The CSM's Prompt Library: Eight Prompts That Earn Their Place in the Workflow

The wish-prompt problem

There are roughly four hundred thousand articles titled "ChatGPT Prompts for Customer Success" indexed by Google. Most of them contain prompts like:

"Act as a customer success manager. Help me write an email to a customer who is unhappy."

That's not a prompt. That's a wish. The model produces something that sounds like what a customer success manager might write — sincere, hedged, vague. The CSM who sends it loses the customer faster than if they'd written it themselves in three sentences.

A working prompt for CS does three things the wish-prompts skip:

1. Specifies the actual constraint. Not "help me write an email" — but "this customer has been escalating since last Tuesday, our champion left in March, the renewal is in six weeks, and I need a response that doesn't commit to anything the company can't deliver."
2. Loads context the model can't infer. The customer's specific situation, the political reality of the account, the language the customer uses in their own emails.
3. Defines what success looks like. Not "make it good" — but "produce three drafts ranging from formal-direct to warm-deflective, each under 120 words, no marketing language."

This piece is eight prompts that do all three. They're organized by the Decision Matrix quadrants — three in Automate, three in Augment, two in Inform — so you know exactly when to reach for each one. There are no Leave-Alone prompts. By definition, those tasks don't get AI prompts.

The point isn't the prompts. It's the pattern. Once you see the format, you can build prompts for any CS task that matters.

The format

Every prompt in the library follows the same template:

• Title — one line describing what it does
• Quadrant — Automate, Augment, or Inform
• When to use — two sentences of trigger context
• The prompt — copy-paste ready, in a code block
• What good output looks like — what you should expect on the page
• Failure modes — two or three things the model tends to get wrong
• Variation hook — what to change for different account contexts

The format itself is the credibility move. Vague prompts produce vague output. Structured prompts produce structured output. If your prompts don't look like the ones below, your output won't either.

Section A — Automate

These produce internal-only artifacts. The customer never sees the result. Errors are recoverable by the CSM reading the output before filing it.

Prompt 1: Post-call internal note from a transcript

Quadrant: Automate
When to use: You just got off a 30–60 minute customer call. You have the transcript (from Otter, Gong, Whisper, whatever). You need a structured internal note in your CRM in 3 minutes, not 30.

You are converting a customer call transcript into a structured internal CSM note.

The note is for internal CRM use only. The customer will never see it.

Read the transcript below. Then produce exactly these five sections in this order:

1. DECISIONS MADE — bullet list of any decisions reached during the call. Each bullet under 15 words. Include who made the decision if it's clear from the transcript.

2. ACTION ITEMS — bullet list of action items, each in the form "[Owner] will [action] by [deadline if mentioned, else 'TBD']."

3. STAKEHOLDER SIGNALS — note any change in attendees, sentiment, or political dynamics. Specifically flag: a new stakeholder mentioned, a stakeholder departing, a frustration that wasn't there before, or an unexpected enthusiasm.

4. RISK FLAGS — anything in the conversation that suggests renewal risk, expansion blocker, or escalation potential. Two sentences max per flag.

5. NOTABLE QUOTES — up to 3 verbatim quotes from the customer that are worth keeping. Use direct quotes only; no paraphrasing.

Constraints:
- Be specific. "Customer was happy" is not useful; "CFO said the integration with NetSuite was the deciding factor" is useful.
- If a section has nothing to report, write "None this call." Do not fabricate.
- Total length under 400 words. Be ruthless.

TRANSCRIPT:
[paste transcript here]

What good output looks like: Five labeled sections, each scannable in 5 seconds. The action items section uses verb-first phrasing. The risk flags section either has 1-2 specific entries or "None this call." No marketing language.

Failure modes:

• The model summarizes the call instead of structuring it. If you see narrative prose in section 1 or 2, the prompt didn't bite — re-run with stronger formatting instruction.
• The model invents stakeholders or action items not in the transcript. Always read the output against the actual call; if anything looks novel, it's hallucinated.
• The model softens "risk flags" into "opportunities." Tell it explicitly: risks are risks, not euphemisms.

