YUCHI WANG · CUSTOMER SUCCESS × APPLIED AI
A customer wants to launch on Friday. One platform works. Another has a blank offerwall. Turn fragmented updates into a focused follow-up and a useful technical handoff.
Independent project · Synthetic customer records · No affiliation with Unity
Explore real engine outputs below.
This is a replay, not a live AI chat.
START HERE / THREE MINUTES, THREE QUESTIONS
A suggested review route.
Open any stop to inspect the evidence.
Android has a failed device check and missing confirmations. Inspect the specific evidence requests, then compare the next question after fresh checks pass.
Discuss: Which new record changes the action, and what remains unanswered?
Open the Android follow-up → Compare the fresh-test update → Read the publisher teaching example ↗A local model completed four requested calls, then described iOS as having “full compliance.” Compare the raw answer with the scoped checklist and the later field review.
Discuss: What did the checks catch, and which parts of the explanation remain unverified?
Inspect the original recorded mistake → Inspect the recorded corrections →Choose a next action and write your reasoning in Practice. Reveal the evidence, then record what you would change. The original response stays separate from the reflection.
Discuss: Which source supports your choice, and what evidence would change your mind?
Try a decision before the guidance → Follow the learning exercises ↗Yuchi selected the project direction and CSM workflow. Implementation, synthetic cases and documentation used substantial AI assistance. The examples show inspectable work; personal mastery and independent customer validation are not claimed.
01 / MORNING REVIEW
Snapshot: 14 Oct 2026 · 09:00 UTC
More than 3 days since recorded
contact; unknown history also included.
Includes platform, review step, language and the comparison toggle. Practice notes are excluded. The case panel always uses the original snapshot; the toggle changes only the separate comparison below.
Original snapshot · this platform only · Markdown
02 / WHAT CHANGES WHEN EVIDENCE CHANGES?
First, confirm two Android settings. Then add fresh device and callback checks. See how the evidence changes the next customer question.
A counterfactual branch. Original records stay intact.
03 / OBSERVE AN ACTUAL MODEL RUN
Compare platform follow-ups, a callback handoff, two inbound messages, and a queue-first attempt with no supplied account IDs. Inspect rejected proposals, recorded corrections, inappropriate tool inputs, and the difference between completed calls and useful customer questions.
Recorded local inference · synthetic customer data · no model runs in your browser · exploratory development attempts, not a benchmark.
Opens this recording and its review. Practice notes and case selections are excluded.
04 / CHECK THE EVIDENCE BOUNDARY
A valid iOS record cannot certify Android. A recent contact with another customer cannot reset this account's clock. Explore the checks that keep those mistakes out of a review packet.
Recorded results from eight synthetic data probes. Declared source ownership and structured claims are checked; the underlying reports are not independently verified.
Inspect what the engine and local host parser allow. A message can contain instructions; that does not make it permission to change facts or send a reply.
Six authored probes, no model inference. These results show specific code boundaries and remaining gaps, not a prompt-injection success rate.
Read the method and limits ↗05 / FROM CASES TO PRODUCT FEEDBACK
A CSM can bring recurring evidence gaps to Product without calling them product defects. Explore the affected accounts, inspect the records, and frame a question worth validating.
06 / MEASURE THE LIMITS
Exact category-set matches on 18 reused synthetic messages. The earlier result was 14/18. Inspected negation failures informed the new rule; this is not a fresh holdout or independent improvement estimate.
A later repair stops “blanket invoice” from matching “blank.” Inspect the before/after probes ↗. Indirect wording still fails. A later explicit-denial repair ↗ excludes selected denied symptoms while preserving a separate positive issue. A bounded callback-denial repair ↗ also preserves separate positive symptoms. Qualified negation, quoted topics and unsupported callback-denial forms remain counterexamples.
The triage baseline uses lexical retrieval. A real MCP host supplies the model; this replay does not call one.
Select a preserved callback-denial probe. Compare categories before the repair with the current rule output, then inspect what still fails.
Sixteen inspected development messages: eight category changes, six positive controls and two remaining mismatches. These are project-authored expectations, not an independent accuracy score. This page replays recorded tool results; it does not run a model.
A matching category does not establish a root cause, resolve a customer issue or approve a message for sending.
07 / PRACTICE THE JUDGMENT
Try six decisions from the actual demo evidence. Choose an action, explain your reasoning, then compare it with the project policy. Keep a reflection you can discuss in an interview.
Self-directed practice, not an assessment. Written reasoning is not graded. Download unfinished notes or reviewed reflections before leaving; reloading clears this page's notes. The page requests a browser warning while practice work exists, but the warning is not guaranteed. Use Save progress to resume later, or download readable notes to discuss. No notes are sent to a server. Guidance is available in the public source.
Share the decision only. Choices, notes and revealed guidance are not included.
This clears choices, original reasoning and later reflections for all six decisions in this page. Download your notes or save a progress file first if you want to keep them. Existing downloaded files are not deleted.
Save a local progress file to continue later. Readable Markdown notes are a separate download. Neither file is uploaded by this page.
Inspect the saved responses below. This preview does not reveal guidance or change your current work.
To cancel, choose Keep current work or press Escape while focused inside this preview. Focus returns to the saved-file chooser.
Resuming replaces all practice answers and notes currently in this page. Save any work you want to keep first. A progress file is editable and does not verify who wrote the responses or when guidance was read.
08 / BUILT TO BE EXPLAINED
This project connects customer communication, process controls, and applied AI. Each draft keeps a trail back to evidence. Each metric includes its limits. Start the learning guide with one customer decision, one AI mistake, and one experiment on a copied dataset. Keep your prediction and reflection as evidence of your own learning.
Explore the learning path ↗“It worked on iOS. Why is Android still blank?” Practice a short answer that separates a recorded failure from missing confirmation. Then compare your wording with the source-linked teaching example and explain what new evidence would change your next step.
Open the customer-education example ↗Download the one-page teaching handout (PDF) ↓
AI-assisted implementation. Personal mastery exercises are provided; completion is not claimed.