an event, a person asking, a date, a change) | `trigger` |   | a doubt, a risk, a cost or a reason to wait or say no | `objection` |   | what is better after (a result, time saved, something they can now do) | `outcome` |   | what was wrong before (the problem, the workaround, the feeling) | `pain` |   | none of these | untagged: keep in the appendix only |    Triggers and objections win ties because they are rarer and decide more: "after the board asked for pipeline by channel" is a trigger even though it implies a pain. When the same words carry opposite meanings in different items ("we finally trust the report" and "we don't trust the report"), split them into 2 phrases by the meaning of each sentence and tag each; never let a pain count toward an outcome. The script's `tag_hint` is a hint, not the tag. Also list every `alternative` buyers name (a competitor, a spreadsheet, the CRM's own reports, doing nothing) with its item count. 8. **Weigh each phrase.**    | Condition | Strength |   | --- | --- |   | At or above the minimum, in 2 or more source types | `strong` |   | At or above the minimum, in 1 source type | `medium` |   | Below the minimum, or a standout | `weak` (keep, labelled) |    Then apply, in order:   - A phrase found only in `post-sale` items is labelled `post-sale`: use it for onboarding, retention and expansion copy, not for acquisition headlines, unless it also appears pre-sale (tickets describe using the product, not choosing it).   - A phrase found only in rep summaries drops one strength level (the rep chose the words).   - Learnings that down-weight a source apply here. 9. **Find the language gap.** Compare the strong and medium phrases with `## Messaging house` and `### Words to use`:    | Status | When |   | --- | --- |   | `match` | the company already says it in the buyer's words |   | `different words` | the company says the same thing in its own words (show both) |   | `missing` | buyers say it in 3 or more items and no pillar covers it |   | `unused` | a pillar no buyer phrase supports: it may matter only to the company | 10. **Pick quotes.** For each strong or medium phrase, pick 1 or 2 verbatim lines. Format: `"<quote>" (<role or buyer>, <source>, <month>)`. Cut with `...` and remove names; change nothing else inside quotation marks. Name a person or company only when the quote is a cleared claim in `## Claims and proof`, and then cite that claim's source. If the reader asks for names on uncleared quotes, keep them anonymous and list the quotes to clear in open items. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No buyer text, or the only material is the company's own words || `THIN` | Fewer than 20 items after step 4, or fewer than 3 phrases reach the minimum. Under 20 items, 2 matching phrases can be chance; the bank is a list of quotes to read, not a pattern to build on || `MINED` | 20 or more items and 3 or more phrases at or above the minimum | Before the verdict, check: every count is items, not mentions; no name, email or reviewer company is left in a quote; no quote was reworded; every phrase has a tag and a strength. Fix and recheck once. Named stop: `BLOCKED` asks for buyer text and lists the sources that would work. No other stops. A tag with no phrases (often objections, when only happy reviews were given) is an open item that names the source to add (lost-deal notes, churn notes). ## Learn Append to `## Learnings` in the profile: - A phrase from the bank was used in a headline or ad and a result came back (a test, a click-through change, demo rate from the reader's data): a `fact` with the result and its source.- The reader says a source misleads (tickets dominated by one large customer, reviews gathered by an incentive): a `proposal` to exclude or down-weight that source.- The reader moves a phrase to another tag or merges 2 themes: a `fact` for this company's bank.- A quote is cleared for named use: ask the reader to add it to `## Claims and proof` with `company-profile`, and record a `fact`. Example: `- 2026-10-24 | message-mining | proposal | Down-weight tickets from accounts over 1,000 seats | Why: 40% of tickets came from 2 accounts (reader's help desk export, Oct 2026) | Status: proposed` ## Output ```Verdict: <MINED | THIN | BLOCKED> · Coverage: <complete | partial (what's missing) | blocked> · Counting: <script | by hand>Items read: <n> (<n> per source) · Dropped: <n> (<reasons>) · Older than 24 months: <n> · Personal data removed: <counts>Minimum items for a repeat: <n> · Learnings applied: <list or none> Phrase bank| Tag | Buyer phrase | Items | Share of items | Sources | Roles | Strength | Example (anonymous) |(grouped: trigger, pain, objection, outcome; strongest first in each) Alternatives named| Alternative | Items | What buyers say about it (verbatim) | Language gap| Pillar or word we use | Buyer phrase | Status | Standout lines (each said once)Open: <tags with no phrases and the source to add; quotes to clear; sources not mined>``` Then the handoff block. ## Hand off ```handoffskill: message-miningversion: 1.0.0verdict: MINEDcoverage: partial (no lost-deal notes, so objections are thin)for: <product, segment or page the mining was for>outputs:  items: <n>  top_trigger: "<phrase>" (<n> items)  top_pain: "<phrase>" (<n> items)  top_objection: "<phrase>" (<n> items)  top_outcome: "<phrase>" (<n> items)  alternatives: [<alternative> (<n>), <alternative> (<n>)]  language_gap: ["buyers say '<phrase>', we say '<phrase>'"]next: [messaging-house, landing-page-critic, creative-brief]open: [lost-deal notes for objections]``` `messaging-house` reads the phrases and the gap; `landing-page-critic` reads `top_pain`, `top_objection` and `top_outcome` to check the page speaks to them; `creative-brief` and `ad-themes` read the triggers as hooks. For `THIN`, `next` still lists them, and `open` says the bank is directional. ## Failure modes | Problem | Do this || --- | --- || Every review is 5 stars and short | Mine them, but say objections are missing; ask for lost-deal notes or churn notes || One buyer wrote a third of the material (a long transcript) | Count that buyer once per phrase, as with any item; say their share of the words || Text in another language | Mine it in that language; give the phrase and a plain translation; never merge across languages || The export has no source column | Treat all items as one source; strength can't be `strong`; coverage `partial` || The reader wants quotes with names for a page | Keep them anonymous; list the quotes and ask the reader to clear them through `company-profile` || A phrase is the product's own feature name repeated by buyers | Keep it out of pains; it can be an outcome only with a result next to it | ## Influenced by - Joanna Wiebe's message mining (Copyhackers): read buyer text line by line, without summarizing, and write from the phrases that repeat. Described here: https://wistia.com/learn/marketing/joanna-wiebe-conversion-copywriting- What we changed: counts are per buyer with a repeat bar that scales with the pile, phrases carry a fixed 4-tag table and a strength, pre-sale and post-sale text are kept apart, the bank is compared with the company's own messaging, and quotes stay anonymous unless cleared.