Variation hook: For accounts where the company language differs from the CSM's voice, add at the end: "Use the customer's actual terminology when quoting them — do not normalize it."

Prompt 2: QBR prep outline from account history

Quadrant: Automate
When to use: QBR in six days. You need to plan what to actually cover. The prompt produces a slide-by-slide outline based on the customer's history and current state.

You are building a QBR outline for an internal prep document. The outline structures the meeting; the CSM will write the actual slide content.

Read the account context below. Produce a QBR outline with exactly 7 slides:

1. OPENING — one suggested theme for the meeting (a tension or question, not a brag).
2. WINS SINCE LAST QBR — 3 specific wins, each linked to something the customer cares about. If a "win" can only be described in product feature terms, leave it out.
3. METRICS THAT MATTER — the 3 metrics this customer specifically asked about last QBR, plus the one new metric you should introduce this quarter.
4. OPEN COMMITMENTS — what we said we'd do last QBR. Honest status: done, in progress, blocked, or dropped. Do not soften "dropped."
5. WHAT'S CHANGED ON THEIR SIDE — anything in the account context that suggests their priorities have shifted.
6. ASK — one thing we want from them in the next 90 days. One, not a list.
7. NEXT STEPS — placeholder; CSM will fill this in based on conversation.

Constraints:
- Each slide gets a one-line "what this slide answers for them" question. Not a topic — a question.
- If their last QBR commitments include items that are "dropped," call them dropped. The QBR is more credible when the bad news is on the page.
- No slide titles like "Strategic Alignment" or "Partnership Vision." Use plain English.

ACCOUNT CONTEXT:
[paste account summary, last QBR notes, current usage data, recent activity]

What good output looks like: Seven slide titles, each with a single-line question underneath. The "wins" section excludes anything that's actually a feature pitch. The "open commitments" section is honest about what didn't happen.

Failure modes:

• Model defaults to corporate-speak slide titles. Re-prompt with: "slide titles must be plain English questions or statements, no nouns like 'Alignment' or 'Vision.'"
• The "wins" section gets padded with weak entries. Tell it: "three wins maximum, and if you can't find three real wins, list two."
• The "ask" section becomes a list of three things. The prompt says one. Enforce it.

Variation hook: For QBRs with new senior stakeholders attending, add: "Slide 1 should also include a 90-second framing of how the customer uses our product today, for any attendees who weren't in previous QBRs."

Prompt 3: Account-update digest for your manager

Quadrant: Automate
When to use: 1:1 with your manager in 30 minutes. You need a digest of what's changed across your portfolio this week. Sorted by ARR descending.

You are producing a weekly portfolio-change digest for a CSM's 1:1 with their manager.

The output is internal-only. The manager will scan it in under 2 minutes.

For each account in the activity log below, produce one line in this exact format:

[ARR band] | [Account name] | [What changed this week] | [Status]

Where:
- ARR band: $0-100K, $100-500K, $500K-1M, $1M+
- What changed: one specific change. Not "had a meeting" — but "introduced new sponsor in finance team." Maximum 12 words.
- Status: one of GREEN / YELLOW / RED. GREEN means on track for renewal. YELLOW means watching. RED means actively at risk.

After the per-account lines, add a "MANAGER ATTENTION" section listing the accounts (max 3) the manager should ask about first, with one sentence why.

Constraints:
- Sort by ARR band (highest first), then by status (RED first within band).
- "What changed" must be a change, not a summary of state. If nothing changed, omit the account from the digest.
- Do not use the word "good" in any status description.

ACTIVITY LOG:
[paste CRM activity, recent emails, usage data summaries]

What good output looks like: A scannable table-style output, longest accounts first, with a manager-attention section at the bottom. Status words used exactly (Green/Yellow/Red), no hedging.

Failure modes:

• Model lists every account whether or not anything changed. Re-prompt explicitly: "omit accounts where nothing changed this week."
• Status words drift toward hedging — "trending green," "leaning yellow." Lock them: "one of exactly GREEN, YELLOW, RED."
• Manager-attention section becomes 5-7 items instead of 3. Cap it explicitly.

