Why Claude Starts from Scratch Every Single Time
The Default Problem
Most people prompt Claude the same way every time. Type a prompt, get an output. And every single time, Claude starts from scratch.
No memory of how you work. No idea what good looks like for you. No understanding of your ICP, your voice, your copy rules, or your standards.
The output you get is Claude's default, not yours.
What This Costs You
Every time you get an output you have to heavily edit, you're paying for Claude's defaults instead of your expertise.
The gap isn't between what Claude can do and what a human can do. The gap is between what Claude produces by default and what Claude produces when it knows exactly how you work.
That gap closes with skills.
What a Skill Actually Is
Skills in Plain English
A skill is not code. It's not an app. It's not a prompt in the usual sense.
It's an instructional manual written in plain English that teaches Claude how to do something your way.
Think of a carpenter given a detailed specification for building a chair: every measurement, every joint, every finish. Claude follows that specification exactly, every time, without you explaining it again. That's what a skill is.
Where to Find Skills
Go to claude.ai → Settings → Capabilities → Skills.
You'll see a list of active skills. Click "Go to Customize" to manage them. Skills can be uploaded as a folder or installed from Anthropic's growing free library.
How Skills Load Without Killing Your Context
Skills use three levels so they don't burn your context window:
Level 1 — Frontmatter: Always in Claude's system prompt. Just enough to know when to use the skill.
Level 2 — SKILL.md body: Loaded when the task matches. Contains the full instructions.
Level 3 — Linked files: Additional references Claude navigates only when needed.
You can have dozens of skills active without paying the context cost of all of them simultaneously.
Type 1 - Skill Booster: Fill a Capability Gap
What It Solves
A Skill Booster is for when Claude is genuinely weak at something — not terrible, just not good enough for your standards.
Your skill fills that gap. Claude's default output: generic, no structure, looks like slop. Claude's output with your skill: clean, structured, follows your framework.
This is capability uplift. You're patching a specific weakness with your expertise.
GTM Examples
Cold outreach copy: Ask Claude for a cold email → generic, template-feel, doesn't convert. Add a skill with your actual copy rules, your proven openers, your CTAs, your voice → output that sounds like you and follows the framework that works.
Landing page copy: Claude's default landing pages are structurally weak. Encode your conversion principles, your hierarchy rules, your friction checklist → every page follows your framework without explanation.
ICP research: Claude summarizes companies generically. Encode your scoring criteria, your signal hierarchy, what you actually need to qualify a prospect → consistent outputs every time.
How to Spot a Skill Booster Opportunity
You need a Skill Booster when:
- You consistently heavily edit Claude's output in a specific area
- The same type of task keeps producing generic results
- You have written rules or frameworks you're re-pasting into prompts constantly
The signal: you're compensating for the same gap every time.
Type 2 - Encoded Reference: Teach Claude Your Way of Working
What It Solves
An Encoded Reference skill is different. Here, Claude isn't necessarily bad at the task, but you have a very specific way of getting the output you want.
Your workflow. Your criteria. Your preferences. Your taste.
You're not uplifting a capability. You're encoding how you think.
GTM Examples
LinkedIn posts: Claude can write posts. But your format — specific hook types, sentence structure, CTA patterns, word limits, what you never say — that's yours. Encode it.
Content briefs: Claude can produce briefs. But your specific sections, your ICP targeting logic, your angle framework — encode that.
Meeting notes: Claude can summarize. But you don't want a summary. You want who owns each action item, by when, what priority — nothing else. Encode that.
The difference from a Skill Booster: Claude could produce an acceptable version without the skill. You're making it produce your version.
The Real Value
Encoded Reference skills are where most GTM operators can win fastest. You already have the expertise. You already know what good looks like. You're just not encoding it.
Every time you edit Claude's output to match your standards, you're doing work a skill should be doing for you. The one-time cost of encoding your process eliminates that work permanently.
The Skill File Structure
What Goes in a Skill Folder
Every skill is a folder:
your-skill-name/
├── SKILL.md # Required — the core instruction file
├── scripts/ # Optional — executable code
├── references/ # Optional — docs Claude loads as needed
└── assets/ # Optional — templates, examples
For most GTM skills, you only need SKILL.md. The other folders are for technical workflows that need to run scripts or reference large external documentation.
The Rules That Matter
File name: Must be exactly SKILL.md, case sensitive. No variations accepted.
Folder name: kebab-case only.
✅ linkedin-post-writer
❌ LinkedIn Post Writer
❌ linkedin_post_writer
No README inside the folder. All documentation goes in SKILL.md or references/.
SKILL.md Structure
Every SKILL.md starts with YAML frontmatter, then the instructions:
---
name: your-skill-name
description: What it does. Use when user asks to [specific phrases].
---
# Skill Name
## Instructions
[Your rules here]
## Examples
[Input → output examples]
The frontmatter (between the dashes) is what Claude reads to decide whether to load this skill. Get that right before writing the instructions.
