Lighting the Way to AIExact recipes for real AI jobs

small business · ~20 min

Draft 2–3 sendable refund or complaint response paths from your real policy—without inventing refunds, timelines, or legal admissions.

Last verified: 2026-09-08

Not legal advice. Use only policy and amounts you authorize. AI does not approve refunds—you choose and send.

Top picks

Best Overall

Claude

Clear option framing; holds policy boundaries.

Open Claude

Best Free

ChatGPT Free

Fast multi-path drafts for one ticket.

Open ChatGPT

Best Power-User

Claude + second pass in ChatGPT

Draft paths; policy-drift / hostility check in the other. Final choice of path is yours—AI does not approve refunds.

Open Claude

Protocol

  1. Step 1 of 5

    Paste policy + complaint facts (5 min).

    Your published refund/return/cancellation policy (or internal decision rules), order/service facts you can share, what the customer asked for, and which paths you are willing to consider (e.g. full / partial store credit / decline with care). Do not paste full card numbers or government IDs. This recipe is the decisioned response: 2–3 full paths for one real complaint, each matching policy you pasted.

  2. Step 2 of 5

    Lock policy-only rules.

    Paste the rules first so the model cannot invent eligibility, amounts, or goodwill credits.

    Copy-paste prompt

    Help me draft 2–3 SENDABLE response options for a customer refund/complaint.
    Hard rules:
    - Use ONLY my policy and case facts. Do not invent refund eligibility, amounts, timelines, shipping labels, or goodwill credits.
    - Each option must state clearly what I am offering from my policy (or declining) and the next step.
    - No legal admissions of wrongdoing; no insults; no blaming the customer.
    - Label options (e.g. Path A Full refund / Path B Partial / Path C Decline with care) and note which policy clause each path relies on—or [NEED FACT] if the clause is missing.
    - Not legal advice. I choose and send.
    Confirm in one sentence, then wait for my pack.
  3. Step 3 of 5

    Fill pack (step 3a) and generate (step 3b).

    Paste policy and case facts, then send the generate command.

    Paste pack

    MY POLICY (paste):
    [ ]
    CASE FACTS (order/service date, product/service, issue as customer stated, what I verified):
    [ ]
    WHAT CUSTOMER ASKED FOR: [ ]
    PATHS I AM WILLING TO CONSIDER: [e.g. full refund / partial refund $X I set / store credit / remake / decline]
    AMOUNTS I AUTHORIZE IF APPLICABLE: [or NEED FACT]
    CHANNEL: [email / chat / ticket]
    VOICE: [warm / plain / formal]

    After you send the filled pack, copy the Generate command in the next box and paste it into the SAME chat so the AI runs the decision tree / produces the output.

    Generate command — copy and send this next (same chat)

    Draft 2–3 full sendable paths now. For each: subject (if email), body, policy basis, and a 3-bullet “verify before send” checklist. Mark [NEED FACT]. Do not invent policy.
  4. Step 4 of 5

    Second-model check (5 min).

    In the other tool, paste the options plus your policy:

    Second-model check

    Here are refund/complaint response options from another assistant plus my policy.
    Find: (1) offers not allowed by my policy, (2) invented amounts/timelines, (3) hostile or defensive tone, (4) missing next steps.
    Return only corrections. Not legal advice.
  5. Step 5 of 5

    Human choose and send.

    Pick one path. Process the refund in your commerce/finance system yourself—do not treat the chat draft as approval.

When not to

  • Do not invent a more generous policy than you published or can honor.
  • Chargebacks, regulated industries, or counsel-instructed matters—follow those processes.
  • Do not paste payment card PANs, CVVs, or government IDs into consumer AI tools.
  • Skip if you have no policy and no authorized amounts yet—decide those first.

Dated sources

Related recipes

Copied