Variation hook: If your portfolio is consistently >30 accounts, change the per-account format to two lines (account + status on line 1, change on line 2) for readability.

Section B — Augment

These produce customer-facing drafts. The CSM is the final author. Human reviews and edits every output before it ships.

Prompt 4: Renewal-proposal email draft

Quadrant: Augment
When to use: 60 days from renewal. You need to send the initial proposal email. You want three drafts at different formality levels to pick from.

You are drafting a renewal proposal email for a CSM to send to their primary stakeholder at a customer account.

Read the context below. Produce THREE drafts of the email, labeled:
- DRAFT A: Formal-direct (for a CFO or procurement-led decision)
- DRAFT B: Warm-direct (for a longstanding champion relationship)
- DRAFT C: Brief-collaborative (for accounts with strong day-to-day engagement)

Each draft must:
- Be under 120 words
- Open with one specific reference to the customer's situation or recent interaction (not "I hope this finds you well")
- State the renewal terms clearly: dates, scope, price
- Propose a specific next step with a specific timeframe
- Close without marketing language

DO NOT use any of these phrases in any draft:
- "I hope this finds you well"
- "as discussed"
- "exciting opportunity"
- "we value our partnership"
- "looking forward to"

CONTEXT:
[paste account name, current contract terms, proposed renewal terms, stakeholder name and recent interactions, any relevant relationship context]

What good output looks like: Three labeled drafts, each genuinely different in tone (not just word substitutions). Each under 120 words. None contain the banned phrases.

Failure modes:

• All three drafts sound the same with mild tone shifts. Re-prompt with: "the three drafts should feel like they could be from three different people."
• The banned phrases creep back in disguised forms ("I trust this finds you well" instead of "I hope this finds you well"). Read carefully.
• Drafts over 120 words. Cap is real; enforce on re-run.

Variation hook: For accounts where pricing changes are part of the renewal, add a constraint: "each draft must reference the pricing change once, neither apologetically nor defensively."

Prompt 5: Sponsor-departure response

Quadrant: Augment
When to use: Your champion just emailed you saying they're leaving the company. Your response needs to do three things at once: congratulate them, ask for a warm handoff, and not try to extract a reference call.

You are drafting a response to an email from a CSM's customer champion announcing they are leaving the company.

The response must:
1. Genuinely congratulate them on the new role (assume the new role is mentioned in their email; if not, congratulate the move).
2. Ask for a warm handoff to whoever is taking over the relationship, framed as helpful to them and their company, not extractive of you.
3. Acknowledge the relationship without performing it.

The response must NOT:
- Ask for a reference call (we will not ask the departing champion for a reference, ever).
- Express disappointment, anxiety, or self-interest about the renewal.
- Use the phrase "stay in touch."
- Promise a coffee, lunch, or call "to celebrate."

Format:
- Single email, under 100 words.
- Tone: warm, brief, professional. Not effusive.

CONTEXT:
[paste the champion's departure email, plus 1-2 sentences about how long they've been your sponsor]

What good output looks like: A single short email. Genuine congratulations, a clean handoff ask framed around the customer's transition (not your renewal), and no extraction attempts.

Failure modes:

• The model will try to ask for a reference call anyway, sometimes disguised ("I'd love to keep you as a reference point for our product"). Read for any extractive ask and remove.
• The model adds "let's stay in touch" 80% of the time. Remove.
• The handoff ask gets framed around the CSM's renewal needs rather than the customer's transition. Reframe to be customer-helpful.

Variation hook: If the champion is leaving for a competitor's customer, omit the warm-handoff ask entirely; the response is shorter.

Prompt 6: Pricing-pushback response

Quadrant: Augment
When to use: Customer pushed back on a renewal price increase. You need a response that acknowledges the friction, restates the value framework, and proposes a structural alternative instead of a discount.

