Replace your CSP · Chapter 02 of 06
Your CSP renders one number nobody in the room can defend. Sturdy reads every email, call, and ticket the account produced and Claude grades five dimensions with the customer's own words attached. Not a score. A diagnostic.
This is chapter two of six. One screen at a time: last chapter was the timeline, this one is scorecards and health.
The full whitepaper and the complete Claude project instruction set. No install. No sales call.
What your CSP shows you
What moved? The support tickets from the spring?
What this project shows you
Rule applied: no state dimension is red, so the composite does not escalate to red. One execution red alongside two state yellows → composite yellow. The rule is printed on the artifact, with counts.
The actual construction
Open up any CSP health score and this is what you find. Not one input is the customer's own words. The pricing dispute, the champion who stopped replying, the sentence where someone wrote "we're not seeing the value": all of it sits in threads the score never read.
Measures whether people open the software, not whether it works for them. A team can log in daily and be actively furious.
Treated as negative, which inverts the truth. Engaged customers file tickets. Customers who gave up file nothing.
Single-digit response rates, self-selected, weeks late. A quarterly NPS cannot see a relationship dying in March.
The most informed input and the least objective. Every CSM is optimistic about their own book. That's being human.
What replaces it
There is no 0-to-100 axis in this artifact. The project runs in Claude, reads the complete communication record over Sturdy's MCP connection, and returns five graded dimensions with a composite verdict. Warning lights, not a gauge.
68 and 71 are not meaningfully different, but they look different, and people act on the difference. A grade makes a claim exactly as strong as the evidence supports: something is wrong here, in this system, and here is what tripped it.
Two accounts at the same blended score can need opposite plays: one has an unhappy customer, one has a slow support queue. Averaging destroys exactly that distinction. Split into what's happening to the customer and what you're doing about it, the two can't hide from each other.
Every color traces to a written threshold in the project instructions. When a CSM disagrees, they open the evidence: either they learn something about their account, or the rule is wrong and they fix the line that produced it. Neither outcome is available from an integer.
A stalled integration is a work item with an owner and a date, not a customer deciding they're done. A rollup that treats those identically teaches the team to ignore both. Only the customer's condition escalates the composite.
Why the output holds up
Same account, same window, same grades, every run. Put this month next to last month and any difference is a change in the relationship, not a change in how the model felt that day.
Inbound share below 20% is red. Friction rate above 14% is red. Plain numbers in a document your team owns, attached to quantities that mean something. When a grade disagrees with reality, you adjust one line and re-run. That's the intended workflow, not a workaround.
Two to four cited threads per dimension: subject, date, and what each one shows, with the verbatim sentence where the wording carries the meaning. A grade you cannot open is a grade you cannot act on.
A quiet healthy account and an account with broken ingestion look identical in three colors. Gray means insufficient data: a capture gap gets a flag and a fix, not a false red that sends someone into a save motion against a customer who was never leaving.
Your customers have customers. A cancellation in a thread might be theirs, not yours, and to a naive classifier the two are identical. The diagnostic filters third-party language before grading, reports raw and attributed counts, and names what it excluded so you can check its work.
The full series
Same argument, one workflow per chapter, each with its own instruction set. Start with the screen your team feels most.
No more manual logging. The live account timeline.
● Live — read itFive graded dimensions with the evidence attached. This page.
● You are hereClaude drafts and maintains the plan, from evidence of what's actually worked.
Coming soonA daily, prioritized action briefing — done and not-done, in one view.
Coming soonAt-risk renewal reporting anchored at the portfolio level, not one account at a time.
Coming soonA live account picture, assembled on demand. The finale.
Coming soonFree, no strings
The complete "After the health score" whitepaper plus Appendix A: the entire Claude project instruction set — signal map, extraction sequence, attribution filter, every numeric threshold, and the rollup rule. Paste it into a project connected to Sturdy and run it on an account you know. Nothing needs editing before first use.
## Hard rules 1. The score is deterministic. The narrative is not. 2. Never fabricate a number. 3. Gray is a real state, not a failure. 4. Resolve signal IDs at runtime. Never hardcode. 5. Apply the attribution filter. ## Step 6 · Set colors ### Sentiment load - RED if crit_rate_pct > 12, or any attributed Cancellation / Value Issue names the account… ### Engagement vitality - RED if inbound_share_pct < 20, or days_since_last_inbound > 30… ## Step 7 · Roll up the composite 1. Two or more dimensions gray → GRAY 2. Any state dimension red → RED 3. …
Whitepaper, instruction set, and the HTML dashboard shell. We'll email all three. No sales call attached.
Instructions land in your inbox. Run them against your own accounts today.
"An account is at 68. Last month it was 74. Nobody in the room can say what moved."
Now somebody can.