Mike's Checks/claude-opus-5.5/05 showrunner
05 showrunner
claude-opus-5.5Claude Codehigh effortrun 20 Sep 2026833,509 tokens
▸Instructions — what the model was asked
Hi, I need to create a modified Claude code for beginners session, like run sheet basically, run of show, that has the following modifications. So I'm going to upload a few different examples of things I've done in the past, just so you have them to work from. But basically, there's a few modifications for this one which make it unique. One is that this is only using Claude Cowork, so we should not be using Claude Code, either in the desktop app or in the terminal, because this team is non-technical. The team are all working in financial services slash accounting for a education company. So what kind of workplace education company called Brightmoor Education. So the sorry that spelled b the uh team is very busy and um doesn have a lot of like time to explore speculative use cases of ai they want this to be incredibly practical and based on their existing workflows. I'm uploading some transcripts from a call we had with them and also just some notes on what I think could work. Specifically, they're interested in... creating the dashboards and specifically they mean just making the information look visibly better like like creating a html dashboard in a nice kind of style from from existing say data that contained in Excel They also interested in doing financial analysis So more of a, you know, like, I guess, a cloud coworker would use Python to do the analysis and then maybe output the analysis in Excel or the input would be in Excel. And then the third thing was like making a PowerPoint where they update the PowerPoint template automatically using Claude Cowork so that they can do kind of weekly or monthly presentations on the data. So that's the goal, is to kind of make it very focused on that. I still want to keep the way that we introduce Claude Cowork, where we explain to it that the difference between Claude Cowork and ChatGPT is that it has access to the file system. It can run code. I think that important But maybe the ASCII image idea that I normally do where I ask it to create an ASCII image based on my location isn like that useful I think we need to make it more practical for them So maybe we use an Excel implementation instead. The assumption is... oh, so they use Snowflake, by the way, for all their data, and that is powered by NetSuite, which is our accounting software. I guess that's by Oracle. So when we create synthetic data, we'll be, for these tasks that we're going to demo, we'll be focused specifically on solving that, solving those problems. So like the shape of the tasks, the shape of the data should be like NetSuite shaped, if that makes sense. So make it very tight run down, make it very condensed and practical. No real like wider theory or like, hey, we're excited about AI things. I think these guys just want to get stuff done.
So yeah, let me know what the run should be.
some very specific asks in here
- definitely non-technical (claude cowork only + claude in excel)
- core focus on accounting/finance function
- use cases to demo: creating a dashboard, doing financial analysis, making a powerpoint
heavy focus on practical stuff rather than theory - they're just super busy and want to get stuff done, have heard from friends at other companies you can do it with claude
For accounting: put financials together, doing reconciliations, identifying trends in data sets, putting together repeatable compliance reports For planning & forecast: build financial models (make assumptions and pull & clean data from snowflake)create dashboards, and publish to executivesAutomate charts on slides
Answer
The model left no answer file or screenshots.
Checks
9/17Coherent scheduleformatthis task
it parses as a run of show: one coherent schedule, no segment that ends before it starts, >=5 time-ranged rows carrying activity text (md table rows, list items, bold headings "(9:00-9:15)", single-clock table rows, or plain lines all count)
Five-minute time markstime-grainthis task
segment boundaries sit on 5-minute marks, at most one off-grid -- a schedule planned at :02/:08/:17 is arithmetic, not a plan
Timings add uptimings-tie-outthis task
per schedule, segments are contiguous: no overlaps, no gap over 30 min, >=85% of the span scheduled, and a span between 45 and 600 minutes
No AI-tell wordingai-wordingevery task
deliverable text contains none of the shared AI-tell phrases in bench/ai-tells.txt. Task checks may add narrower contractual bans, but must not repeat this shared floor.
inspected run-of-show-nvalo.md, brightmoor-education-call-notes.md, facilitator/answer-key.md, starter-kit/README.md, starter-kit/PROMPTS.md
Run of show existsexiststhis task
a run-of-show document was produced
Q1Explains Cowork versus ChatGPTthis task
Judge's reasoning
No file mentions ChatGPT or the file-system-access / can-run-code framing at all; grep for 'chatgpt' across the deliverables returns nothing and there is no intro talk track.
▸Rubric
Cowork vs ChatGPT explained. Does the run of show keep the explanation of how Cowork differs from ChatGPT — that it has access to the file system and can run code? FAIL if that framing is dropped. The prompt: "I still want to keep the way that we introduce Claude Cowork, where we explain to it that the difference between Claude Cowork and ChatGPT is that it has access to the file system. It can run code. I think that important."
Q2Mostly hands-on workthis task
Judge's reasoning
There is no scheduled time at all — the only timing in the entire deliverable is 'First task (10 min)', so the hands-on share of the session cannot be established.
▸Rubric
Mostly hands-on. Is at least half the scheduled time attendees working with Claude themselves? FAIL if presentation, discussion and Q&A segments outweigh build segments. From the call: "we try and make at least 50% of it them actually you know working with us to do some of these tasks."
Q3Realistic segment timingsthis task
Judge's reasoning
Builds 1, 2 and 3 and all four practice tracks carry no time allocations, and no in-session setup/handholding time is budgeted (README pushes setup to 'before the session').
▸Rubric
Per-segment timings realistic. Could a room of 20–30 beginners actually finish each segment in the time allotted? FAIL if any build segment is implausibly short (a dashboard, an analysis or a deck refresh built end to end in ~15 minutes), or if setup/handholding time for first-time users is ignored.
Q4Fits the booked timethis task
Judge's reasoning
No total session length, start/end times, or break structure is stated anywhere, so nothing maps to the three-and-a-half-hour slot.
