Five messages. One clear queue.

Before you start

• A Claude conversation you can access. Haiku 5.5 is optional and account-dependent; use an available model if it is not offered. Naming a model does not select it.
• Start with the five fictional messages in the prompt. For your own batch, use unique IDs and remove account numbers, credentials and unrelated personal details before pasting.
• This labels a small batch; it does not connect to an inbox, send replies or measure quota. INMC monitors compatible Claude Code plan readings, not Claude web-chat activity.

How to use it

1. Copy the complete prompt and run its fictional batch once. No API key, plugin or inbox access is required.
2. Read the four category definitions and the rule for mixed or unclear requests. Keep the IDs so a person can trace every row.
3. Compare the output with the expected labels below. Wording of short evidence quotes may vary; the categories and review flags should agree.
4. Replace only the messages block with up to five of your own messages and run it. Review uncertain rows before taking any customer action.
5. Stop when the IDs, labels and counters pass. If one row fails, correct only that row with the missing context; do not expand the whole task.

ILLUSTRATIVE EXAMPLE

Make it concrete.

Fictional input, supplied inside the copyable prompt. S-01 reports duplicate charges; S-02 reports sign-in trouble; S-03 reports a failed CSV export; S-04 requests a dark theme; S-05 combines a charge issue and sign-in trouble. No real customer case or measured saving is claimed.

Expected output

• S-01 → billing; evidence: “two charges”; review_needed: no.
• S-02 → access; evidence: “cannot sign in”; review_needed: no.
• S-03 → bug; evidence: “Export failed”; review_needed: no. This is a reported failure, not a verified root cause.
• S-04 → other; evidence: “dark theme”; review_needed: yes. Feature requests are outside the three defined issue categories.
• S-05 → other; evidence: “charged twice and cannot sign in”; review_needed: yes. The message spans billing and access.
• Final counters: processed=5 and review_needed=2.

Check the result

• All five original IDs appear once, in order; no extra message is invented.
• Only billing, access, bug or other appear. S-04 and S-05 are marked for review.
• Each evidence quote is present in its source message and contains no more than eight words.
• The two counters match the rows. No reply, refund, account action or hidden model switch is claimed.

If the result is off

• An ID is absent or duplicated: fix the input IDs before rerunning; do not let the model invent an identifier.
• A mixed request receives a confident single category: point out the two issues and request a correction of that ID only, using the existing other/review rule.
• A feature request is labeled bug: restate that bug requires a reported feature failure. Review the correction before using the queue.
• Haiku 5.5 is not in your model picker: use a model available to your account. This guide makes no promise about plan limits, speed or cost.