A prompt is something you type once. A skill is a file you write once and reuse, an instruction set that teaches an assistant how your team wants a specific job done, loaded whenever that job comes up.
For outbound, the difference matters because the same handful of tasks recur weekly: defining who to target, reviewing copy before it goes out, triaging what comes back. Writing those instructions properly once beats re-explaining them every Monday.
Below are five, as plain markdown you can copy. They work in Claude, and the same text works as a system prompt or reusable instruction block in most assistants.
1. ICP Definer
Use when: you are about to build a list and your targeting criteria live in someone’s head rather than on paper.
# ICP Definer
When asked to define or refine an ICP, produce exactly these sections.
## Firmographic
Industry, headcount band, revenue band, geography, company stage.
Ask for the attribute if it is missing. Do not guess.
## Qualifying condition
The one thing that must be true for this company to need the product.
Not "they do outbound". Something checkable, like "they run more than
20 sending mailboxes" or "they have 3+ SDRs."
## Persona
Job titles that buy, titles that influence, titles that use.
Note which can sign.
## Disqualifiers
Explicit exclusions: company types, sizes or situations that look
like a fit and are not.
## Evidence
For each criterion, state whether it came from the user, from
closed-won data, or is an assumption. Flag assumptions clearly.
## Rules
- Never invent criteria to fill a section. Ask.
- If the user's stated ICP conflicts with their closed-won data,
say so plainly.
- Keep the whole output under 300 words.The “Evidence” section is what makes this useful. Most ICP documents mix observed fact with wishful thinking and never distinguish them.
2. Cold Email Refiner
Use when: reviewing a sequence before launch.
# Cold Email Refiner
Review the supplied cold email against these eight rules.
For each, state PASS or FAIL with one line of reasoning.
Then give one rewritten version.
1. First line is about the recipient, not the sender.
FAIL if it opens with "I" or the company name.
2. Under 120 words.
3. One ask, not two.
4. No more than one link. Zero is better in a first touch.
5. Every merge field has a fallback that reads naturally.
6. The message still makes sense if every merge field is empty.
7. No unverifiable claims (percentages, rankings, "leading").
8. Subject line under 6 words, lowercase, no colon.
After the rewrite, state the fill-rate assumption: which merge
fields must populate for this to work, and what happens if they
do not.
Do not soften the assessment. A FAIL is more useful than a pass.Rule 6 is the one that catches the most. Covered at length in personalisation that fails safely.
3. Signal Triage
Use when: a batch of signal-sourced leads arrives and you need to decide which are worth working.
# Signal Triage
For each lead with an attached signal, output one line:
LEAD | SIGNAL TYPE | STRENGTH | WINDOW | ACTION
Strength:
- STRONG: job change to an ICP-fit company, promotion where the
person previously evaluated us, hiring surge with a leadership change
- MODERATE: single hiring signal, competitor engagement, topic post
from someone senior
- WEAK: single post engagement, headcount drift, new search match
Window: days remaining before the signal is stale.
Engagement signals: 14 days. Job postings: 42. Job changes and
promotions: 90.
Action:
- WORK NOW for STRONG, or any signal with under 7 days remaining
- QUEUE for MODERATE
- HOLD for WEAK unless stacked with another independent signal
Flag any lead where two independent signals fired. Note that the
shorter window governs timing.
Do not write outreach copy. Triage only.The last instruction matters, without it, assistants tend to helpfully draft emails you did not ask for.
4. Account Researcher
Use when: preparing for a call, or multi-threading a target account.
# Account Researcher
Given a company, produce a brief of no more than 200 words.
## What they do
One sentence. Plain language, not their marketing copy.
## Why they might need us
Two bullets maximum, tied to something observable, a job posting,
a stated priority, a recent change. If there is nothing observable,
say "no clear trigger found" rather than inventing one.
## Who to contact
2-3 names with titles and why each one. Note which is likely to
own budget.
## Opening angle
One sentence. What the first message should be about.
## What not to say
Anything that would reveal close monitoring, or any claim we
cannot support.
Rules:
- Cite where each fact came from.
- "No clear trigger found" is an acceptable and useful output.
- Never speculate about internal politics or financial position.The “no clear trigger found” permission is deliberate. An assistant that must produce a reason will produce one whether or not it exists.
5. Reply Triage
Use when: working through an inbox.
# Reply Triage
Categorise each reply as exactly one of:
POSITIVE_MEETING (wants to talk, asking for a time
POSITIVE_INFO) interested, wants detail first
QUESTION (specific question about product, price or terms
OBJECTION) stated reason not to proceed
REFERRAL, points to someone else
NOT_NOW (interested, wrong timing, with a stated or implied date
NEGATIVE) clear no
OOO, out of office
UNCLEAR, does not fit the above
For each, output: CATEGORY | SENTIMENT | SUGGESTED OWNER | URGENCY
Route to a human, always:
- Anything categorised UNCLEAR
- Any QUESTION touching security, compliance, data handling,
pricing or contract terms
- Any reply reading as annoyed or hostile
- Any reply from a named target account
Draft responses only for POSITIVE_MEETING, REFERRAL and OOO.
Do not draft for the rest.That routing block is the substance. The categorisation is easy; deciding what not to answer is the part worth writing down.
How to use these
As a file. Save each as a .md and load it when the task comes up.
As a system prompt. Paste the body into whatever instruction field your assistant offers.
Connected to your platform. If your assistant has access to your outbound tool over MCP, these become operational rather than advisory, the ICP definer can build the list, the triage can act on the actual inbox. Agentic outbound covers that shape.
Writing your own
Four things the five above have in common, and they are what makes a skill file work rather than read as a nice idea:
Specify the output format exactly. “Output one line per lead in this format” beats “summarise the leads.”
Give explicit permission to say nothing. “No clear trigger found is an acceptable output” prevents the most common failure, which is helpful invention.
State what not to do. “Do not write outreach copy.” “Do not draft for the rest.” Negative constraints do more work than positive ones.
Include the routing rule, not just the categories. The value is in what gets escalated, not in the labels.
And keep them short. A 2,000-word skill file is a document nobody maintains; a 200-word one gets edited when it is wrong.
The honest limitation
These make an assistant consistent, not correct. A skill file that encodes a bad ICP produces consistently bad lists, faster.
So the first one on the list is first for a reason. Everything downstream (triage, research, copy review) inherits whatever the ICP got wrong.
The short version
Five files covering the repeated work in outbound: defining an ICP with evidence separated from assumption, reviewing copy against eight rules, triaging signals by strength and decay window, researching accounts with permission to find nothing, and categorising replies with explicit escalation rules.
Copy them, edit them, keep them short. The parts that make them work are the exact output format, permission to return nothing, and the negative constraints.
And fix the ICP first, because everything else inherits it.
Next: 12 prompts that run your outbound week, or scoring leads before enrolment.
Frequently asked questions
What is a Claude skill?
A markdown file containing reusable instructions for a specific task, loaded when that task comes up. Unlike a one-off prompt, it is written once and reused, so output stays consistent across people and sessions.
What outbound tasks suit a skill file?
Repeated work with a definable output: ICP definition, cold email review, signal triage, account research and reply categorisation. Anything you explain the same way every week.
How do you write a good skill file?
Specify the exact output format, give explicit permission to return nothing rather than inventing, state what not to do, and include routing rules for what must escalate. Keep it short enough that people actually maintain it.
Do skill files make AI output accurate?
They make it consistent. A skill encoding a flawed ICP produces consistently flawed lists, so the targeting definition has to be right before the downstream skills are worth anything.