Meta-review: the session that built this

One Claude Code session, 13:19 to 14:25 on 13 Sep 2026. Everything on this site, the repo, the plugin, and the five simulated fellows came out of it. This page is the same kind of artefact the workshop asks you to share, applied to itself.

By the numbers

Human messages	30	about a third typed while Claude was still mid-task
Median human message	~12 words	the longest was the pasted brief; most were one line
Assistant messages	~300
Tool calls	~140	file writes, shell, web search, browser, artifact, Vercel
Wall clock	66 min	of a 3-hour cap
Things built	7	task repo, probe + checks, skill, share plugin, comparison generator, live site, 5 simulated runs
Things thrown away	2	a six-field card board, then a web uploader; both replaced by files in a repo

What the human did

Almost every human message was a steering move, not a specification. In order:

Set the frame and the constraint. "I'll need to save process logs." Claude started a timestamped log before anything else, and every later decision has a line in it.
Asked for evidence before deciding. "search for best practices first." Eight minutes, five searches, four sources that survived.
Overruled the AI's recommendation. Claude argued for a single-skill verification workshop. "I still think peer-sharing is valuable, particularly for researchers who may tend to work independently." That call set the whole design, and the AI's alternative became the worked example inside it.
Simplified, twice. Claude built a structured card board. "No, simpler. I'd like to allow people to share their skills, upload their logs." Claude built an uploader. "Is there not a simple way…?" The answer was files in a repo.
Named the real comparison. "Compare the process logs and skills in one view? Particularly user input." That one line is the reason the page shows prompts first and replies as dropdowns.
Demanded reusability. "So that each user can just download the claude plugin and it allows sharing?" Turned a script into something installable.
Verified. "verify the code." 28 checks, one of which caught the test author's own wrong expectation.
Tested with users, then more users. "simulate the test run, and share." Then "simulate a few more users." The simulation found a leak in the exercise: the data generator's docstring spoiled the answer.
Caught a consent gap. "Does the plugin allow selecting which skills to share or confirm with user?" It didn't. Now it previews, asks, and asks again before pushing.
Fixed what only a human sees. "The cards jut off the right of the window."

Where the AI was wrong, and how it surfaced

Severity: high would have shipped a wrong research claim or exposed private data · medium a broken tool or an invalid exercise · low wasted time or cosmetic.

Severity	Mistake	Caught by
high	Published the human's session log with the pasted work-test brief in it, then made the repo public.	The human, six minutes later. Repo re-privated, text redacted, git history rewritten, old deployments purged.
high	Predicted the probe would learn negation; the first run said template. Later runs said negation and length.	Running the check instead of trusting the hypothesis. Both readings kept on the page.
high	Copied every skill silently, no consent.	The human asking one question.
medium	Introduced a variable-shadowing bug while adding reply capture.	The next run crashed. Fixed in a minute.
medium	Wrote a spoiler into the exercise's data generator.	Five simulated fellows all "finding" the confound suspiciously fast.
medium	Missed the human's mid-turn messages in the log extractor.	Writing this page and noticing the best prompts were absent.
low	Recommended against the human's design.	The human. The design was better for this cohort; the AI's idea survived as content.
low	Over-built the sharing mechanism, twice.	The human, with one-line prompts.
low	Wrote a test with a wrong expected count.	The test failing; the code was right.
low	Estimated timestamps instead of reading the clock.	A `date` call. Log corrected.

Sorting by severity says the important thing. All four high-severity errors were caught by a human asking for a check, or noticing, not by the tool. The tool caught its own medium ones by crashing or failing a test. Nothing caught the high ones except someone deciding to look.

The pattern, if you want to copy it

Log from minute one. The log is what makes the rest reviewable, including this page.
Let the AI argue, then decide yourself. The AI's case was reasonable and evidence-backed. It was still the wrong frame for this cohort, and only the human could know that.
Say "simpler" early and often. Three of the most valuable messages were under ten words.
Verify, then simulate. Unit-level checks caught code bugs. Simulated users caught a design bug the checks could not see.
Compare what people typed. Six runs, same model, same data, one caveat versus five checks. The difference was the prompt. See the comparison.

Caveats

The five "fellows" are Claude sessions, not people. All of them could read the docstring that spoiled the confound before it was removed. One had the sanity-check skill available and used it. The human and the AI on this page are the same pair that built the page; take the self-assessment accordingly, and check the raw prompts and replies and the full timestamped process log yourself.