Playbook case study page skill

Playbook: Case-Study Page Name-Drop

All rules here are best practice, not law. Override any of them when the campaign calls for it; note the best practice once and move on.

Use when: the campaign wants a first line that proves you looked at the prospect's site, and the prospect publishes customer stories — B2B software, agencies, consultancies, MSPs, staffing, manufacturers with named clients.

Do not use when: you want a tech detail (playbook-tech-on-website), their pricing model (playbook-pricing-page), or an ad they are running (playbook-ad-library).

One-line output: case_study_line = "your work with Intercom on outbound pipeline", rendering as Saw your work with Intercom on outbound pipeline.

1. Trigger and scope

Does this company publicly brag about a named customer, and if so which one, in words you can put in an email today?

The path list is ten, not four. /customers, /case-studies, /customer-stories, /success-stories, /case-study, /clients, /our-work, /portfolio, /testimonials, /stories.

This matters: in the live test one hit came from /testimonials, the ninth path. Any implementation that probes only the first four loses that row to a paid fallback.

It does not judge whether the customer is impressive, does not read individual case-study PDFs, and does not look for customers not published on the prospect's own site. If the client list is private, the correct output is an abstain.

Two things make this signal safe:

1. The model never writes the sentence. It only picks the customer and quotes the page. A script assembles the clause.
2. Every claim must appear verbatim in the page text you fetched, so a hallucinated logo cannot reach an inbox.

2. Output contract

Inputs required per row

Field	Type	Required?
domain (bare, lowercase, no www, no scheme, no path)	string	yes
company_name	string	no, used for logging and page-owner context

Output fields

Field	Type	Example	Null?
case_study_line	string, ≤16 words	your work with Notion on global spend management	no, use ""
client_name	string	Notion	no, use ""
case_study_url	string	https://ramp.com/customers	no, use ""
evidence_quote	string	Notion unified global spend management across 10+ countries	no, use ""
confidence	high / low	high	no
ship_ready	boolean	true	no, defaults false
verified	boolean	true	no, defaults false
reject_reason	string	non-customer-context	"" when accepted

Only ship_ready = true rows get pushed. verified = true with ship_ready = false means the line is real but held for review.

Coverage expectation

6/7 rows usable (86%). Broken down: 5 produced a real line (4 ship-ready, 1 held at low confidence), 1 correctly abstained, 1 wrongly abstained on a carousel-layout page.

⚠️ Plan on the ship-ready number, not the usable number: 4/7 (57%) shipped without an operator touching them.

Expect lower coverage on SMB and trade segments, where most companies publish testimonials with no company named. Treat 50 to 60% as the planning number for a mixed B2B list, and 80%+ only for software and agency segments.

Copy-fit rules

• Slots into Saw {{case_study_line}}.
• The value always begins with the lowercase words your work with , then the customer's name in the customer's own capitalization, then optionally on plus a 2 to 5 word topic.
• No trailing period, no leading capital, no em dashes.
• The prospect's own company name never appears inside the value. The reader knows who they are.
• The value is assembled by a script, not written by the model, so these rules hold by construction rather than by asking a model nicely.

Downstream gate

If empty: drop the clause with spintax and keep the row. Never render Saw . — put the clause in its own optional sentence so the paragraph still reads when blank.

⚠️ confidence = low does NOT auto-ship

Low-confidence rows are held, and strict-quote mode is the default. The coverage cost is accepted on purpose.

A low row is a bare logo with no story attached — which is exactly the shape an investor logo, a press logo, a partner logo, and a CMS placeholder logo all take. The verbatim gates in §6 cannot tell those apart from a customer logo.

So every low row ships ship_ready = false and waits for an operator eyeball. To ship them, read a 20-row sample, confirm they are customers, then enable it deliberately. A low row never carries a topic clause even when it does ship.

