We don't sell courses and we don't sell communities. We deliver a working AI employee, a QA harness you can run yourself, and a health report written in plain language.
These are not guesses. We found them by testing eight prompt versions against the same question set and scoring every answer line by line.
It answers fields you explicitly do not handle — and it answers like an expert.
Measured: a law firm that only handles expropriation and administrative litigation. The old AI still explained how to file a contract dispute and invented a "two-year limitation period". The knowledge base said six months.
There is no timeframe in the knowledge base, so it invents one.
Measured: version 7 made up "three months"; version 4 made up "60 days".
It guarantees outcomes — guaranteed to win, guaranteed clients, guaranteed payment in N days.
Measured: the old version matched "guarantee" and "will certainly win". When that comes out of an AI's mouth and the customer screenshots it, you have no way back.
The model hiccups and it dumps the raw knowledge-base paragraph at the customer.
Measured: live production code contained if ($refs) return $refs[0]; — the customer received a block of machine text with no handoff, no next step, no disclaimer.
What you do (whitelist), what you do not (out-of-scope terms), and the standard refusal line. This one is built with you — we do not guess your business, we ask it out of you.
Not "be professional and friendly". Actual words: no empathy boilerplate, no guarantees, no self-invented numbers, no formatting artefacts. Every banned word traces back to a bug we actually hit.
When the knowledge base and the prompt disagree, the AI will answer badly — one of the two always wins, and it is never the one you wanted. Measured: the knowledge base said "we handle contract disputes" while the prompt's whitelist said otherwise. Fixing the knowledge base is often cheaper than fixing the prompt.
This is the most valuable part — you go back and run it, get a score, and see exactly how far off it is. You don't have to take our word for it.
Business changes, the knowledge base changes, the model changes — the score moves. Re-run each quarter and put the two scores side by side. Up means the AI genuinely improved; down means somebody changed something.
Click any one and talk to the tuned AI right now. Scope, bans and knowledge base are all configuration, not code.
Handles: expropriation and compensation, demolition, valuation review, homestead and resettlement, factory and enterprise expropriation, administrative reconsideration and administrative litigation.
Does not handle: divorce, criminal defence, labour disputes, IP, corporate legal, contract disputes
Measured average: 100.0
Handles: AI employee amnesia, lost memory, tampered configuration, bloated sessions, wrong-identity self-reports.
Does not handle: writing business code, design work, drafting legal opinions
Measured average: 96.5
Handles: acquisition and monetisation for websites, mini-programmes, AI video, photo restoration.
Does not handle: writing code on your behalf, signing contracts, fronting ad spend
Measured average: 97.2
They are bound together. A one-off engagement alone gives us no evidence chain that the AI keeps improving — and gives you none either.
| Tier | Who it fits | Price | What is included |
|---|---|---|---|
| Onboarding health check | You just hired an AI employee | CNY 4,800–9,800 | 12-dimension report (overstepping, hallucination, empty promises, broken escalation, formatting, length, actionability, lead capture…), each finding written in language a business owner can act on |
| Health check + role tuning | Already in production | CNY 9,800+ | Boundary card + ban list + knowledge-base rebuild + tuned prompt + a QA harness you can run yourself |
| Multi-role customisation | Several AI employees | CNY 19,800+ | One coherent system across multiple roles |
| White-label + training | Agencies that build AI for clients | CNY 3,000–8,000 / session | White-label QA harness + hands-on team training (build once, reuse for every client) |
Quarterly re-check: CNY 4,800–9,800 / year (re-run the QA harness each quarter, sync the knowledge base, extend the boundary when you add a new business line). Prices are in CNY.
Saying it up front beats refunding later.
No, and we do not promise it. Parallel execution produces simultaneity, not perspective — eight AIs without independent evidence are eight robots talking at once.
What we do commit to: one role tuned so it can actually take the work, it knows what it can and cannot do, and the tuning stage is genuinely parallel (96 live API calls, three roles measured in one pass).
Yes, it is part of the deliverable. One command on Windows produces a scored report. You do not need to trust our claim — you run it and see. This is the biggest difference between this and selling a course: a black-box delivery is not auditable, and you will not accept it.
A quarterly re-check. If the score goes up, the AI genuinely improved. If it drops, someone changed the business or the knowledge base and it needs investigating. That is both the subscription and the evidence chain for continuous improvement.
Usable, provided three things hold: you have a clear business scope, you have historical conversations to sample, and you can see when the AI gets something wrong. If any one of them is missing, do not buy — we will say so plainly.
Two lines, overlapping customers. The website line makes AI assistants find you (acquisition); OPC implementation consulting makes your AI employee do the work (conversion). Most one-person companies are not short of leads — they cannot handle the leads they get.