Whitepaper and LLM instruction set · Sturdy × Claude
The Portfolio Cockpit is an LLM instruction set that builds a CSM's daily working surface from what customers actually said, ranked by the money and dates in your CRM. Every claim on the page carries the customer's own sentence.
Part 1 · The old way
Every Customer Success Platform ships one. On paper it is exactly the right surface. Many are quietly abandoned after onboarding. That is a design problem, not a training problem.
The problem: what a portfolio view is built from, and why nobody can defend the score.
Our solution: detection from customer language, prioritization from the CRM, evidence under every claim.
The setup, the activity record, and the honest limits.
Part 1 · The old way
Customer language lives in email, Slack, support conversations, meeting transcripts, and call notes. It is where customers explain what they need, what is frustrating them, what they have committed to, and what they might buy next. It is some of the richest intelligence a company has, and it is remarkably difficult to use at scale.
Most systems built to understand customers were designed around structured data: product usage, CRM fields, survey scores, renewal dates. Until recently there was no practical way to read all the language, so companies asked people to summarize the conversations themselves. That decision sits underneath most of the problems with modern Customer Success tooling.
A CSM carrying twenty to sixty accounts does not have a reporting problem. They have a triage problem.
Every morning, some portion of the book needs attention and the rest does not. The job is to find that subset quickly, act on it, and avoid discovering problems for the first time on a renewal call eight weeks later.
Part 1 · The old way
The page is assembled from telemetry, surveys, CRM fields, and whatever the CSM remembered to log. Telemetry tells you what someone clicked, not what they think. Surveys are intermittent. CRM fields hold only what someone took the time to enter. The critical context sits untouched inside the conversations.
Thresholds, alert rules, and playbooks are authored months before the situation they are meant to interpret. A rule written in advance cannot know that the person raising a concern is the same executive who stopped replying three weeks ago, or that the customer just started using the word "procurement."
A page that shows the same eleven amber accounts every morning teaches the user that opening it produces nothing new. A daily surface needs a diff: what changed, what moved, what went quiet, what came due.
An account appears green because usage is strong, even though the champion said on a call that their budget is under review. The customer already told the company what was happening. The system was not listening.
Part 1 · The old way
One of the strongest risk signals is the absence of communication. An account that stops responding generates no alert, no threshold breach, and no new sentence for a rules engine to classify. It stays green right up until the renewal does not happen.
Composite scores blend usage, surveys, support activity, and custom weighting few users understand. Ask a CSM why an account is yellow and the answer is that it is what the system says. Once the score cannot be defended, it stops being treated as information.
A score without explanation is decoration. A score with evidence is useful.
Part 1 · The old way
Does the CSM still open it on day thirty without being told to, and can ten minutes with it end in three clear decisions? Everything below was designed against that standard.
Part 2 · The new way
The most important structural decision in this build is separating them. The Sturdy Corpus supplies detection. The CRM supplies prioritization. Money and time.
Part 2 · The new way
Detection
What the customer said, in their own words.
Prioritization
How much money is attached and how soon it matters.
The CRM ranks findings by how much money is attached and how soon it matters. Useful, but downstream of the finding itself.
Part 2 · The new way
Every element either changes what the CSM does today or explains why an account is in the state it is in. Nothing else belongs on the page.
Cancellation language appeared three days ago and the account is single threaded on one contact.
We're going to need to pause the renewal until we've reviewed alternatives with procurement.
Dana Whitfield, VP Operations · Email · 3 days ago
Security is a critical signal and this is one business day old. Fast, complete answers here usually close the issue outright.
Our infosec team flagged the SSO configuration during the annual review and wants a response this week.
Ana Duarte, IT Security Manager · Zendesk · 1 day ago
Support promised an ETA seven days ago and it has not been sent. The same account raised a value issue two days ago, so the open commitment is compounding.
Six months in and my team still exports everything to a spreadsheet to get the view they need.
Marcus Feld, Senior Director · Slack · 2 days ago
Next three actions · sample data
Not ten, and not a ranked backlog. If the CSM completes only these three items, they have done the highest-value work available that day.
Book value splits into active, in renewal window, under churn notice, and unscored. Renewal ARR splits into committed, forecast, and at risk, because at risk is the number the CSM owns, explains, and reports upward.
New signals, risk-tier moves, accounts entering the renewal window, conversations going quiet, commitments coming due. This is the panel that creates the daily habit, and "no material change" is a good result to report.
Sorted by ARR, colored by score. Every score expands into its reasons and the customer language behind them. The color prioritizes. The evidence holds up in a forecast call.
Up to five cards, one per account, each with the quote, speaker, role, channel, timestamp, source, and suggested next step. "Budget Issue, Meridian Freight" tells the CSM very little. The sentence where the customer says their budget came back flat tells them what conversation to have.
Silence weighted by ARR and renewal proximity. An account with meaningful ARR, a renewal in sixty days, and no response in a month surfaces even if its health score is green.
Active contacts over ninety days, flagging single threading, executive changes, contacts going dark, and bouncing addresses. One active contact is one resignation away from a competitive evaluation with no internal advocate.
Inbound asks the customer is waiting on, and outbound commitments our side has not met. An unanswered request is a service failure. An unmet commitment is a credibility failure.
Expansion signals, reference candidates, and recurring feature requests, held to the same citation standard as the risk panels. Expansion interest recorded nowhere else is revenue the company may never capture.
