playbook-fundraising

name

playbook-fundraising

description

Produces a copy-ready sentence about a company's funding round, and the list-building filters that select companies which have ever raised or raised recently. Triggers on "raised a round", "recently funded", "companies that raised", "Series A list", "VC backed", "post funding outreach", "who just raised". Outputs funding_line, a COMPLETE sentence including the leading "Saw" and the trailing period, so an empty value renders as nothing at all.

Playbook: Fundraising

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 angle depends on the prospect having money to spend or a mandate to grow — either a TAM filtered to funded companies, or a first line that names their round.

Do not use when: you want headcount growth or open roles as the signal (playbook-hiring-surge). If you want the round as a pure list filter and no copy at all, you only need lane A in §3.

One-line output: funding_line = "Saw you raised $52M in the Series B." — the whole sentence lives in the variable. §2's downstream gate explains why that matters more than it looks.

1. Trigger and scope

Three separate questions get called "fundraising", and they route to different sources. Do not mix them up: one is a solved free problem and one is not solved at all.

• Lane A, ever raised. "Only show me companies with venture backing." A search filter applied when the list is built. Free, reliable, no per-row work.
• Lane B, the copy line. "Write a first line about their round." A per-row lookup producing funding_line. This is the main deliverable and the tested path.
• Lane C, raised in the last 30 days. "Who closed this month?" The people-database funding indexes lag 3 to 4 months, measured. Lane C has no proven end-to-end path and is documented honestly below rather than faked.

Not covered: valuation claims, investor-name personalization, and anything reported only as a rumor.

2. Output contract

Inputs required per row

Field	Type	Required?
domain (bare, lowercase, no www, no scheme)	string	yes
company_name	string	no, used only for the abstain log and QA

Output fields

Field	Type	Example	Max	Null?
funding_line	string	Saw you raised $52M in the Series B.	90	no, use the abstain value
funding_clause	string	you raised $52M in the Series B	80	no, use ""
funding_evidence_url	string	a funding-round URL	300	no, use ""
funding_confidence	enum	high / low	4	no

Abstain value: "". funding_confidence is low whenever the line is empty.

Coverage expectation

8/10 rows usable (80%), where "usable" counts a correct abstain on a company that genuinely has no recent round.

Read the second number too: only 2/10 rows produced a non-empty line, because the test set was adversarial — three bootstrapped companies and three whose last round is years old. On a list already filtered through lane A with a 365-day funding window, every row has a funding event by construction, so the non-empty rate is governed by the equity-stage filter and the 12-month window rather than by source coverage. That rate was not measured. Measure it on your first real campaign before letting copy depend on the variable being present.

Copy-fit rules

Two fields, two different jobs. Keep them straight.

• funding_clause is what the model writes: lowercase first letter, no trailing period, under 80 characters, 5th-grade reading level, no em dashes. It must read grammatically inside Saw <funding_clause>.
• funding_line is what gets pushed: deterministic code wraps the clause into "Saw " + funding_clause + ".". When the clause is "", the line is "" too. The model never writes this field, a formula does, so the wrapping cannot be forgotten or hallucinated.
• Neither field ever names a month, a season, a year, or a date, and neither names total funding, valuation, or investors. §7 explains why, with measured error rates.

Downstream gate — read this before wiring the email body

⚠️ Spintax is a RANDOM chooser, not a conditional.

{Saw {{funding_line}}. |} picks one of its branches at random on every row, with no knowledge of whether the variable is populated. On rows that abstain it renders Saw . about half the time, and on rows that DO have a line it throws the personalization away about half the time.

Spintax can never be used as an if-populated gate for any variable, in this playbook or any other. This generalizes past fundraising and is probably the single most useful thing in this file.

There are exactly two mechanisms that work. Pick one per campaign:

1. Default: the self-erasing variable. Put the entire sentence in the variable — which is why funding_line carries the leading "Saw" and the trailing period. The body line is just {{funding_line}} on its own line. An empty variable renders nothing, so an abstaining row simply loses that sentence and the rest of the email is untouched. QA step: after upload, preview one abstaining lead and confirm the paragraph collapses cleanly without leaving a double blank line.
2. Split at upload time. Partition the list into populated and empty, and upload two campaigns or two variants — one with the funding sentence hard-coded into the body, one without. Use this whenever the funding angle changes the whole email rather than one line, and whenever the non-empty rate is low enough that a single body would read oddly for most recipients.

Only exclude abstaining rows entirely when the campaign angle is the round.

The usual word-level spintax block for wording variety is unrelated and still required. Spintax is for wording variety, never for conditional content.

