clefdecide中文

Clef examples: 10 decision recipes with real results

Each recipe was run on Clef-Flash on 2026-10-02. The numbers below are the probabilities it actually returned, not illustrations. Copy a request, swap in your own state, and set your own thresholds.

1. Ticket routing

choice

Checkout has been failing for every customer for the last hour.

Question: Which team should handle this request?

Options: Billing: payments, invoices and refunds / Technical: outages, errors and configuration / Sales: plans and upgrades

Most likely: Technical: outages, errors and configuration (confidence 92.7%)

AnswerProbability
Billing: payments, invoices and refunds1.7%
Technical: outages, errors and configuration97.5%
Sales: plans and upgrades0.8%

clef-flash · 186 input tokens

How to act on it: Assign the ticket automatically when confidence is above 0.8; otherwise let a person pick from the top two teams.

See the API request
{
  "model": "clef-flash",
  "state": "Checkout has been failing for every customer for the last hour.",
  "questions": {
    "answer": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": {
        "a": "Billing: payments, invoices and refunds",
        "b": "Technical: outages, errors and configuration",
        "c": "Sales: plans and upgrades"
      }
    }
  }
}

2. Is it urgent?

noul · yes / no

Checkout has been failing for every customer for the last hour.

Question: Is this support request urgent?

Probability of yes: 0.97

AnswerProbability
Yes96.8%
No3.2%

clef-flash · 169 input tokens

How to act on it: Page the on-call engineer when the probability is above 0.9.

See the API request
{
  "model": "clef-flash",
  "state": "Checkout has been failing for every customer for the last hour.",
  "questions": {
    "answer": {
      "type": "noul",
      "instructions": "Is this support request urgent?"
    }
  }
}

3. Severity score

score

Checkout has been failing for every customer for the last hour.

Question: How severe is the customer impact?

Options: No impact / Minor / Major / Critical

Weighted score: 2.68 (confidence 44.9%)

AnswerProbability
0 · No impact1.8%
1 · Minor1.5%
2 · Major24.1%
3 · Critical72.7%

clef-flash · 186 input tokens

How to act on it: Sort the queue by the weighted score. Anything above 2.5 needs attention now.

See the API request
{
  "model": "clef-flash",
  "state": "Checkout has been failing for every customer for the last hour.",
  "questions": {
    "answer": {
      "type": "score",
      "instructions": "How severe is the customer impact?",
      "criteria": [
        "No impact",
        "Minor",
        "Major",
        "Critical"
      ]
    }
  }
}

4. Review sentiment

choice

The strap broke after two days, but support replaced it fast.

Question: What is the overall sentiment of this review?

Options: Positive / Negative / Mixed

Most likely: Positive (confidence 52.3%)

AnswerProbability
Positive81.2%
Negative4.3%
Mixed14.5%

clef-flash · 172 input tokens

How to act on it: A top answer near 50% means the text is genuinely mixed. Treat it as mixed or send it to a person instead of forcing one label.

See the API request
{
  "model": "clef-flash",
  "state": "The strap broke after two days, but support replaced it fast.",
  "questions": {
    "answer": {
      "type": "choice",
      "instructions": "What is the overall sentiment of this review?",
      "criteria": {
        "a": "Positive",
        "b": "Negative",
        "c": "Mixed"
      }
    }
  }
}

5. Product photo

noul · yes / no

A product listing photo.
Example product photo: a white ceramic mug

Question: Is the whole product visible, with nothing cut off by the frame?

Probability of yes: 0.96

AnswerProbability
Yes95.9%
No4.1%

clef-flash · 824 input tokens

How to act on it: Send photos below 0.8 to a person before they go live.

See the API request
{
  "model": "clef-flash",
  "state": "A product listing photo.",
  "questions": {
    "answer": {
      "type": "noul",
      "instructions": "Is the whole product visible, with nothing cut off by the frame?"
    }
  },
  "images": [
    "data:image/jpeg;base64,…"
  ]
}

6. Spam and scam detection

noul · yes / no

Congratulations! You won a $500 gift card. Claim it within 24 hours at bit.ly/claim-now

Question: Is this message spam or a scam?

Probability of yes: 0.81

AnswerProbability
Yes81.0%
No19.0%

clef-flash · 184 input tokens

How to act on it: Block above 0.9, send 0.5–0.9 to a review queue, deliver the rest.