3. Source chain (cost-tagged)

#	Source	Cost	Call	Result
1	Direct path guess on the prospect's own origin	FREE	HEAD over all ten paths, browser UA, follow redirects, GET with Range: bytes=0-4095 when HEAD is refused	6/7 found the page here
2	Homepage anchor scan	FREE	parse <a href> and anchor text for case stud, customer stor, success stor, our work, clients, probe the first same-origin match	⚠️ untested, 0 hits. It ran on exactly one domain and returned nothing. It is design intent, not a measured result
3	Page fetch and text flatten	FREE	strip script and style, keep <img alt> as [logo: X], collapse whitespace, cap at 18,000 chars	mechanical
4	Model customer picker	CHEAP ($0.18/1k)	§6 prompt, minimal reasoning effort, JSON mode	5/6 fetched pages yielded a verified customer
5	SERP site:{domain} case study	CHEAP	one simple query, never OR chains	untested
6	Rendering proxy	METERED	only when step 3 returned under 200 characters	untested. Cap it to rows that actually failed, never list-wide
7	An entity-search API	METERED	last resort	untested

Steps 1 to 4 are the tested path and the only ones with a verdict.

Before you escalate, check the list size

Case-study finding is a research task, and for research a strong reasoning subagent generally beats a vendor browsing agent or a database when volume is modest and quality matters. In practice:

• Steps 1 to 4 stay first regardless of size. Ten free HEAD requests found the page on 6 of 7 rows. Nothing outranks free and measured, and a subagent is a worse tool than a HEAD request for a question a HEAD request answers.
• The leftover rows are where it bites. On a modest tail — the handful that fell through on a list of a few hundred — hand those rows to subagents, one per company. A subagent follows the trail when the obvious path 404s, reads the actual page, and abstains honestly; a one-shot vendor agent pads.
• Steps 5 to 7 are the SCALE answer, for when the tail is thousands of rows running unattended.
• Verification does not relax. A subagent's customer name faces the same verbatim gates. If the name is not in the fetched page text, it does not ship. "The agent said so" is not evidence.

Rejected alternatives

• A metered web agent as the finder. It costs credits per row to do what ten free HEAD requests already do at 6/7. Keep it for the leftover rows only.
• LinkedIn scrapers. Customer stories live on the company website, not on LinkedIn.
• Third-party review sites. Their customer lists are not the prospect's own claim, so the opener stops being "I read your site" and becomes "I read a directory" — exactly the tell you are avoiding.
• Asking the model to write the finished sentence. Measured: it lowercased proper nouns, copied a percentage out of the few-shot example onto a different company, and produced ungrammatical clauses.

4. Verification

VERDICT: PASS 6/7 (86% usable, 4/7 ship-ready) | claims 5/5 (100% correct) | ten-path HEAD sweep → model picker at minimal reasoning → deterministic line assembly | p50 2.4s/row | ~$0.18/1k.

Claim-bearing playbook: 70% usable AND 90% claim correctness. The single failure was a wrong abstain on a carousel layout — a coverage miss, which costs volume and nothing else. Naming an investor or a press logo as a customer is the failure the 90% exists to stop, and it is why strict quoting and the low-confidence hold are both defaults rather than options.

Page discovery alone scored 7/7 (6 found, 1 correct negative).

What this verdict does NOT cover

• The Clay recipe. Never run end to end.
• The homepage anchor scan. 0 of 6 discoveries.
• Chain steps 5 to 7. Never executed.
• The non-customer logo class. No row in the sample had an investor wall, a press strip, an integrations grid, or a CMS placeholder as its most prominent entity. The placeholder and context gates were unit-tested against a real page dump and a synthetic investor wall — a code check, not a live campaign result.
• ⚠️ "Verification" that re-runs the same check is not verification. A second pass that re-fetches the same URL with the same UA and runs the same substring check is a reproduction, not a second source. It can only confirm the string is on the page — which the code gate already guarantees — so it structurally cannot detect the partner, press, or placeholder-logo error class.

Read the 6/7 as "6/7 rows produced a line whose name is verifiably printed on the prospect's own page", not as "6/7 rows named a confirmed customer."

5. Clay implementation

• clay-table.md — the column build, including the domain-normalization trap that produces a 100% miss rate with no obvious cause.
• clay-workflow.md — the CLI-buildable version.