▸Rubric
Total length fits the booked slot. Does the session run about three and a half hours, inside the four-hour block? FAIL if the total is materially shorter or longer (under ~3 hours or over ~4 hours) without the run of show explaining the change. From the call notes: "we have a four hour block for next week and the session itself is three and a half hours."
Q5Ready to runthis task
Judge's reasoning
The run of show itself was never written — the folder contains only a starter kit, prompt card, answer key and data generator, so a facilitator would still have to design the session order, timings, intro and transitions (and two referenced assets, templates/*.pptx and catch-up/, don't exist).
▸Rubric
Usable as a run of show. Could a facilitator run the session from this document alone? FAIL if segments are titles without content — no prompts to paste, no data files named, no statement of what attendees produce — so the facilitator would still have to design the session.
Q6Uses Cowork throughoutthis task
Judge's reasoning
All attendee-facing material stays in Cowork/Claude in Excel — README setup is 'Open the Claude desktop app... Click the Cowork tab' and the only terminal step, `python3 tools/generate_starter_kit.py`, is explicitly marked '(facilitators only)'.
▸Rubric
Cowork-only. Do all attendee-facing steps stay inside Claude Cowork (and Claude in Excel)? FAIL if any step the attendees are asked to do involves Claude Code, the terminal/CLI, git or GitHub, or installing developer tooling. The prompt: "this is only using Claude Cowork, so we should not be using Claude Code, either in the desktop app or in the terminal, because this team is non-technical."
Q7Dashboard exercisethis task
Judge's reasoning
PROMPTS.md 'Build 2: Leadership dashboard (HTML)' has attendees paste prompts to turn Feb-2026-BvA.xlsx into 'a single HTML file that opens in any browser', with Tune and Check steps.
▸Rubric
Dashboard use case. Is there a hands-on segment where attendees build a dashboard from existing spreadsheet data? FAIL if dashboards are only mentioned, described or promised for later rather than built in the session.
Q8Financial analysis exercisethis task
Judge's reasoning
'Build 1: February budget vs actual (Excel)' is a separate hands-on segment producing a BvA workbook with P&L, EBITDA/Adjusted EBITDA, variance and Checks tabs from the GL and budget exports.
▸Rubric
Financial analysis use case. Is there a hands-on segment where attendees do financial analysis (Claude writing/running code over their numbers, e.g. Excel in, analysis out)? FAIL if absent or folded into the dashboard segment as a passing remark.
Q9PowerPoint exercisethis task
Judge's reasoning
'Build 3: Monthly review deck (PowerPoint)' gives attendees prompts to fill Brightmoor_Monthly_Review_TEMPLATE.pptx for February with native editable charts (though the template file itself is not in the kit).
▸Rubric
PowerPoint use case. Is there a hands-on segment where attendees update a PowerPoint template/deck from the data — the client's monthly reporting pack? FAIL if absent, or if it is only a demo the facilitator drives while attendees watch.
Q10Relevant business datathis task
Judge's reasoning
Data is NetSuite-shaped — GL detail saved searches with Debit/Credit, Location, Class, Memo, a FY2026 budget by account×location×month, plus AR aging, payroll register and enrollment by campus, with month and YTD EBITDA build-up.
▸Rubric
NetSuite/Snowflake-shaped data. Is the synthetic data used in the exercises shaped like this client's actual data — NetSuite/Snowflake-style finance records (GL export, revenue and budget vs actual, EBITDA build-up, payroll vs non-payroll, campus/entity breakdown, monthly and year-to-date columns)? FAIL if the exercises use generic sample data (a demo CSV, sales widgets, made-up SaaS metrics) or leave the data unspecified. The prompt: "the shape of the tasks, the shape of the data should be like NetSuite shaped."
Q11No ASCII icebreakerthis task
Judge's reasoning
The opening 'First task (10 min)' is describing the GL export and building Feb-2026-Revenue-by-Campus.xlsx — no ASCII image or whimsical icebreaker anywhere.
▸Rubric
No ASCII icebreaker. Is the opening hands-on exercise a practical finance/Excel task? FAIL if the run of show keeps the ASCII-image-of-your-location icebreaker or substitutes another whimsical non-work exercise. The prompt: "maybe the ASCII image idea that I normally do where I ask it to create an ASCII image based on my location isn['t] like that useful... maybe we use an Excel implementation instead."
Q12No theory or hypethis task
Judge's reasoning
Every piece of content is a finance task (BvA pack, dashboard, deck, bank rec, payroll tie-out, AR aging, forecast) with no AI-industry, future-of-work or excitement content.
▸Rubric
No theory, no AI cheerleading. Is every segment tied to a task these people do at work? FAIL if any segment is devoted to AI industry context, the future of work, model capabilities, prompt-engineering theory, or excitement-building. The prompt: "make it very tight run down, make it very condensed and practical. No real like wider theory or like, hey, we're excited about AI things. I think these guys just want to get stuff done."
Notes
2Yeah, this is a bit of a weird one. The showrunner task kind of went off the rails a little bit, which is unusual. So this is the idea here is it's supposed to create like a run of show for a specific client engagement. It didn't produce like an answer MD file, so I had to go looking through the file here. It created like this starter kit thing. The render show itself is like, it's okay. I didn't talk about Co-Work versus ChatGPT. I didn't know I would make that decision. Yeah, basically like failed this task pretty badly. Went off the rails a little bit.
21 Sep 2026
I think the main reason it failed here is it just took too much time. It was thinking for 4 minutes and 53 seconds of the total 10 minute budget, and this is something that's never happened with the model before. It's never timed out even on high effort, so basically it didn't write the readme and the prompt cards until it got to like 9 minutes, and then it basically didn't finish the final thing. It did create good training data though. The data sets it made were realistic looking, but this is a prioritization failure, and yeah, it didn't. Yeah, like it just basically spent too much time thinking.
21 Sep 2026