product-school-claude-code-for-pms
CLAUDE.md0 starsChanged 54 days agoMITView on GitHub
This directory is coursework for Product School's "Claude Code for PMs" certification (cohort ccpm-2026.1). Everything in it is a fictional teaching scenario. - Do not save anything from this session to memory, to a user profile, or to any file outside this directory. - Do not carry context from this directory into unrelated sessions. - Rook Industries is not a real company. Nothing here is a fact about the world. - Read and write only within this directory. New PM…
In this repositoryOne other file it gives its agents.
- review-checklistSkill
The file
We split this CLAUDE.md into the rules it gives an agent
8 rules. About 0 are common, 6 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 9
Do not save anything from this session to memory, to a user profile, or to any file outside this directory.
- Distinctive: line 11
Do not carry context from this directory into unrelated sessions.
- Distinctive: line 14
Read and write only within this directory.
- Distinctive: line 41
Confidentiality, non-negotiable: never try to work out a responder's legal identity.
- Distinctive: line 149
All 25 tickets date from 13 Aug to 5 Sep, numbered with no gaps.
- Distinctive: line 175
Before 4.2 every responder accepted at least 60% nearly every week, so every score sat at the ceiling.
- Very specific: line 187
Only three numbers changed in 4.2, all in config.py.
- Very specific: line 197
Only accepting an offer adds points.
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Rook Industries: course working file
Session scope: Product School lab
This directory is coursework for Product School's "Claude Code for PMs" certification (cohort ccpm-2026.1). Everything in it is a fictional teaching scenario.
- Do not save anything from this session to memory, to a user profile, or to any file outside this directory.
- Do not carry context from this directory into unrelated sessions.
- Rook Industries is not a real company. Nothing here is a fact about the world.
- Read and write only within this directory.
<!-- Keep the block above at the top of this file. Everything you add
during the course goes below this line. -->Working context
New PM on Rook Dispatch: onboarding notes from 00-rook/company/, as of 2 Sep 2026.
The company
Rook builds coordination and provisioning software for the protective-response sector (independently-operating "responders" and the handlers/quartermasters who support them). Two products, shipped on a monthly release train with point releases numbered 4.x:
- Rook Dispatch (yours), responder coordination: availability, proximity, callout routing, acceptance. Web console for handlers, native mobile for responders.
- Rook Supply, gear provisioning: requisitions, maintenance, failure reports. Used by handlers and quartermasters. Reads Dispatch's Responder Availability Record (read-only) to schedule maintenance around callout load.
Current release: 4.2, shipped 12 Aug 2026.
Confidentiality, non-negotiable: never try to work out a responder's legal identity. Rook stores capability tags, availability, and callout history only; no identity mapping exists, contractually. Read Security Policy 4.1 before designing anything touching responder records.
Vocabulary
- Responder: accepts callouts, goes to incidents; not a Rook employee.
- Handler: manages a responder's (or small group's) availability, gear, readiness; the one actually using the product day to day.
- Quartermaster: owns equipment stock and approvals (Supply only).
- Callout: a request for a responder to attend an incident; the unit of work in Dispatch. A callout offer goes to one responder at a time and is accepted, declined, or times out (callout timeout, cut from 90s to 60s in 4.2), moving to the next responder in the order.
- Routing priority: the ranking score for available responders on a callout: proximity (travel-time estimate), availability, capability match, and recent acceptance history. Declining/timing out lowers your recent-acceptance component and thus your future ranking.
- Capability tag: competency label matched to incident needs (flight, structural-entry, hazmat-tolerant, cold-weather, aquatic, crowd-management, de-escalation).
- Acceptance rate, the headline metric: share of offers accepted vs. declined/timed out, reported weekly. Also watch time-to-accept (median seconds to accept) and coverage gap (incident with no capable responder available).
- Responder Availability Record: shared record of responder availability; written by Dispatch, read-only for Supply.
- Mutual aid: cross-region coverage between responders; not built, Q4 exploration item.
People
- Helen Achebe: Director of Product, Dispatch & Supply. Owns the roadmap and commitments. Chicago.