6. Locked prompt

Model: gpt-4o-mini inside Clay, a nano-class reasoning model outside. Reason, measured: nano spends its completion budget on hidden reasoning unless minimal effort is passed, and Clay's integration does not expose that parameter. Without it, this exact prompt returned finish_reason=length with empty content on 6 of 6 rows. So nano wins outside Clay on price (~$0.18/1k vs ~$0.50/1k) and mini wins inside Clay on the only thing that matters there: returning content at all.

Params outside Clay: max_completion_tokens=1200, reasoning_effort="minimal", JSON response format, no temperature, flex tier for overnight batches.

You read a company's own case-study, customer-stories, or testimonials web page and pull out ONE named customer they publicly brag about.

You do NOT write the sentence. A script builds the sentence from your fields. Your only job is to pick the right customer and prove it is on the page.

Return JSON only, with exactly these keys:
{"client_name": "...", "evidence_quote": "...", "detail_phrase": "...", "confidence": "high|low"}

What each key means:
- client_name: the name of ONE customer company named on the page, spelled and capitalized exactly as the page spells it. Never the page owner. Never a partner directory, a press outlet, a review site, or a person's name. Pick the customer with the clearest story on the page.
- evidence_quote: 4 to 20 words copied character for character from the page text. It MUST contain client_name. It is what proves the customer is really named there.
- detail_phrase: 2 to 5 plain words naming the TOPIC of the work, and every word of it must appear inside evidence_quote. It is a noun phrase like "global spend management" or "first response time". It is never the customer's name, never a sentence fragment like "time back" or "hours saved", and never ends in a small word like back, up, out, on, of, to, with. If evidence_quote holds no clean topic, return "" here. An empty detail_phrase is a good answer, not a failure.
- confidence: "high" when the page clearly presents that company as a customer, "low" when it is only a logo with no story.

Rules:
- Never invent a customer, a number, or a result. If the page names no customer company, return "" for client_name and "" for the other fields.
- A page of unnamed praise ("Great service. - Dave R.") names no customer company. Return "".
- The examples below show SHAPE ONLY. Never copy a customer name, a number, a percentage, or a phrase out of the examples into your answer.
- detail_phrase never contains the page owner's own name, never contains a number the page does not state, and never sounds like bad news for the customer.
- 5th-grade reading level. No em dashes anywhere in the output.
- Ignore navigation text, cookie banners, menu items, and blog titles. Those are not customer stories.

Examples:
Input: page owner = zendesk.com, page text = "Customer stories. How Uber cut first response time by 30%. Read the Shopify story. [logo: Slack]"
Output: {"client_name": "Uber", "evidence_quote": "How Uber cut first response time by 30%", "detail_phrase": "first response time", "confidence": "high"}
Input: page owner = someagency.com, page text = "Our work. [logo: Peloton] [logo: Casper] [logo: Warby Parker]"
Output: {"client_name": "Peloton", "evidence_quote": "[logo: Peloton]", "detail_phrase": "", "confidence": "low"}
Input: page owner = acmeplumbing.com, page text = "Testimonials. Great service, fast and friendly. - Dave R. Highly recommend! - Sarah T."
Output: {"client_name": "", "evidence_quote": "", "detail_phrase": "", "confidence": "low"}

PER-ROW DATA (appended last)
Page owner company: <company name>
Page owner domain: <domain>
Page URL: <case study url>
Page text:
<page text>

Note the "SHAPE ONLY" rule. It is there because the model copied a percentage out of a few-shot example onto an unrelated company during testing.

The clause is assembled by code

line = "your work with " + client_name
if every content word of detail_phrase appears inside evidence_quote
   and detail_phrase is not the client name
   and detail_phrase does not end in a dangling word (back, up, out, on, of, to, in)
   and detail_phrase contains no finite verb (is, are, helps, helped, ...)
then line = line + " on " + detail_phrase
if line is longer than 16 words, drop the detail

Verifier pass: five code checks, not a second model call

1. client_name appears verbatim in the fetched page text.
2. evidence_quote appears verbatim in full in the fetched page text.
3. evidence_quote contains client_name.
4. client_name is not a CMS placeholder or generic noun (Startup, University, Company, Partner, Client, ...).
5. client_name does not appear only inside investor, press, partner, or integration wording — check a ~600-character window before each mention, cancelled by customer wording in the same window.

