# Cold-email reset kit

Cerrito — original framework. Unpublished draft, September 25, 2026.
This is a test design, not a winning campaign or a guarantee. No outreach was run to produce this kit.

## 1. Research before writing

Audience / role / company situation:
Offer and useful next step:
Exclusions (stage, role, incompatible situation, previous opt-out):

For each account:
- Verified public event:
- Source URL and publication/retrieval date:
- Exact relevant observation, in your own words:
- Inference about the business problem (label as inference):
- Why this role might own the problem:
- Evidence against fit:
- What remains UNKNOWN:
- Reviewer and decision: CONTACT / EXCLUDE / NEEDS EVIDENCE

Do not infer budget from funding, pain from headcount, or a broken process from a job opening. Do not claim to have made an audit or artifact unless it exists.

### Reusable research prompt

Using only the supplied public sources:
- Find evidence this account has the problem our offer addresses.
- Return the source URL and date for every factual claim.
- Separate observation from inference.
- List reasons this account may be a poor fit.
- If evidence is missing, return UNKNOWN. Do not write an opener.

Human review: open the sources; check whether the evidence supports the actual message. A model-generated citation is not verification.

## 2. The one-line summary test

If an inbox assistant reduced the message to one sentence, what would it say?

- Recipient's relevant problem:
- Useful thing offered:
- Reason it matters now:
- Single next step:

Remove decorative praise, unverifiable claims and unnecessary asks. This is a clarity exercise, not an AI-ranking optimization technique.

## 3. Bounded experiment card

Hypothesis:
One variable being tested:
Account selection / exclusions:
Allocation method (prefer random assignment where feasible):
Keep each account in one variant; do not send both variants to the same person or company.
Variables held comparable (role, stage, trigger, offer unless tested, schedule, follow-ups):
Maximum accounts / total send cap / total cost cap:
Fixed observation window after first contact:
Date all included prospects will have completed that window:
Primary metric (define before launch):
Positive-reply classification:
Economic floor (maximum fully loaded cost per qualified attended meeting, or other business-specific criterion):
Stop conditions for delivery problems, complaints, bad targeting, inaccurate claims:
Owner for replies and opt-out suppression:

A small pilot finds obvious problems; it may be underpowered to distinguish plausible performance differences. Preserve counts, uncertainty and the observation window. Do not keep expanding a losing test to chase an arbitrary reply rate.

## 4. Outcome scorecard

Use one record per cohort/variant. Record source system and export time.

P = distinct prospects contacted:
S = messages sent, including follow-ups:
B = confirmed bounced messages (exclude temporary delay notices):
H = human reply messages, including repeated replies:
R = distinct human respondents:
I = distinct interested prospects under the pre-written rule:
M = distinct prospects booking a meeting:
A = distinct prospects attending a meeting:
Q = qualified opportunities:
W = won customers, when mature:
Cash spend (software/data/other):
Operator hours (research/review/sending/reply handling):
Chosen hourly cost assumption:

- Human reply-message yield = H / S × 100
- Prospect response = R / P × 100
- Interested prospects = I / P × 100
- Attended prospects = A / P × 100
- Cash cost per attended prospect = cash spend / A
- Fully loaded cost per attended prospect = (cash spend + hours × hourly cost) / A

Zero denominator: report not applicable. Unmatured cohort: report pending. Do not substitute message counts for unique people. Separate auto-replies, opt-outs, negative replies, referrals and ambiguous replies from positive interest. Delivered means only what your source actually measures, not proof of being read.

## 5. Keep / change / stop

KEEP TESTING: comparable mature cohorts produce qualified outcomes within the economic floor. Repeat before expanding; no universal percentage guarantees success.
CHANGE: response categories expose one fixable issue. Name that issue and change one variable.
PAUSE / STOP: technical failures, complaints, inaccurate assumptions or repeated mature cohorts fail the predefined floor. Fix the issue or redirect the effort.

Current decision:
Evidence and counts:
Uncertainty / alternative explanations:
Next action and owner:
Next review date:

## Research context

The accompanying article links the primary benchmark and provider sources. Those are commercial samples with different definitions; this kit does not combine them into an industry average.