3. Source chain (cost-tagged)

Lane A, ever raised / raised within 6 to 12 months (list building)

#	Source	Cost	Call	Hit rate
1	A people/company search API's funding filter	FREE	add a company_funding.stage array of equity stages to your ICP filters; for recency add a funding_date of 365 or 180	76% of a US software 50-500 segment had raised at any stage, 65% at an equity stage

Equity stage set: Pre seed, Seed, Series unknown, Series A through Series E-J, Angel, Corporate round, Convertible note, Equity crowdfunding.

Lane B, the copy line (per row) — THE TESTED PATH

#	Source	Cost	What it does	Result
1	Company lookup by domain, read the funding block	FREE	{"filters":{"company":{"websites":{"include":["acme.com"]}}},"page":1} with a browser User-Agent	6/10 domains returned a funding record, 9/10 resolved to the correct company
2	domain-equality guard (code)	FREE	assert the returned company.domain equals the requested domain	removed 1 wrong company out of 10. Mandatory, never skip
3	equity-stage and 12-month filter (code)	FREE	drop Secondary market, Private equity, Debt financing, Grant, Undisclosed, Non equity assistance, all Post IPO *, Product crowdfunding. Then require the newest survivor to be ≤12 months old	removed 3 misleading "latest rounds" out of 6. Mandatory
4	Model phrasing	CHEAP, $0.22/1k	§6. Wording only, never arithmetic or eligibility	2/2 eligible rows produced clean copy

⛔ There is no step 5. If the lookup returned no company record at all, the row abstains and the line is empty, which is the correct outcome for a signal you cannot evidence. Do not add a paid company-enrichment fallback: it costs roughly 8 credits per company and does not carry funding fields anyway.

Lane C, raised in the last 30 days — NOT SOLVED

#	Source	Cost	Reality
1	A funding-signals feed	cheap, per record	dates ran through yesterday, so freshness is real. But only 9 of 25 rows carried a clean company name, and none carried a domain
2	Model cleanup plus domain resolution	CHEAP	read company name + source URL + round type + amount, return {company_name, is_closed_round, is_acquisition}, then resolve name→domain
3	A web-research "find all" API	EXPENSIVE	not tested. Write a stated reason into the run notes before firing it, and cap the run
4	A funding database's UI export, by hand	expensive (seat)	reliable but manual. One-off urgent lists only, never a recurring campaign

Rejected alternatives

• A people-database funding-date filter for lane C. Measured: a 90-day window over the entire US Software Development market returned total_count = 1. The index lags 3 to 4 months.
• Semantic "entity search" APIs for lane C. 0 of 20 returned companies had a round inside the requested window. It returned public companies, companies its own payload labelled Unfunded, and one labelled Acqui-Hired. Entity search ranks on static attributes and has no event-date awareness, so a date-bounded objective degrades into "companies that look fundable".
• total_funding as a copy fact. One company came back at $550.9M where the press reported $452M, because these totals fold in debt facilities and secondary sales.
• raised_at as a copy fact. Off by 15 days for one company, one month for another, two months for a third.

4. Verification

VERDICT: PASS 8/10 (80% usable) | claims 8/8 (100% correct) | chain = company lookup by domain → domain-equality guard → equity-stage + 12-month filter → model line writer | p50 3.0s/row | ~$0.22/1k.

This is a claim-bearing playbook, so it carries two bars: 70% usable AND 90% claim correctness.

The 90% is coverage correctness, not coverage — scored as true claims / rows where a claim was emitted. A company that never raised, correctly abstained on, is a correct outcome and leaves the denominator entirely. It is not a miss. The bar binds on truthfulness when the signal exists, because a fabricated round is the one failure nobody can fix after it sends.

Both misses above were coverage misses (no record for one domain; a two-month date error pushed one round outside the window), not false claims — which is why claim correctness is 8/8.

⚠️ Lane A's end-to-end list quality was not graded. Lane C is untested end to end and must not be described as working.

5. Clay implementation

• clay-table.md — the 7-column build.
• clay-workflow.md — the CLI-buildable version.

6. Locked prompt

Model: the one this prompt was graded on. Params: max_completion_tokens=3000 (never max_tokens on a reasoning model), no temperature, response_format={"type":"json_object"}, flex tier for batch.

⚠️ No prompt-cache discount applies here. The static prefix is ~520 tokens and caching engages at 1,024+. cached_tokens was 0 on all 20 measured calls. Do not budget for a cache discount.

The prompt's JSON key is funding_line because that is byte-for-byte what was graded — but it holds the clause. Code wraps it into the pushed sentence. Do not edit the key to "fix" the naming; that silently changes the tested prompt.

You write one short clause about a funding round for a cold email.

You will be given a JSON record for one company. Every fact in it has already been
checked. Your only job is wording. Do not do arithmetic and do not add facts.