Any failure means abstain, after one retry (the model occasionally quotes an adjacent block, or picks a placeholder logo when a real customer sits next to it). Checks 1 to 3 caught a fabricated detail during the live test.

⚠️ Honest limits, read this before trusting a line

Checks 1 to 3 prove the string is ON the page. They do not prove the entity is a CUSTOMER.

An investor wall, a press strip, an integrations grid and a placeholder logo all live on the same page and all pass a substring check. This is not hypothetical: a real captured page served [logo: Startup], [logo: University], [logo: Anthropic], [logo: Google]. Before checks 4 and 5 existed, a model returning client_name = "University" with evidence_quote = "[logo: University]" passed all three checks and shipped as Saw your work with University.

Checks 4 and 5 are reject-only heuristics, so a false positive costs one abstained row and never a bad line. They are not a proof of customerhood, and no live-test row exercised an investor or press wall. That residual risk is exactly why low-confidence rows are held.

Truncation guard: finish_reason=length means retry, never abstain. It is the single most likely failure on this prompt.

7. Edge cases and failure modes

Symptom	Cause	Fix
Every domain returns "no page found" from a script that works by hand	A default runtime User-Agent is being 403'd or challenged	Send a browser User-Agent on every request
A page "exists" at /customers on a company that obviously has none	Soft 404: the origin 302s unknown paths to / and returns 200	Reject any candidate whose final path is /, and reject final paths that no longer match the case-study pattern
/customers lands on a support login or help center	Consumer sites use "customers" to mean account holders	The path pattern check plus the ignore-navigation instruction. Rows that slip through show up as low with no detail
Model returns empty content with finish_reason=length	Reasoning burned the completion budget	Minimal reasoning effort, cap 1200, treat length as a retry. Inside Clay, where the parameter is unavailable, use a mini-class model
The line names a customer not on the page at all	Hallucination	The verbatim gates catch this
The line names an entity that IS on the page but is not a customer — an investor (Saw your work with Sequoia.), a press logo, an integration partner, or a CMS placeholder (Saw your work with University.)	Investor walls, press strips, partner grids and unfilled CMS slots sit on the same page as the customer logos. The verbatim gates cannot catch this	Partially mitigated, not solved. Checks 4 and 5 are heuristics and neither was exercised live. low rows are the exposed class and are held. Before enabling low-confidence shipping for a client, read 20 rows of that client's output yourself
The line carries a result belonging to a different customer	Numbers next to the wrong logo, or the model copying a number out of the few-shot example	The detail is dropped unless every content word appears inside the evidence quote for that same customer
The line reads your work with Air Tutors on Air Tutors	The model returned the customer name as the topic	A code check rejects a detail equal to the client name
The line reads your work with Google on time back	The detail is a sentence fragment	Reject details ending in a dangling word or containing a finite verb
A real case-study page yields nothing (wrong abstain)	Carousel or logo-wall layouts put names and quotes in separate DOM blocks, so no short quote contains both	A known limitation, and strict quoting stays the default. This is a coverage miss, which costs volume only. Lenient quoting widens exactly the non-customer class above, so it stays off until that class has been live-tested
Page text comes back under 200 characters	JS-rendered marketing page	Escalate that row only to a rendering proxy. Never list-wide
Two runs on the same company return different customers	Expected. Any named customer is a valid answer	Not instability. If you need determinism, cache the first accepted value per domain

Hard rules

• Every HTTP request sends a browser User-Agent, or CDN-fronted prospect sites read as missing rather than as blocked.
• Politeness is per origin, not global. These are your targets' web servers, not a vendor API you pay for. Keep per-origin concurrency at 2, never above about 4; a global cap is a blast-radius control on your own egress and can be higher.
• Throughput is fetch-bound, roughly rows × page-seconds / global-concurrency. Raising model concurrency alone does nothing.