Use cases
Support
Diagnose customer errors, confirm product behavior, hand off to engineering and keep known issues in memory.
Merlin recipes for support teams. Turn a customer error into a diagnosis with code references, confirm what the product actually does before replying, hand a verified problem to engineering as a Linear issue, draft the customer reply, keep known issues in channel memory and schedule the open-escalations digest.
How support prompts work
Support sits between the customer's report and the engineers who can explain it. The recipes below run in the support channel and the escalation channel. Each prompt is a Slack message with the sanitized evidence attached. Merlin posts progress in that thread and the result lands there: a diagnosis, a fact from the code, a Linear issue with the evidence already in it, or a reply draft.
What you need
| Requirement | Why it matters here |
|---|---|
| Merlin in the support and escalation channels | Diagnosis and handoff happen in the thread that already holds the evidence. |
| A public runtime with the product repository | "What does the product do" is answered from the code, not from memory. |
| Linear connected | Handoffs become issues with the log context and findings attached. |
| A channel whose audience may see the material | Customer logs and account details stay with the people allowed to see them. |
Support members rarely need their own runtime. Ask engineering which public runtime carries the repository.
Prompts to paste
Diagnose a customer error
Attach the sanitized log and ask for the code path and the evidence, with production off limits.
@Merlin identify the code path that produces this export timeout. Explain the likely cause and the evidence. Do not change production. If the cause is supported, draft a Linear issue for engineering.
"If the cause is supported" keeps a guess from becoming a ticket. When an engineer joins the thread, they can narrow it:
Check the v2 endpoint too and include the request size at which the behavior changes.
Confirm what the product does before replying
A customer asks whether a limit or behavior is real. Check the code instead of the help center.
@Merlin what happens today when a workspace has more members than its plan allows? Answer from the code and cite the file. Say if the help center disagrees.
Hand off to engineering
When the cause is agreed, create the issue with everything engineering would otherwise ask for.
@Merlin create a Linear issue in the Platform project from this thread: the sanitized reproduction, the log excerpt, the code path you found and the affected customer count from the thread. Link back here.
The issue carries the investigation, so engineering does not restart it. The workspace's issue-creation policy decides when task work is represented in Linear.
Draft the customer reply
Once the facts are known, ask for the reply in the tone the team uses.
@Merlin draft a reply to the customer in this thread: what happened, the workaround, and that a fix is tracked. Under 120 words, no internal ticket ids, follow the support-tone skill.
A skill holds the tone rules and the phrases the team avoids, so every reply follows them.
Find whether others hit the same thing
Before escalating, check for a pattern.
@Merlin search Linear for open issues about export timeouts in the last 90 days and tell me if this report matches one of them.
Schedule the escalations digest
Escalations get lost in busy channels. Schedule the list.
@Merlin every weekday at 8:30am, post the Linear issues labeled "customer-escalation" that are still open, grouped by age, with the customer count from each issue.
Scheduled automations deliver to Slack and run on behalf of their creator. Fixed times run in the workspace timezone for a channel automation and in your personal timezone for a direct-message automation.
Keep known issues in channel memory
When a known issue has an agreed workaround, save it so the next person who asks gets the answer immediately.
@Merlin remember for this channel: CSV exports over 50,000 rows time out. The workaround is filtering by month, and the fix is tracked in PLAT-412.
When the fix ships, tell Merlin to forget it or update the entry. Memory in a private channel is visible only to that channel's members.
Limits
Do not paste credentials or full private datasets. Merlin and the chosen providers process the task context needed for the investigation.
A likely cause is not a confirmed fix. Preserve uncertainty and ask for a reproducible case when the evidence is incomplete.