We run two small AI tools right now: a resume screener and a resume builder. Both burn real LLM costs on every request.
How do you account for that? That's what a credit system is for. But from day one, we knew it couldn't serve just these two tools. Every new product can't ship its own credit logic. So we built it as infrastructure: whatever business runs on top, there's only one ledger underneath.
Four rules for the credit system
1. Separate business logic from money. The credit system doesn't care what your feature does — screening resumes, building resumes, whatever. It sees one thing: a request came in, deduct N credits. The more complex the business, the simpler the credit system must be.
2. Centralize pricing. All credit consumption rates are unified, priced by request cost, controlled globally. New product wants to launch? It plugs into the credit system. No exceptions — otherwise costs spiral.
3. Pre-charge, then refund. Credits are deducted when a task starts, so we never work for free. If the task fails, credits go back. The user sees two records: one charge, one refund. They pay for results, and they see we didn't cheat them. That's the most primitive form of trust.
4. The minimum unit is 299, not 300. Deliberate. If a task costs 300 credits, we charge 299 and leave 1 behind. That single credit is worthless — but it means the user's balance never hits zero. They remember there's still something left with us.
The core value is still the product. But financial design adds a small layer of retention on top.
The system works. The products don't.
Now the launch story. Both tools finished their concentrated promotion window a while ago. Here's what actually happened.
Our method: work backwards. Set the deadline first, then split phases. Way more effective than "let's just start and see." Three phases: concentrated launch, catch-up, review.
During the launch phase, we covered all channels with both video and text, cross-promoting both tools. Day one data came in and I already knew: this probably isn't going to work. So from the 20th, we moved into catch-up early, overlapping with the launch phase — private channels, secondhand marketplaces, anything that pulls numbers fast.
The review phase isn't just a retrospective. If the data is acceptable, we use it to attract new users and drive referrals. If it's not, we stop investing in both tools, keep them running, and move to the next build.
Data doesn't lie. If the product has no pull, don't stay emotionally attached.
What killed the tools wasn't a competitor. It was the LLM itself.
The resume builder launched into beta and got nearly zero users. The reason is blunt: writing a resume is something you can just ask an LLM to do. Nobody wants an extra entry point for that.
The resume screener got some usage, but HR in mainland China is hardwired to recruiting platforms. Asking them to screen resumes in an external tool goes against their instincts. And it faces the same extinction risk — once models can batch-process resumes efficiently on their own, the tool is meaningless.
So the call is clear: if the resume builder shows no traction by end of October, we delete it at the code level. The screener stays available but gets no more ad spend.
The beta forced us to see who the real customer is
The biggest value of this beta was forcing clarity on something we'd left fuzzy.
Our original logic: resume tools attract job seekers to feed the talent pool; the screener attracts HR to build a demand-side pool; add human optimization later; close the loop. It worked on the slide deck. It doesn't work in a real workflow —
HR runs recruiting, but HR doesn't decide whether to outsource, and HR doesn't own the budget. The boss does.
That's aiomniu's real customer. HR and project managers are executors. The boss is the source of demand.
So what we build next has to appeal to business owners: cost reduction, efficiency, new revenue, growth, compliance. The credit system holds the cost line. The beta data drew the life-or-death line for products. And the boss's ledger — that's the thing we actually need to learn to read.
Top comments (0)