AI products I shipped — and the method behind them.
Aaron Zhang · Product Manager @ TikTok LIVE
→ / space / swipe · F fullscreen
the operating belief
I don't ask AI for magic.
I teach it how I work.
Own the problem — own the AI.
Ch.1 What I shipped · Ch.2 How I work
CH.1 · SHIPPED · three AI products live in production
Three products, one thesis: put AI where the work actually happens.
LIVE
Campaign AI
An AI teammate for campaign ops — data, diagnosis, next-best action.
workflow-embedded
LIVE
Smart Diagnosis & Outreach
Finds campaign issues before anyone asks — and fixes them in one click.
workflow-embedded
LIVE
Creator Coach
Turns expert playbooks into real-time next steps for every creator.
workflow-embedded
how I read AI product maturitycopilot · you ask, it helps→workflow-embedded · runs inside the workall three are here→AI-native · exists only because of AIwhere I'm headed
Take the AI out and all three still have a platform underneath — so none of them is AI-native yet. Knowing which rung you are actually on is the job; the coach products I'm building next are the ones designed AI-native from day one.
shipped on TikTok LIVE · 2025–2026
CH.1 · CAMPAIGN AI — my role: 0→1 product owner
An AI teammate that knows the data and the playbook.
Campaign AIillustrative mockup
How is my campaign doing this week?
Participating creators▲ on track
Reach rate▼ below benchmark
Rewards spendon pace
Next best action → refresh the outreach push for the lagging audience. Drafts ready to review.
Ask it in plain language. It reads the live data, works out what is actually going on, and hands back the next move — not just a chart to interpret.
where it sits today
workflow-embedded
It has moved past the chat box: it lives inside the campaign workflow, so the answer shows up where the decision is actually made.
CH.1 · CAMPAIGN AI · how it runs
From a question to the next action — in one loop.
You ask (or it notices)→Reads the live data→Weighs history & comparable campaigns→Checks config, audience & pacing→Recommends the next action
the diagnosis is the hard part — not one benchmark check, but rules, history, comparable campaigns, config and audience weighed together
50%+
of campaign operators use it
10+
messages per active user each week
From dashboard-digging to a 30-second answer.
Not a chatbot bolted on — a teammate in the workflow.
CH.1 · SMART DIAGNOSIS — my role: 0→1 product owner
It doesn't wait to be asked.
Watches every live campaign→Detects the anomaly→Notifies the owner→One-click fix
the diagnosis arrives before anyone asks
Config alertillustrative mockup
A config issue puts exposure at risk
An unnamed live campaign is set to under-deliver. Detected automatically — before the owner noticed.
One-click fix available
CH.1 · SMART DIAGNOSIS · impact
Thousands
of campaigns covered each quarter
85%+
fix adoption — flagged issues actually resolved
Tens of millions of impressions protected.
…and when reach lags, it drafts the outreach plan.
Three dimensions at once — phase × language × audience segment. Who to reach, when, and in which language: a whole plan, ready to review.
70%+
of campaign messages AI-created
10%+
lift in reach rate
outreach planillustrative mockup
phase ↓ / audience →
New
Active
Lapsed
Phase 1
×3 langs
×2 langs
×3 langs
Phase 2
×2 langs
×3 langs
×1 lang
Phase 3
×1 lang
×2 langs
×2 langs
CH.1 · CREATOR COACH — my role: 0→1 product owner
It hears the worry first — then does the math.
Creator Coachillustrative mockup
now on · a flagship campaign
Can I still make the next round?
Coach
1 · hears the intent
You can still catch up.
2 · diagnoses
You're #78 with 12,580 points — about 340 short of the top-50 cut-off.
Gifts
Match
Tasks
6,200
3,800
2,580
3 · hands back the move
3h left. Set a live goal now, then start a match — the fastest way to close the gap.
Set goalStart matchSee tasks
Ask me anything…
A creator doesn't want a dashboard. They ask "am I still in it?" — and the answer has to come back as a decision, not a number.
1Meet the intent, not the words. "Can I still make it?" is a worry before it is a query. Answer the worry first, or nothing after it lands.
2Diagnose against something checkable. Rank, the exact gap, and where the points came from — so the advice can be argued with instead of believed.
3End on one doable move. Not "engage more" — a specific action, sized to the time actually left.
CH.1 · CREATOR COACH · impact
Tens of thousands of creators coached daily.
95%+
positive rating
10%+
revisit rate
Participation and revenue both moved — A/B-validated.
Expertise that reached the top creators — now it reaches everyone.
CH.1 · in one page
3
AI products live
Thousands
of campaigns served each quarter
Tens of thousands coached daily.
Revenue lift, A/B-validated.
Shipped, adopted, measured.
CH.2 · HOW I WORK · the magician's climb
AUTONOMY ↑
L1 · ASK
Apprentice
学徒
It answers. You still do the work.
L2 · ACT
Spellcaster
法师
It does the task. You check it after.
L3 · EMBED
Enchanter
阵法师
It runs the workflow. You hold the gate.
L4 · GOVERN
Archmage
魔导师
It runs the system. You set the goal & limits.
