CSM COPILOT — PRESENTER GUIDE

Purpose: show one useful customer decision, one recorded AI mistake, and a
learning step you can explain yourself. This is a rehearsal aid, not evidence
that you have completed the practice. Suggested pacing has not been measured.

BEFORE THE CONVERSATION

Extract the whole package. Follow START_HERE.txt to serve it over localhost.
Keep project-brief.pdf open as a backup. The static demo replays saved outputs;
it does not run a model or contact customer systems. The customer records are
synthetic. External sources and hosted sharing links require internet access.

At http://127.0.0.1:8000/ open these routes in separate tabs:
  ?case=lantern_android&view=evidence&lang=en#workspace
  ?run=qwen4-batch#model-runs
The paths above belong after the trailing slash. On the hosted portfolio, use
the same paths after https://yuchi-wang02.github.io/csm-copilot/ instead.

THREE-MINUTE CUSTOMER STORY

0:00–0:30 — Frame the problem.
Show Lantern Games and the snapshot label. Explain that a publisher wants to
launch but Android evidence is incomplete. State what decision the CSM needs
to move forward. Do not claim this is a real customer or measured business gain.

0:30–1:10 — Trace the decision.
In Evidence, expand "Trace checklist decisions to source records" and inspect
"Non-test device verification". Contrast its recorded failure with unknown
currency approval. Explain why a fresh device test and a missing approval
confirmation are different requests. An iOS check does not establish Android
readiness, and a complete scoped checklist is not launch authorization.

1:10–1:50 — Prepare the next action.
Open Follow-up, then Handoff. Show a concrete request and a missing intake
field. The handoff needs SDK version, time window and redacted logs; it retains
the customer's recorded region and reproduction details. Explain which owner
needs which information. A generated draft sends nothing and does not count
as a new customer contact.

1:50–2:30 — Inspect an AI mistake.
In the qwen4-batch recording, open "Check the claim against the recorded result".
Contrast the model's "iOS's full compliance" with the scoped checklist result.
The model completed four calls, but its wording overstated what they proved.
Explain the division of work: deterministic tools produce workflow results;
an optional model selects tools and writes an explanation that still needs review.

2:30–3:00 — Explain your contribution and learning.
Describe the project direction you selected and the substantial AI assistance
used for implementation. Give one decision or code path you can explain in
your own words. If you have completed a practice or source-change experiment,
show your actual prediction, observation and reflection. Otherwise describe
the next exercise as planned work. End here and invite a question.

OPTIONAL FOLLOW-UP: A REAL CITATION USED FOR THE WRONG CLAIM

Open ?practice=citation_purpose#practice. This sixth exercise shows a customer
report used in both an Android readiness citation list and a technical handoff.
Write your own reasoning before revealing guidance. Ask: what exact claim does
the report support, and what evidence supports the failed device check?

Then inspect ?run=qwen4-morning-field-review#model-runs. The host rejected one
readiness citation even though its source ID and platform were real. The same
report legitimately supplied handoff details. Source existence, scope and
claim support are separate questions. The original failure remains recorded.

For a broader engineering question, expand "Compare the morning-review
attempts". Separate recorded setup, tool execution and answer review. Seven
calls without a final answer differs from a returned but rejected answer.
The attempts changed several controls and are not a controlled model benchmark.

YOUR PREPARATION — LEAVE BLANK UNTIL YOU DO THE WORK

The customer decision I can explain:

The source record that supports it, and its limit:

My prediction before changing one input in a separate working copy:

What I actually observed, and what surprised me:

What I would investigate next:

The Practice section keeps initial reasoning separate from later reflection.
Save progress JSON before leaving; readable Markdown notes are a separate
download. Resume previews earlier work before replacing the current session.
Compatible five-decision files add the sixth decision blank in this release.
Six-decision files cannot be resumed in the older v0.30.0 application. Files
remain editable and do not authenticate authorship or personal mastery.

KEEP THE TAKE-AWAY SIMPLE

Share project-brief.pdf for the overview, the ZIP for a fixed interactive
snapshot, or the hosted portfolio for the evolving version. Both PDFs and this
guide can be read without running Python. Share-link buttons in the local app
point to the public site, which may have advanced beyond this package.

Independent, AI-assisted portfolio project by Yuchi Wang. No affiliation with
Unity. Tests and recorded failures support inspectable engineering claims;
they do not establish independent model reliability or production impact.