Return JSON only, exactly these keys:
{"funding_line": "...", "evidence_url": "...", "confidence": "high|low"}

Rules:
- If eligible is false, return "" for funding_line and "low" for confidence. Nothing else.
- The clause must read grammatically inside this sentence: "Saw <funding_line>."
- Start with a lowercase letter. No trailing period. No em dashes. No quote marks.
- 5th-grade reading level. Under 80 characters.
- Use amount exactly as written in the record. Never change the number.
- If stage is not empty, name it. If stage is empty, say "round" instead.
- Never name a month, a season, a year, or a date. Never say "recently".
- Never mention total funding, valuation, or investors.
- Copy evidence_url from the record. Set confidence to the value in the record.

Examples:
Input: {"company":"Attio","amount":"$52M","stage":"Series B","eligible":true,"evidence_url":"https://www.crunchbase.com/funding_round/attio-series-b","confidence":"high"}
Output: {"funding_line":"you raised $52M in the Series B","evidence_url":"https://www.crunchbase.com/funding_round/attio-series-b","confidence":"high"}
Input: {"company":"Deel","amount":"$300M","stage":"","eligible":true,"evidence_url":"https://www.crunchbase.com/organization/deel","confidence":"high"}
Output: {"funding_line":"you closed a $300M round","evidence_url":"https://www.crunchbase.com/organization/deel","confidence":"high"}
Input: {"company":"Northwind Labs","amount":"","stage":"Series A","eligible":true,"evidence_url":"https://www.crunchbase.com/organization/northwind-labs","confidence":"high"}
Output: {"funding_line":"you closed the Series A","evidence_url":"https://www.crunchbase.com/organization/northwind-labs","confidence":"high"}
Input: {"company":"Acme Widgets","amount":"","stage":"","eligible":false,"evidence_url":"","confidence":"low"}
Output: {"funding_line":"","evidence_url":"","confidence":"low"}

PER-ROW DATA (appended last, as the user message)
<the eligibility record>

Verifier pass: not needed, and deliberately so. The usual reason to add one is that the model might hallucinate a date or an amount. Here it is structurally incapable of it: eligibility, the 12-month arithmetic, the stage filter and the rounding all happen in code before the call, and the prompt forbids dates, totals, valuations and investors outright. The only facts in play are a stage string and a pre-formatted amount string, both copied verbatim. Spend the verifier budget on lane C instead, where headline parsing genuinely can invent a round.

Truncation guard: finish_reason=length with empty content means retry, never abstain. A reasoning model burns reasoning tokens before emitting content and can hit the ceiling on a trivial prompt. Retry up to 3 times.

What an unlocked prompt looks like: the v1 of this prompt scored 4/10 and shipped you closed the Series E-J back in October and you have raised north of $550M to date. Both are exactly the failures the locked rules now forbid.

7. Edge cases and failure modes

Symptom	Cause	Fix
Lookup returns a completely different company	A website filter matches other listed websites too, so a vendor that lists your target as a customer wins the match. Observed live, with a total count of 1 and no warning	Assert the returned domain equals the requested domain. There is often no strict-domain filter available — a company_domain filter 400s
A first line calls a secondary sale a fundraise	The funding-events array includes Secondary market rows, sometimes with a large recent amount. One company had a $300,000,000 secondary	Filter to the equity stage set before sorting for the latest event. Existing shareholders selling is not money the company raised, and a founder will notice
A first line calls a PE minority stake a fundraise	Databases record a PE firm's purchase of a stake as a Private equity round	Exclude Private equity by default. Buying shares from existing holders is not money into the company, so calling it a raise is a false claim. Toggleable per campaign for a PE-adjacent offer; note the override
Copy contains Series E-J	That is a database bucket label for late-stage rounds, not a real round name	Blank it and fall back to the word "round". Same for Series unknown
The month in the copy is wrong	raised_at is systematically early. Measured: off 15 days, one month, and two months on three companies	Never put a month or a date in the line. The locked prompt forbids it
The total raised does not match the press	Database totals fold in debt facilities and secondary sales	Never quote total funding. Quote the round amount, which checked out exactly on 4 of 4
A company that clearly raised gets an empty line	No record at that domain, or the round is newer than the index (3 to 4 month lag)	Expected. Abstain is correct, and the whole-sentence variable renders as nothing. Do not "fix" it by loosening the guard
An email goes out reading Saw .	Somebody wrapped the clause in spintax as an if-populated gate

Hard rules

• Pace company lookups at about 1 request per second. Zero rate-limit errors were observed at 0.8 req/s across 37 requests.
• Funding-signal feeds usually charge per record returned. A limit: 200 call spends 200 credits.
• Spintax is a random chooser and must never gate a personalization variable.
• Never ship a model you have not graded on this prompt. A cheaper model may look better on paper; if the abstain rule, the character cap, the lowercase-first rule and the no-date rule were never verified on it, it is untested. A $0.11 saving per 1,000 rows is not worth that.