Lumen
Docs Start free
Start free

Prompt library

Prompts that already survived production

Every prompt below runs in a real customer workspace. We have kept the exact wording, the model it targets and the answer it returns — including the parts where the agent pushes back. Copy one and adapt it to your data.

Showing all 12 prompts

Ticket triage with macro grounding

Support lumen-pro
Prompt

You are a support agent for an e-commerce brand. Read the ticket, find the closest matching macro in the `support-macros` collection, and draft a reply in that macro's voice. If no macro scores above 0.7, say so and escalate instead of guessing.

Output

Matched “Late delivery — courier delay” at 0.86.

Hi Amara — your parcel is sitting at the Rotterdam hub and the courier has it moving again tonight. I have refunded the €4.95 shipping fee and added a €10 credit for the trouble. New ETA is Thursday.

1,240 tok€0.0041

Pull request reviewer

Engineering lumen-pro
Prompt

Review this diff against the conventions in `eng-handbook`. Report only issues that would block a merge — correctness, security, and public API changes. Skip style. For each finding give the file, the line, and a one-sentence fix.

Output

Two blocking findings:

  • auth/session.go:88 — token compared with ==; use a constant-time compare.
  • api/v2/users.py:214 — removes email from the response; that is a breaking change to a public field.
3,910 tok€0.0129

Contract deviation check

Legal lumen-pro
Prompt

Compare the attached agreement to `MSA-standard-v4`. List every clause that differs materially, cite the section number of both documents, and rate the commercial risk as low, medium or high. Do not summarise clauses that match.

Output

Three material deviations.

  • High — §9.1 caps liability at 1× fees (standard §8.1: 2×).
  • Medium — §4.2 Net-60 payment terms (standard: Net-30).
  • Low — §12.4 auto-renews with 90 days' notice (standard: 60).
28,400 tok€0.0937

Text to SQL, read-only

Data lumen-pro
Prompt

Using the schema in `warehouse-ddl`, answer the question with a single Postgres query. Read-only: no INSERT, UPDATE, DELETE or DDL. Always add a LIMIT. Return the SQL first, then one sentence explaining what it counts.

Output

SELECT plan, count(*) AS churned
FROM subscriptions
WHERE cancelled_at >= now() - interval '30 days'
GROUP BY plan ORDER BY churned DESC LIMIT 50;

Counts subscriptions cancelled in the last 30 days, grouped by plan.

2,180 tok€0.0072

Pre-call account brief

Research lumen-flash
Prompt

Build a one-page brief for tomorrow's call. Pull the last four touchpoints from CRM, any open support tickets, and their public product changes in the last quarter. End with three questions worth asking. Mark anything you could not verify.

Output

Halcyon · renewal in 41 days

  • Usage up 3.2× since March; two seats short of the Scale tier.
  • Open ticket #4471 on SSO group sync — 9 days, unresolved.
  • Unverified: rumoured Series B, single blog source only.
5,600 tok€0.0056

Incident timeline from logs

Engineering lumen-pro
Prompt

From these logs and the on-call Slack thread, reconstruct the incident timeline in UTC. Separate what we observed from what we inferred. Do not name a root cause unless a log line supports it — otherwise list the top two hypotheses.

Output
  • 14:02 — p99 latency crosses 4s on checkout-api. Observed.
  • 14:06 — connection pool saturated at 200/200. Observed.
  • 14:19 — rollback of v2.31.0 restores latency. Observed.

No log line proves causation. Hypotheses: the new N+1 query in v2.31.0, or the replica failover at 13:58.

18,700 tok€0.0617

Churn reason clustering

Support lumen-flash
Prompt

Group these 340 cancellation notes into at most six themes. Name each theme in the customer's own language, not ours. Give a count and one verbatim quote per theme. Put anything that fits nowhere into "Other" rather than forcing it.

Output
  • “Never got it set up” (94) — “We bought it in March and never finished onboarding.”
  • “Too expensive for what we use” (77)
  • “Switched to a bundled tool” (61)
  • Other (23) — no shared theme.
41,200 tok€0.0412

Evidence check with citations

Research lumen-pro
Prompt

For each claim in the draft, find supporting evidence in the `papers` collection. Label each claim Supported, Contradicted, or Unsupported, and quote the sentence you relied on. Never mark a claim Supported from the abstract alone.

Output
  • Supported — claim 1. Okafor 2024, §4.2: “retrieval reduced hallucination rate from 11.4% to 3.1%.”
  • Unsupported — claim 4. Nothing in the corpus addresses multilingual recall.
  • Contradicted — claim 6. Lindqvist 2023 reports the opposite effect at scale.
62,900 tok€0.2076

Release notes from commits

Content lumen-flash
Prompt

Turn these 84 merged commits into release notes for customers, not engineers. Three sections: New, Improved, Fixed. Lead each line with the benefit. Drop refactors and dependency bumps unless they change behaviour.

Output

New

  • Share a saved view with your whole workspace in one click.

Fixed

  • CSV exports keep leading zeros in postcodes.
7,400 tok€0.0074

Metric anomaly explainer

Data lumen-pro
Prompt

Signups fell 18% week over week. Break the drop down by channel, country and device, and tell me which segment explains most of it. Check whether tracking changed before blaming demand.

Output

82% of the drop sits in one segment: paid search, Germany, mobile.

Before you treat it as demand — the consent banner shipped on the 14th and mobile events fell 71% the same hour. Likely measurement, not signups.

9,300 tok€0.0307

Security questionnaire draft

Legal lumen-pro
Prompt

Answer this vendor security questionnaire from `trust-center` and last year's completed questionnaires. Answer only what the sources support. For anything unsupported, write “Needs review” and name the person who owns that control.

Output

47 of 61 answered from source. 14 flagged.

  • Q22 encryption at rest — AES-256, KMS-managed keys. SOC 2 §4.1
  • Q39 sub-processor list — Needs review, owner: R. Sandhu.
15,800 tok€0.0521

Framework migration planner

Engineering lumen-pro
Prompt

Scan the repository and plan the upgrade in stages that each ship independently. For every stage give the files touched, the rollback, and an honest estimate. Flag anything that cannot be done incrementally.

Output

Four stages, 11 days of work.

  • Stage 1 — codemod 214 imports. Fully automatic, revert with one commit.
  • Stage 3 — not incremental. The router rewrite touches every route at once; budget a freeze.
88,400 tok€0.2917

Run any of these against your own data

Create a workspace, connect a source, paste a prompt. The free tier covers roughly 40,000 support replies a month.

Start free Read the API docs