Triage customer support tickets and prepare fixes or replies for your approval
Have Codex triage incoming customer support tickets and prepare a reply or a proposed fix that waits for your approval before anything goes out.
What the agent did
Source-reported example · as described by Hacker News user mike_hearn (full name not stated in the source), describing his own part-time side business; his Hacker News profile lists Hydraulic, maker of Conveyor
The author (Hacker News user mike_hearn), who runs a part-time side business (his Hacker News profile lists Hydraulic, maker of Conveyor, a tool for distributing desktop apps, and in the thread he says Tesla is a Conveyor client), built his own setup about six or seven months before the comment and now runs two Codex instances he calls "AI employees".
Each runs in a permanently rolling session in its own UNIX user account, is woken by incoming email (via Maildir) and by scheduled timers, and has accounts in his bug tracker and CI systems.
For support, they triage and work on customer support tickets: he says that sometimes he wakes up to find the fix or response for a customer-filed ticket already prepared and waiting for his approval.
When the setup started, the agents had very few privileges and showed him what they were going to do for each task; he checked carefully before giving them more freedom.
Reported: the setup has been running "more than six months" ("about six/seven months" in the original comment) and "it's been great". Reported: "Sometimes I wake up and the fix/response for a ticket filed by a customer is already there waiting for my approval." No ticket counts, time saved or accuracy figures are given.
Limitations
This is a short Hacker News comment thread, not a write-up: there are no ticket volumes, response times, accuracy figures or examples of a drafted reply. The author says the agents "sometimes make mistakes" and he still checks their work; he started them with very few privileges and widened access only after checking their work over time.
He also says he recently began letting them interact directly with customers in specific scenarios, which goes beyond draft-and-approve; this listing covers only the approval workflow. "Fix/response" can mean a code fix as well as a reply, since the agents also do the coding for his software business.
The setup is custom (dedicated UNIX accounts, Maildir wiring, systemd timers, bug tracker and CI accounts) and the author does not publish its configuration or instructions. With a single rolling session an agent can only work on one thing at a time; he added a second agent for parallelism but says one was sufficient 95% of the time.
He reports one memory glitch where an agent forgot it had been given credentials to a service. The comment describes other jobs (bug-backlog cleanup, coding, server upkeep, release help) that are not part of this listing.
Try it in your business
What you'll need
Codex access; a Linux machine with a dedicated user account per agent; email delivery into Maildir with a trigger that wakes Codex on new mail; scheduled wakeup timers; agent accounts in your support ticketing or bug-tracking system and CI; command-line tools or API access to those services; your own review and approval of each prepared fix or reply
Codex runs in dedicated UNIX user accounts with custom Maildir email triggers, systemd wakeup timers with POSIX locks, and accounts in the author's bug tracker and CI; this is a self-built harness, not a packaged app connection.
The original is a custom-built system. The prompt below is a simpler starting point and won’t reproduce it.
Starting prompt
The author didn't publish their prompt, so this is our adaptation.
How ours differs from the original setup: Same execution agent (Codex). The author's agents run continuously in their own UNIX accounts and are woken by email and timers; our prompt assumes you trigger Codex per ticket or on a schedule you set up.
It is draft-and-approve only: unlike the author's recent change, it never lets the agent contact customers directly, and it never merges, deploys, closes tickets or deletes anything. Ticket labels and the approval-queue location are our additions, since the source does not describe them.
You help me run customer support for my small software business. When a new support ticket or support email arrives [ticket system or mailbox], read it and any related history, then triage it: label it as a bug, question, feature request or duplicate, note its urgency, and link any existing ticket it matches. For questions, draft a reply to the customer. For bugs, investigate in the code repository [repo], reproduce the problem if you can, and prepare a proposed fix on a separate branch with a short explanation, without merging or releasing it. Put each draft reply and proposed fix in my approval queue [where] with a one-paragraph summary and anything you're unsure about. Do not send anything to customers, merge, deploy, close tickets or delete anything; I will review and approve every reply and fix myself.
Replace the [bracketed] parts with your own files and tools.
Get examples like this every Tuesday and Friday
Practical ways to put AI agents to work in your business. We check the original sources, explain the setup and point out the limits. Free.
Unsubscribe any time. Privacy policy
Related use cases
Draft the weekly recap email to a client
Have Grok Bot write the weekly recap email to a client, then clean it up, as the author says he does.
Our suggested starting prompt — not the author's original · untested
You'll need: Grok Bot with enough access to, or notes about, the week's work for the client (the post does not say which); your usual recap format or an earlier recap to follow; your own read-through and clean-up of each draft before it goes to the client.
Draft this week's recap email to [client name]. Cover what we did for them from [date] to [date], what is in progress and what we need from them next. Use [where the week's work is recorded: notes, a task list or a folder I give you] and my last recap, [paste or attach], as the format and tone to follow. Where you are unsure whether something happened, mark it 'check' instead of stating it. Do not send anything to the client; give me the draft to review and edit.
Check local prospects' email security and draft personalized outreach for your approval
Have Grok Bot research local businesses' public email-security records and draft personalized first-touch emails for your approval, the work the author says his AI team did.
Our suggested starting prompt — not the author's original · untested
You'll need: Grok Bot, or a Grok Bot team set up with roles like the author's; a list of the businesses to research (the post does not say whether the team found them or he supplied them); somewhere to keep the drafts and findings; your own approval of every email before anything is sent.
Here is a list of local businesses: [name, website, contact]. For each one, read only the DNS records the domain already publishes publicly and note whether it has SPF, DMARC and DKIM records and anything that looks missing or misconfigured. If you cannot tell (for example DKIM may depend on a selector you cannot see), write 'unknown' instead of guessing. Do not probe, scan or log in to anything. Then draft a short first-touch email to the owner, in plain English and in my tone, that names what you found for their domain and offers [my free check or service]. Put the findings in a table (business, domain, what you found, how sure you are) and save each draft with it. Do not send any email or contact anyone; give me the table and the drafts for my review and approval.
Categorize bank transactions, number vouchers, match payments to invoices and draft year-end accounts
Have Claude Cowork do the bookkeeping from a bank statement and spreadsheet, match payments to invoices and draft year-end accounts, as the author describes in a LinkedIn post that he says Claude wrote.
Our suggested starting prompt — not the author's original · untested
You'll need: Claude Cowork with access to your bookkeeping spreadsheet (and a backup copy), your bank statement export and your sales invoices; last year's accounts prepared by your accountant, for the format; a rules file kept alongside the spreadsheet; your own answers to its questions; your accountant's review.
You are helping me with bookkeeping in Cowork. Files: bank statement [file], bookkeeping spreadsheet [file], sales invoices [folder], last year's accounts from my accountant [file]. Step 1: make a backup copy of the spreadsheet before changing anything and tell me where it is. Step 2: read the spreadsheet and describe my categories, how voucher numbers run and which kinds of transaction never get a number; wait for me to confirm. Step 3: fill in only blank rows where the category and voucher number are clear from the history; for anything new or ambiguous, list a question for me instead of guessing, and never overwrite a filled row. Step 4: check your work against the backup: row counts, and gaps or duplicates in the numbering; report what you find. Step 5: match each client payment to a sales invoice and list any invoices still outstanding and any matches you are unsure about. Step 6: save the rules we agreed, and my answers, to [rules file] and show me what was added. Then, only after I have reviewed all this, draft this year's accounts in the same layout as last year's, section by section, from the categorised data, and flag anything that needs my accountant's judgement (for example depreciation, tax and opening balances) instead of filling it in. Do not send anything to my accountant or anyone else.