You are drafting a response to a customer who has pushed back on a renewal price increase.

The response must propose ONE structural alternative — not a discount. Structural alternatives include:
- Term length change (multi-year for the lower rate)
- Scope change (drop a module they're not using)
- Payment cadence change (annual vs quarterly billing)
- Effective date change (delay renewal by 60 days to align with their fiscal year)

The response must NOT:
- Offer a percentage discount or dollar concession
- Use the phrase "let me see what we can do"
- Concede before naming the alternative

Format:
- Email body, 100-150 words.
- Open by acknowledging their specific objection (not "thanks for the feedback").
- State the alternative in one paragraph.
- Close by inviting a 20-minute conversation specifically about the alternative.

CONTEXT:
[paste the customer's pushback email, current pricing, proposed pricing, and any context about why the increase is happening]

What good output looks like: A focused email that takes the conversation off the discount axis and onto a structural axis. The alternative is specific (named term length or named module to drop), not vague.

Failure modes:

• Model offers a discount anyway. ("We could potentially explore a small adjustment to the pricing.") Catch and remove every time.
• Model proposes multiple alternatives, diluting the move. The prompt says one alternative. Pick the strongest and delete the rest.
• "Let me see what I can do" or its variants. Banned. Find and remove.

Variation hook: If the customer's pushback explicitly references competitor pricing, add: "do not engage with the competitor pricing reference directly; redirect to value delivered against their specific outcomes."

Section C — Inform

These produce inputs to a human decision. The CSM is the decision-maker. The model is a thinking aid.

Prompt 7: Renewal-forecast challenge

Quadrant: Inform
When to use: You've placed an account at 85% likely to renew. Before the forecast goes to your manager, you want the strongest possible case against your estimate.

You are stress-testing a CSM's renewal forecast for a specific account.

The CSM has scored this account at [X]% likely to renew. Your job is to argue the opposite case — make the strongest argument you can that the actual probability is significantly lower (or higher, if asked to argue up).

Read the account context below. Then produce:

1. THE COUNTERCASE: 3-4 specific reasons the actual probability could be 15-25 percentage points lower than the CSM's estimate. Each reason must reference something specific in the context, not generic risk factors.

2. THE EVIDENCE THE CSM IS UNDERWEIGHTING: 2 things in the account context that the CSM may have noted but not fully internalized.

3. THE SINGLE STRONGEST POINT: of all the reasons above, which one would you put on the page if you only had one shot to change the CSM's forecast?

Constraints:
- Be specific. "The champion could leave" is a generic risk; "the champion's LinkedIn shows recent activity around a new role" is specific.
- Do not produce a "balanced" view. Your job is the countercase.
- If the case is genuinely strong, say so in the final point: "On reflection, the CSM's estimate looks defensible because..."

ACCOUNT CONTEXT:
[paste account summary, recent interactions, usage data, stakeholder map, current forecast and reasoning]

What good output looks like: A focused contrarian view. Three to four specific reasons grounded in the account context. A single strongest point at the end. If the case is weak, the model says so.

Failure modes:

• Model produces generic risks (champion could leave, budget could be cut) that aren't tied to the specific context. Re-prompt: "every reason must reference something specific in the context provided."
• Model hedges and produces a "balanced" view instead of the countercase. Be explicit: "this is a steel-manning exercise, not a balanced one."
• Model only argues down. Flip and run it again with "argue the actual probability could be higher than estimated" to triangulate.

Variation hook: For high-stakes forecasts ($1M+ ARR), run this prompt three times with three different framings (champion-departure risk, competitive risk, internal-restructure risk) for a fuller stress test.

Prompt 8: Vendor-pitch deconstructor

Quadrant: Inform
When to use: A CS tooling vendor sent you their deck or one-pager. You want to translate vendor language into operator language before the demo.

You are translating a vendor pitch into operator language for a CSM evaluating a tool.

Read the vendor material below. Produce:

1. WHAT IT ACTUALLY DOES: in 50 words or less, describe the tool's actual function. Strip every marketing word ("revolutionize," "AI-powered," "seamless"). Describe what the software literally does.