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

We count per buyer with a bar that scales with the pile, so a big export can't fake a pattern. Pre-sale and post-sale words stay apart, and the bank is set beside your own messaging so the gaps show. Quotes stay anonymous unless you've cleared them.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Remove personal details first

    Names, emails, signatures and the reviewer's company come out before anything is read for meaning. The role stays, because it tells you who said it.

  2. 02

    Keep buyer words apart from yours

    Reviews, calls and tickets are mined. Your own site, deck or an AI summary of buyers is not, because those are your words.

  3. 03

    Count per buyer, not per mention

    A phrase one buyer says 5 times counts once. Its script counts the phrases that repeat, with a bar that rises as the pile grows.

  4. 04

    Read for what the count misses

    Buyers say the same thing in different words, so it reads every line, groups the paraphrases and keeps a few vivid lines said only once, labelled as such.

  5. 05

    Tag and weigh each phrase

    Each phrase is a trigger, pain, objection or outcome. Phrases found in 2 kinds of source rate strong. Anything found only in tickets is kept for onboarding copy, not headlines.

  6. 06

    Compare with your messaging

    It shows where buyers say what you say, where they use different words, and what they talk about that no pillar covers.

What you need

  • Required

    Buyer text (reviews, call notes or transcripts, tickets, survey answers)

    A connected review site, CRM, call recorder or help desk → An export you upload, as CSV or text → Text you paste

  • Optional

    Where each item came from, and the buyer's role if you know it

    Columns in your export → A short header line on each item → A note from you

  • Optional

    Reviews of your competitors

    The review site → Text you paste

  • Optional

    Your current messaging

    Your company profile → Your homepage copy

What it can return

  • MINED

    At least 20 items and 3 or more phrases that repeat across buyers.

  • THIN

    Fewer than 20 items, or almost nothing repeats. The bank is a list of quotes to read, not a pattern to build on.

  • BLOCKED

    There's no buyer text, or the only material is your own copy.

Where it stops for you

  • When there's no buyer text, or only your own copy

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells security questionnaire software. It mined 34 reviews and 28 call notes.

Verdict

MINED

Coverage: partial (no lost-deal notes, so objections are thin)

Counting: script

Items read: 59 after dropping 3 that were too short. A phrase counts as repeated at 3 buyers.

TagBuyer phraseBuyersSourcesStrengthExample
Triggersecurity review stalls the deal9Reviews, callsStrong"Our biggest deal sat in security review for 6 weeks." (Head of sales, call, Sep)
Painquestionnaires eat our week12Reviews, callsStrong"Questionnaires eat our week, every week." (Security engineer, review, Aug)
Objectionanswers go out of date5CallsMedium"What happens when our policies change and the answers go out of date?" (CISO, call, Sep)
Outcomesend it back the same day7ReviewsMedium"We send it back the same day now." (Sales engineer, review, Jul)
We sayBuyers sayStatus
Automate security questionnairesquestionnaires eat our weekDifferent words
AI-powered trust centerno matchUnused
none of our pillarssecurity review stalls the dealMissing
Alternatives named

a shared spreadsheet of past answers (11 buyers) and the security team's inbox (6 buyers).

Open

lost-deal notes, so objections can rate strong. 4 quotes are worth clearing for named use.

The handoff block the next skill reads
skill: message-mining
version: 1.0.0
verdict: MINED
coverage: partial (no lost-deal notes)
for: security questionnaire product, all segments
outputs:
  items: 59
  top_trigger: "security review stalls the deal" (9 items)
  top_pain: "questionnaires eat our week" (12 items)
  top_objection: "answers go out of date" (5 items)
  top_outcome: "send it back the same day" (7 items)
  language_gap: ["buyers say 'security review stalls the deal', no pillar covers it"]
next: [messaging-house, landing-page-critic, creative-brief]
open: [lost-deal notes for objections]

What good looks like

  • Every count is buyers, not mentions
  • Each quote is word for word, with no name unless it's cleared
  • Rep summaries are never quoted as the buyer's own words
  • A missing tag, often objections, names the source that would fill it
  • The gap with your messaging is specific enough to act on

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Why count per buyer and not per mention?

One loud review can repeat a phrase 5 times. Counting each buyer once shows what many buyers say, which is what a headline should use.

Can it use our G2 reviews?

Yes. Export them or paste them. Reviews of your competitors help too, because they show what buyers dislike about the alternatives.

Will it put reviewers' names on our quotes?

No. Quotes stay anonymous unless your company profile lists the quote as cleared, with its source. It lists the quotes worth clearing.

What does the script do?

It removes personal details, then counts 2 to 5 word phrases per item and hints at a tag for each. It's optional. Without Python, the skill does the same count by reading and says so.