See the API request
{
  "model": "clef-flash",
  "state": "Congratulations! You won a $500 gift card. Claim it within 24 hours at bit.ly/claim-now",
  "questions": {
    "answer": {
      "type": "noul",
      "instructions": "Is this message spam or a scam?"
    }
  }
}

7. Customer intent

choice

Hi, can I move my plan to annual billing and get the discount you mentioned?

Question: What does the customer want?

Options: Cancel the subscription / Change plan or billing / Report a bug / Ask a product question

Most likely: Change plan or billing (confidence 93.5%)

AnswerProbability
Cancel the subscription0.6%
Change plan or billing97.5%
Report a bug0%
Ask a product question1.4%

clef-flash · 199 input tokens

How to act on it: Route on the top option when confidence is high; otherwise show the two most likely intents to an agent.

See the API request
{
  "model": "clef-flash",
  "state": "Hi, can I move my plan to annual billing and get the discount you mentioned?",
  "questions": {
    "answer": {
      "type": "choice",
      "instructions": "What does the customer want?",
      "criteria": {
        "a": "Cancel the subscription",
        "b": "Change plan or billing",
        "c": "Report a bug",
        "d": "Ask a product question"
      }
    }
  }
}

8. Comment moderation score

score

This is the worst update ever, whoever designed it should be fired.

Question: How abusive is this comment toward a person?

Options: Not abusive / Rude / Insulting / Threatening or hateful

Weighted score: 0.76 (confidence 27.5%)

AnswerProbability
0 · Not abusive33.8%
1 · Rude58.1%
2 · Insulting6.9%
3 · Threatening or hateful1.3%

clef-flash · 197 input tokens

How to act on it: Hide comments that score 2.5 or more, flag 1.5–2.5, publish the rest.

See the API request
{
  "model": "clef-flash",
  "state": "This is the worst update ever, whoever designed it should be fired.",
  "questions": {
    "answer": {
      "type": "score",
      "instructions": "How abusive is this comment toward a person?",
      "criteria": [
        "Not abusive",
        "Rude",
        "Insulting",
        "Threatening or hateful"
      ]
    }
  }
}

9. Lead qualification

choice

We are a 40-person agency, budget approved for Q1, we need SSO and an invoice. Can we talk this week?

Question: How ready is this lead to buy?

Options: Hot: ready to buy soon / Warm: interested, needs nurturing / Cold: not a fit right now

Most likely: Hot: ready to buy soon (confidence 49.0%)

AnswerProbability
Hot: ready to buy soon79.4%
Warm: interested, needs nurturing16.9%
Cold: not a fit right now3.7%

clef-flash · 203 input tokens

How to act on it: Send hot leads to sales the same day; put warm leads in a nurture sequence.

See the API request
{
  "model": "clef-flash",
  "state": "We are a 40-person agency, budget approved for Q1, we need SSO and an invoice. Can we talk this week?",
  "questions": {
    "answer": {
      "type": "choice",
      "instructions": "How ready is this lead to buy?",
      "criteria": {
        "a": "Hot: ready to buy soon",
        "b": "Warm: interested, needs nurturing",
        "c": "Cold: not a fit right now"
      }
    }
  }
}

10. Sensitive data check

noul · yes / no

Sure, my card is 4111 1111 1111 1111 and it expires 09/28.

Question: Does this message contain payment card numbers or other sensitive personal data?

Probability of yes: 0.97

AnswerProbability
Yes96.8%
No3.2%

clef-flash · 199 input tokens

How to act on it: Redact or block before the message is stored or sent to other tools.

See the API request
{
  "model": "clef-flash",
  "state": "Sure, my card is 4111 1111 1111 1111 and it expires 09/28.",
  "questions": {
    "answer": {
      "type": "noul",
      "instructions": "Does this message contain payment card numbers or other sensitive personal data?"
    }
  }
}

Try it with your own data

Rewrite any recipe in the playground (3 free runs). For the difference between the three question types, read noul, choice and score. To call the model from code, see the Clef API guide.

FAQ

Are these results real?

Yes. Each recipe was run once on Clef-Flash on 2026-10-02, and the tables show the probabilities it returned. Running again can give slightly different numbers.

How should I pick a threshold?

Run 20–50 of your own real examples, look at where the right and wrong answers fall, and choose a line above which you act automatically. Send everything below it to a person.

Can I ask several questions at once?

Yes. One request can carry up to 64 questions that share the same state.