→ PORTABLE OWNERSHIP
L5 · TEACH
Grandmaster
宗师
You share the method. Anyone can run it.
Find your rung with one test: if AI did your repeated task confidently wrong tomorrow — would you catch it? That's how high you can safely climb.
L1 · ASK · own the meaning
Same data. Same weeks. Opposite answer.
AI found the culprit in seconds — clean, confident, and wrong. It had counted the wrong denominator. Count only the eligible base, and the same weeks read a steady 91–96%. The −19.6% was a definition pretending to be reality.
Fast confidence is not correctness.
one metric · two denominators — which one is real?
L2 · ACT · own the proof
It doesn't hand you a draft — it operates your tools.
campaign-config · adminillustrative mockup
AI agent is controlling this tab — batch-filling campaign config · 14 fields · 6 languages
Campaign namefilled✓
Start / end windowfilled✓
Reward tiersfilled✓
Eligibility rulesfilled✓
Localized copy · 6 languagesfilling…✓
The leap: it doesn't just answer — it drives your tools, doing the repetitive back-office clicking field by field, on its own.
meeting notesdesign filesweekly reports
It acts on its own — so verification is born.
L2 · ACT · so I pointed it at something that mattered
Filling a form is low stakes — if it slips, you see it. So I handed it something where a mistake would travel: the monthly review other people make decisions on.
It built the whole thing in seconds. Early versions were quietly wrong — 3×.
attribution breakdownillustrative mockup
every day of the month · pulled, checked, written
one month — pulled, broken down by cause, charted, and written before coffee
but early versions were quietly wrong — 3×
Ranked wrong — leaderboard sorted the wrong way
Counted twice — one campaign double-summed
Week started Sunday — one rule flipped the conclusion
Demo is easy. Production is right every time.
Fix the rules, not the answer.
L3 · EMBED · own the workflow
The loop runs — and closes — without me starting it.
Something happens→AI drafts→I approve→Task opened→Tracked to closure
weekly reviewillustrative mockup
🧭 Weekly on-call digest
grouped by severity · drafted for approval · tracked to closure
Open threads grouped by severity→ the highest-risk items surface first
One escalation drafted for approval→ nothing escalates without my sign-off
💡 every item is tracked until it closes
I don't kick this off. The workflow watches, drafts, and — once I approve — opens the task and follows it to closure. This is a process I already knew how to run: I wrote the steps down, and now it runs them. I hold the gate; the loop does the legwork.
gates · escalation · closure
L4 · GOVERN · own the system
I stopped telling it what to look at — I told it what I care about.
L3 · embed
I designed the steps. It runs a process I already knew how to run, and stops at my gate.
vs
L4 · govern
I never wrote the steps. It decides what deserves attention out of an open world — in a field I don't know. I set only the goal and the guardrails.
Trend Intelligence Engine · my own side project
🧭 Daily AI digest · one day's run, 2026-07-03
1 major · 2 watch · 2 monitor | 70 sources → 436 raw → 308 leads → 1 digest
Model access is now gated by policy & safety — a frontier model re-deployed behind cyber-guardrails→ build regional fallback & degrade paths
Open models enter coding-agent cost/routing — Codex ↔ GLM · Copilot adds Kimi K2.7→ make model routing a config knob, not a rewrite
💡 read-out: switch models by task to keep cost and risk in check
🚨 tracked with deadlines: legacy model retirement · EU AI Act obligations
I knew nothing about intelligence-mining. I owned the question — "what actually deserves our attention?" — and the verification. AI supplied the domain's tricks. Built on public sources, on my own time.
what's left to own? → the system — the goal & the guardrails.
L5 · TEACH · own the method
The method that caught my −19.6% — now anyone can install it.
I froze the whole "who counts?" checking method into a Skill — a reusable instruction pack, an SOP for AI. A teammate who never saw the trap now gets the corrected answer. Ownership stopped living only in my head.
what it took to ship one safely
A second AI reviewed it and found 6 things I'd gotten wrong. It was tested again and again. A human still had to approve it before it could go live.
the trap at the top
Package your first Skill and you feel unstoppable. That's the most dangerous moment — mastery is a moving target.
Demo → production is still a very long road.
the enabler · the harness
A raw model is a genius on day one — no laptop, no logins, no memory.
Acts · 自己干活
calls the tools, runs the steps
Self-checks · 自己评测
reads the result, picks the next move
Remembers · 持续记忆
keeps the context over time
cheap AI fetches→smart AI judges→cheap AI writes it up→I approve
The model is not the product — the harness is.
three principles I keep
01 · own the meaning
Fast confidence ≠ correctness — own what the number means.
02 · own the proof
Demo → production is a very long road — own the proof.
03 · own your taste
The tools are the same for everyone — your taste is the 10×.
The suit doesn't fly the pilot. The pilot flies the suit.
The tools are the same for everyone. Your taste is the 10×.
No Magic, Only Method.
Own the problem, own the AI.
what I'm building next
I'm working on coach-style AI products — career coaching, study-abroad advising, and more. If you're building in this space, I'd genuinely like to compare notes.