Voc Source Chat

/voc-source-chat — Mine Internal Chat for Customer Voice

Your team already does VoC work. They paste customer quotes into channels named things like #wins, #churn-risk, #feature-requests and #sales-questions, where the quote gets four emoji reactions and is never seen again.

This is the highest-yield source per unit of setup effort in the whole pipeline, and it is the one every VoC tool ignores because it is not a customer-facing system.

Requires

Wiring the tools

This skill deliberately does not hardcode chat MCP tool names. They vary by server, and a skill that declares a tool name your server does not expose fails at its first step while looking like it ran.

To connect yours:

1. Add your chat MCP server to .mcp.json (see .mcp.json.example)
2. Run /mcp in Claude Code to list the tool names your server actually exposes
3. Add those names to this file's allowed-tools

If you skip step 2 and guess, you will get a skill that reads no messages and reports no error.

Execution

Step 1 — Resolve channels

Read chat_channels from your context file. Each entry:

chat_channels:
  - name: "#wins"
    signal_type: praise
    reliability: high
  - name: "#sales-questions"
    signal_type: objection
    reliability: medium

If unconfigured, ask which channels carry customer quotes. Do not scan every channel — most chat volume is internal coordination and scanning it wastes tokens and buries the signal.

Step 2 — Pull the window

Read messages from the configured window. Include thread replies: the quote is often in the parent and the crucial context in a reply three deep.

Step 3 — Separate quote from relay

This is the work, and it is where naive implementations fail. Chat contains three things:

Preserve the distinction all the way through. A paraphrase is real signal about a real conversation, but it is your colleague's words, and it must never be published as a customer quote. Downstream skills rely on this flag.

Extract quoted spans precisely. Look for text inside quotation marks, after said: / told me: / their words:, and inside blockquote formatting. When a message mixes relay and quote, keep only the quoted span as verbatim and put the rest in context.

Step 4 — Attribute

For each signal, resolve:

• Company — usually named in the message or its thread. If absent, check the channel topic and the surrounding messages before giving up.
• Who relayed it — the team member. Useful for follow-up, and useful for spotting when one enthusiastic rep is generating a disproportionate share of your "signal."
• Date — the message timestamp, not the date of the underlying conversation. Note the difference when it is knowable; chat relays often lag the call by days.

Unattributable signals still get written. They become tier U at the profiling step and are excluded from confirmation counts automatically.

Step 5 — Write

One JSON per signal to outputs/voc/sources/chat/<date>-<slug>.json:

{
  "channel": "chat",
  "chat_channel": "#wins",
  "verbatim": "we can't get the data out fast enough",
  "paraphrased": false,
  "context": "relayed after the QBR",
  "company": "Northwind Trading",
  "relayed_by": "person who posted",
  "message_date": "2026-04-22",
  "permalink": "...",
  "signal_type": "pain"
}

Report counts by channel and by signal type, plus how many were unattributable.

What this gets wrong

• Relay bias is real and large. You hear what your team found notable enough to paste. Quiet problems and boring-but-common complaints never make it to chat.
• Enthusiasm skews the sample. One rep who posts every customer comment will dominate the corpus. Check the relayed_by spread before treating volume as prevalence.
• Paraphrase drift compounds. By the time a customer's comment reaches a channel it has passed through one person's memory and framing. This is why the flag matters.
• Chat is a lagging indicator. Something posted Friday may have been said the previous week. Date arithmetic on this source is approximate.
• Reading team channels is a trust matter. Tell your team this is running. A tool that silently harvests colleagues' messages is a bad thing to be discovered doing, even when the intent is benign.