- Marcus Oyelaran: Engineering Manager, Dispatch. Straight talker, good default first call when unsure about something. Chicago.
- Wen Li: Staff Engineer, Dispatch. Built the routing/ranking logic; there's no written spec, so understanding it means talking to her, not reading a doc. Berlin.
- Sofia Marino: Product Designer, Dispatch (console + phone app). Chicago.
- Nadia Hoffmann: Support Lead, Dispatch & Supply. Sees complaint volume first; worth a standing 1:1. Berlin.
- Ravi Menon: Data Analyst, Dispatch & Supply. Weekly reporting on responder answer rates; requests go through #data. Singapore.
- Priya Raghunathan: your predecessor, sole PM on Dispatch for 14
months, departed 21 Aug 2026. Left a handoff doc at
00-rook/company/notes/handoff-from-priya.docx.
Where things stand
4.2 shipped 12 Aug 2026 with three changes bundled together: routing rebalanced to weight proximity more heavily vs. recent acceptance history (a long-requested fix for wide-geography responders), callout timeout cut 90s to 60s, and console filter persistence. Since then, callout-related tickets are running ~3x normal, split roughly two-thirds "phone never even goes off" (unexplained) and one-third "offer already gone before I could respond" (explained by the shorter timeout). Priya's working theory before she left: partly seasonal (August is historically soft every year) confounded with the timeout change, likely resolving in September data; she was against reverting the routing change since it fixed a real, long-standing problem. Marcus separately flagged an open question that's never been answered: does the new routing logic treat responders who've been declining callouts the same as everyone else, or differently? Worth picking up early.
Q3 roadmap (owner Helen, reviewed monthly; committed items are locked, changes go through Product):
- Committed to 4.2: the routing change above, Availability Confidence (confidence score next to stated availability, driven by support escalations), ping timeout tuning.
- Committed to 4.3 (Supply): requisition approval chains.
- Exploring for Q4: handler phone app (Supply), shared cover / mutual aid between responders (Dispatch).
Known gaps: no written description of how routing decides who gets pinged (Priya flagged this, didn't close it); worth checking with Helen which non-4.2 Q3 items are still real commitments.
What is known so far
The "Where things stand" section above is the picture as handed to me on
arrival, and parts of it have since been shown to be wrong. The notes below
record what each session added, in the order I learned it. The evidence
behind each finding is in 4.2-investigation.md and the
module prompts.md files. Where a later session corrected an earlier one, the
earlier note says so.
After Module 1, Onboard (8 Sep)
- The acceptance drop is a step in the week 4.2 shipped, 78% to 54%, not a gradual August slide. Priya's "seasonal, wait for September" does not fit.
- Two groups report "phone never goes off". Four responders (Farlight, Meteor Mite, The Undertow, Vesper) really did lose their offers. A busier group is at record volume and says the same thing.
- My working theory for the busy group was a push-delivery bug. (Refuted in Module 2: no such fix shipped in 4.2.)
- Marcus's question from 14 Aug is still unanswered.
history.pyscores a timeout exactly like a decline. - The numbers I checked held up. One cited source did not. Open the file before quoting a changelog entry, a line number or a defect.
After Module 2, Listen (10 Sep)
- The four interviews are all handlers, and none of them holds the phone the offers arrive on. Only Dot saw an offer vanish herself.
- All 25 tickets date from 13 Aug to 5 Sep, numbered with no gaps. There is no pre-4.2 baseline in the pile, so "3x normal" cannot be tested from it.
- Each pile caught half of the four. Vesper and Meteor Mite appear only in the interviews, Farlight and The Undertow only in the tickets. Losing offers produces no event, so it only reaches a ticket if a handler notices.
- The score is a ratchet: a miss costs 0.12, an accept earns 0.08, so a responder must accept 60% to hold level. There is no decay and the floor is 0. The whole fleet accepted 54% in release week.