2. WHAT IT DOES NOT DO: 2-3 things the deck implies but does not commit to. Look for: "supports," "integrates with," "designed for" — these are softer than "does."

3. THE DATA QUESTION: what data does the tool require from us, and what data does it send out to third parties or its own backend?

4. THE FAILURE CATEGORY: based on the pitch, which failure mode is most likely? Choose from:
   - Vaporware (the feature described doesn't fully exist yet)
   - Pre-trained on generic data (the AI is not trained on your business)
   - Black box (you can't see why it produces what it produces)
   - Workflow lock-in (the tool only works if you adopt their workflow, not yours)
   - Slow time-to-value (the tool requires extensive setup before producing output)

5. THE QUESTION FOR THE DEMO: one specific question the CSM should ask in the demo that will reveal whether the failure mode applies.

Constraints:
- Be ruthless. Vendor decks are aspirational; your job is to find the gap between aspiration and reality.
- If the deck is unusually honest, say so: "this deck makes specific, falsifiable claims; the demo question is straightforward verification."

VENDOR MATERIAL:
[paste deck text, one-pager, or product page content]

What good output looks like: A short, harsh translation. The "what it actually does" section reads like a product spec, not marketing. The failure category is named explicitly. The demo question is specific and verifiable.

Failure modes:

• Model softens its translation toward marketing. Re-prompt: "strip every adjective that isn't a measurable property."
• Model defaults to "Black box" as the failure category for every tool. Force a specific choice: "name the single most likely failure based on what's actually in this material."
• The demo question is generic ("can you walk me through the workflow?"). Force specificity: "the demo question must be answerable with a yes, no, or specific number."

Variation hook: When evaluating multiple competing tools, run this prompt against each vendor's material separately, then compare the "WHAT IT ACTUALLY DOES" sections side by side. The honest tools' descriptions sound similar; the aspirational ones diverge.

The patterns

After eight prompts, the patterns are visible. Five rules that make any CS prompt work:

Specify the audience inside the prompt, twice. Once for the model ("you are writing for a CSM"), once for the output ("the recipient is a CFO who has not seen the product"). The model's default audience is "general business reader" — which is the wrong audience for every CS task.

Constrain the length explicitly. Models default to 300-word emails. Real CS emails are 80–120 words. Tell the model the cap and it will respect it 90% of the time.

Name the failure mode you want to avoid. "Do not use phrases like 'I hope this finds you well' or 'as discussed.'" Models will use them anyway sometimes, but less often when told not to.

Ask for multiple drafts at varying styles, not "just make it good." Three drafts at different tones give you a range to pick from. One draft locks you into the model's default voice.

Tell the model what to do with the context, not just to read it. Don't paste context and ask for output. Paste context and say: "use this to identify the underlying objection, then write a draft that addresses it without naming it explicitly."

The library is yours to modify

These eight are a starting kit, not the complete library. Operators should adapt them to their voice, their company's positioning, their customers' language. The companion Google Doc is set up for exactly that — every prompt has a "your version" space underneath where you save the version you actually use.

The doc also includes a meta-prompt: a prompt for building new prompts in this format. Give it a CS task, and it returns a prompt with the audience, the constraints, the failure modes, and the variation hook already structured. That's how the library grows.

A working copy stays private to you. Build your library. The pattern compounds — your tenth prompt is better than your first because you've internalized what the structure is doing.

What's next

The prompts in this library cover the Automate, Augment, and Inform quadrants of the Decision Matrix. What they deliberately don't cover is the question of how to do this safely — what counts as customer data, what models will do with what you paste in, and where the lines are for customer-facing AI work.

The next piece, AI Safety for Customer-Facing Work, is the operating floor underneath this library. Eight rules that govern how you use any of the prompts above. The matrix tells you what to use AI for. The safety rules tell you how to use it without ending up in your General Counsel's inbox.

Download the full Prompt Library — eight copy-paste prompts plus a meta-prompt for building your own.