Missing renewal dates, missing ARR, past-due opportunities, unresolved contacts, undeliverable addresses, and the value excluded from scoring. Bad CRM data corrupts every panel above it, so the exclusions become a short list of fixes.
Part 2 · The new way
The listed contributions sum to the number displayed. CRM-sourced contributions are tagged, and the score is recomputed from the inputs actually available.
| Input | Points | Source |
|---|---|---|
| Critical signal | 40 each | Corpus |
| Warning signal | 15 each | Corpus |
| Info signal | 5 each | Corpus |
| Repeat of the same risk type within 30 days | 10 each, capped at 30 | Corpus |
| Customer spoke last, no reply after 3 business days | 20 | Corpus |
| Exactly one active contact in 90 days | 20 | Corpus |
| Expansion or Happy signal within 30 days | minus 15 | Corpus |
| Renewal inside 90 days | 25 | CRM |
| Renewal inside 45 days | 40, replacing the above | CRM |
Full weight inside 14 days, half at 30, one fifth at 60, zero past 90. Last quarter's problem does not keep voting at full strength.
An account that cannot be scored renders Unscored, never Green.
Verbatim quote, speaker, role, channel, timestamp, link. Quotes are never paraphrased or assembled from fragments. A card without a citation is a defect.
Part 2 · The new way
Task creation, escalation, and opportunity creation are all possible with the same connectors. Each creates another object the CSM has to manage, and none is central to the job the cockpit does.
The reply lands in the conversation where the CSM can read and edit it. Short, answers the specific point the customer raised, proposes one next step, written in the CSM's voice. Nothing sends automatically.
The corpus reads from every system a customer talks to you in, but we can only write back into some of them. Drafting in the chat works the same way no matter where the message came from.
One activity record, written only after a confirmation showing exactly what will be written. The conversation that never reaches the CRM is the conversation that disappears during handoffs, renewals, and escalations.
The write is narrow. It appears only for discrete, material communication events with a quote and a date that did not already sync from the CRM and are not offered elsewhere on the page.
Part 2 · The new way
With a CRM connected, it produces the full page. Without one, it produces a corpus-only view. Money fields, renewal dates, and CRM hygiene rows are removed rather than estimated. The financial summary becomes a communications-derived view: accounts with coverage, open critical signals, silence, single threading, and no coverage. Scores are recalculated from the inputs actually available.
When data is missing, remove the dependent element and say why.
Part 3 · What you get
Three pieces, and you probably already have two: the Sturdy Corpus, your CRM, and a Claude project holding the instruction set and template. Salesforce is the worked example. The pattern holds for any CRM with an activity object.
Part 3 · What you get
No configuration screen. No field-mapping exercise.
Add the Sturdy connector and, for financial prioritization, your CRM connector. Add an email connector only if you want drafts pushed into a drafts folder.
A new Claude Project with a clear name, such as My Cockpit. It becomes the place the CSM returns to each day.
Paste it into the project instructions field. It defines connector detection, retrieval, risk taxonomy, scoring, citations, actions, and the activity record format.
Upload it to the project files. It controls layout, design tokens, connector-specific behavior, confirmations, and score expansion.
The project detects what is connected, retrieves, computes, populates the data object, and renders the page.
The reasoning layer and the rendering layer meet in one populated data object. Redesign the page without changing how data is gathered, or change the gathering without redesigning the page. Every value traces back to a named field and a known source.
Part 3 · What you get
One completed activity record per logged event, created only after the CSM confirms.
POST /sobjects/Task { "Subject": "<risk type> noted", "Description": "<channel>, <date>. <speaker>, <role>:\n\n\"<verbatim quote>\"\n\nLogged from Portfolio Cockpit. Source: <link>", "WhatId": "<Opportunity id for renewal-scoped cards, otherwise Account id>", "WhoId": "<Contact id, only when the match is confident>", "Status": "Completed", "ActivityDate": "<date of the communication, not today>", "Type": "Other", "Priority": "Normal", "OwnerId": "<the CSM's user id>" }
Not the date it was logged. Backdating keeps the account timeline chronologically honest, which is the point of writing it.
Salesforce derives it from the related record. A direct write fails.
Channel, date, speaker, role, quote, and source. The one interpretive element is labeled as a classification, not presented as something the customer said.
If the connector authenticates as a service account, every logged activity is attributed to the integration user and the activity reporting is ruined.
Part 3 · What you get
Six things it cannot do, stated plainly.
A concern raised on a call and never confirmed in writing is invisible unless transcripts are ingested. Accounts with no communication coverage render Unscored.
There is no stored last-view timestamp. Tell it when you last looked and it computes against that. Otherwise it uses the last three days and labels the panel with the window it used.
ARR and renewal dates come from the CRM or they do not appear. Nothing is inferred from conversation content.
Where a Sturdy account has no confident CRM match, money fields stay empty, the account renders Unscored, and it counts in Data hygiene. The same rule applies to contacts.
With an external id field for source communications, a repeat click updates rather than duplicates. Without one, it warns you once per session.
Start with one portfolio rather than an entire team. Silence windows and score thresholds ship with defaults, and the defaults are not your book.
Sell it as armor rather than surveillance. It identifies risk before the renewal call, creates an evidence trail that supports the forecast, and eliminates two tedious tasks: logging context and composing follow-up.
Everything ships ready to run.
If you would rather not build it yourself, we will stand it up with you against your own communication record in a single session. Email hello@sturdy.ai.
One email, three files, no sequence.
Loading the form. If it does not appear, email hello@sturdy.ai and we will send the files.