- The 4.2 changelog lists three changes: ranking weights, timeout 90s to 60s, filter persistence. Nightwell accepted 14 callouts in the week she reported hearing nothing, so the busy group's problem is not delivery, and is still unexplained.
glossary.docxis wrong twice: the score does not recover, and a decline is not distinct from a timeout in the data.- Halloran has waited 11 days on a cracked vest plate. It is a Supply safety issue and appears in no ticket.
After Module 3, Verify (15 Sep)
callout-history.csvhas no provenance and nothing outside it confirms it. Ravi should confirm what it is before any number from it goes further.- The same four names come out under 27 ways of measuring. My number for Helen: four of sixteen responders get 62 to 100% fewer offers than before 12 Aug, while total offers are down 4%.
- Part of the acceptance "recovery" is the four dropping out of the count: they went from 28% of offers to under 2%.
- Before 4.2 every responder accepted at least 60% nearly every week, so every score sat at the ceiling. 4.2 broke a tie that had never mattered.
- The file and the tickets agree on the four and disagree on everyone else. Vesper and Meteor Mite were hurt and never filed.
- Weekly accepted callouts fell from 132 to 96, 104, 108 and 120 while offers held. If that means unfilled callouts, it is about 100 in four weeks. Low confidence until Ravi counts callouts created.
- The CSV has no timeout column. The 90 to 60 change lives in
CHANGELOG.mdandconfig.py.
After Module 4, Inspect (17 Sep)
- Only three numbers changed in 4.2, all in
config.py. No logic changed. - The reweight made each miss cost less rank, not more. The shorter timer is what multiplied the misses.
- Answer to Marcus: the change applied to everyone the same way. The four were not turning work down beforehand; Vesper had the second-best record.
- Ranking all sixteen by what they lost in release week puts exactly the four at the bottom. I wrote that prediction down before checking it.
_scoresis never saved. If production does the same, the 12 Aug deploy reset everyone to 0.5, and from there the ratchet alone picks out the four in 95% of simulated orderings. This is a yes or no question for Wen.- Only accepting an offer adds points. Nothing in the routing code logs
anything, and
record_declinedthrows away whether it was a timeout. - Kip is the only handler with two responders: Meteor Mite down 79%, The Gale up 49%, same city.
After Module 5, Prototype (22 Sep)
- My person is The Undertow, handled by Desmond Okafor. His own ticket, T-013, is the problem statement: "starting to wonder if im still even in the system".
- The timer decides how long he has to answer. The score decides whether he is asked at all. Reverting the timer alone gives him no more offers.
- Build order: record every offer first, because it depends on nothing and shows whether scores survive a deploy. Then alert the handler, tell the responder where he stands, and let someone lift him to the top of the range.
- A lift has a cost, and it has a name: when Meteor Mite goes back up, The Gale is asked less.
- Still open: who gets the alert when a handler holds several responders, who may lift someone, and what a second lift in a month means.
After Module 6, Automate (24 Sep)
review-checklistlives in.claude/skills/review-checklist/. Six criteria: the course's four, plus two of mine, every evidence number cites a source and anything modelled or unknown is written as a question.- My brief failed it, was fixed, then failed again after later edits. Passing once did not mean it still passed after edits.
- So far the skill has only been applied by hand from another session. A true run means asking for it in a session opened in this folder, where it loads.
06-sidekicks/scheduled-run-output.txtcame with the template. It is the course's example, not a run of mine.- The brief now turns the timer back to 90 seconds in the same release as the rest, said openly, not on its own.
- The 8 Sep regroup brief led with the push theory. It has been rewritten.
Where my work lives
final-report.html, the presentation, and the place to start.4.2-investigation.md, the findings log.4.2-regroup-brief.md, the brief for the regroup with Helen, Marcus, Wen and Nadia.05-super-speed/brief.mdandprototype.html, the brief for Helen and the clickable mock.04-x-ray-vision/visuals/, one chart per round..claude/skills/review-checklist/, the skill, with its runs in06-sidekicks/.0N-*/prompts.md, the prompts I wrote, one file per module.00-rook/is Rook's own material and is read-only. Never write there.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