The Description That Makes or Breaks Your Skill
Why the Description Is Everything
The description field in your YAML frontmatter is how Claude decides whether to load your skill. It runs in Claude's system prompt at all times.
One field, two jobs: tell Claude what the skill does, and tell Claude when to use it.
If Claude doesn't know when to use your skill, it doesn't matter how good the instructions are.
What a Good Description Looks Like
❌ Too vague: "Helps with content."
❌ No triggers: "Creates high-quality LinkedIn posts with proper structure and formatting."
✅ Complete: "Writes LinkedIn posts following [your] format and voice. Use when user asks to write a LinkedIn post, create content, draft a post, write a hook, or build a carousel outline."
The test: does the description include specific phrases you — or your team — would actually say when starting this task?
Trigger Phrases Are the Key
Add the exact phrases that should activate your skill. Think about how you actually start these tasks:
- "Write a cold email for..."
- "Draft a LinkedIn post about..."
- "Create a landing page for..."
- "Help me research this prospect..."
Include both formal and casual phrasings. Under 1024 characters total. No angle brackets in the text.
How to Define Your Skill Without Going Generic
The Most Common Mistake
"Write me great website copy" is not a skill. That's a vague instruction.
A skill has structure. Specific rules. Defined constraints. Concrete examples of what good looks like and what to avoid.
Generic instructions produce generic outputs. The skill is only as good as the specificity you put into it.
Use AI to Figure Out What to Teach AI
Before building any skill, ask Claude: "What are the key decisions and rules I should define for a skill that does [X]?"
Claude pulls the structure out of your head. It asks the right questions. It surfaces decisions you didn't know you were making implicitly.
Then you answer with your actual rules, your actual examples, your actual frameworks. That's when the skill becomes powerful — not the description, but the encoded knowledge inside it.
What to Encode in Every GTM Skill
Rules — what Claude must always do (include a hook, keep under X words, always reference the ICP pain point)
Anti-rules — what Claude must never do (specific banned phrases, formats to avoid, mistakes you keep correcting)
Examples — 2–3 examples of output that's right, 1–2 examples of output that's wrong and why
Context — who the output is for, what platform it lives on, what action it should drive
The more specific your rules, the less editing you do on every output.
Evals: Testing Your Skill Against Real Scenarios
The Problem Evals Solve
Without evals, you build a skill, use it a few times, get outputs that seem okay — but you have no data. No proof. No way to know if the skill is actually doing what it was supposed to do.
It feels like a black box.
Evals tests your skill against realistic scenarios. It runs your skill through different situations and tells you exactly where it's working and where it's breaking.
How to Run an Eval
Skills 2.0 builds and tests your skill automatically during creation. It creates a fictional scenario relevant to your skill, runs it through the full workflow, and generates a sample output before you save anything.
You preview the output, verify it matches your standards, then commit the skill.
For ongoing testing: run the skill against 5–10 representative scenarios — easy cases, edge cases, and common failure modes. Note where outputs deviate from your standard. Improve the instructions. Repeat.
What to Test For
Triggering tests — does the skill load on obvious requests? On paraphrased requests? Does it stay quiet on unrelated tasks?
Functional tests — does the output match your standard? Are the rules being followed? Are the anti-patterns absent?
Edge case tests — what happens when the input is ambiguous, incomplete, or unusual? Does the skill handle it or break?
Fail any of these and you have specific instructions to fix, not a vague sense something is off.
Benchmark: Running the With vs Without Comparison
Why Benchmarking Matters More Than You Think
AI models keep improving. Claude today is better than Claude six months ago. That creates a question nobody was asking: when the model improves, does your skill still add value?
Your skill was built to fill gaps in what the old model couldn't do. But what if the new model already handles it on its own? Is your skill still needed — or is it getting in the way?
Benchmark answers this. Every time Claude gets a significant update, you can test whether your skill is still making a difference.
How to Run a Benchmark
Take a real brief you'd normally use. Run it twice:
- Once with your skill active
- Once with no skill, just Claude's raw default
Same brief. Same model. Same moment.
Compare the outputs side by side. The difference is your skill's actual contribution. If the gap is meaningful, your skill is earning its place. If the gap is small, your instructions may be redundant now that Claude handles it natively.
This is proof, not theory. The side-by-side is the benchmark.
Triggers: Making Sure Your Skill Fires
The Silent Failure Mode
You build a LinkedIn post skill. You say "write me a LinkedIn post." You get generic output. You assume the skill isn't working.
The skill might be working perfectly. It just never got called.
When your skill doesn't trigger, you get Claude's raw default — and you have no idea the skill was skipped. This was the biggest failure mode in Skills 1.0. Triggers tell you exactly whether your skill activated for a given prompt.
How to Fix Undertriggering
If your skill isn't firing when it should:
Add more trigger phrases to the description. Include every variation of how you'd phrase this task, formal, casual, abbreviated.
Add domain-specific terms. If your skill is about cold email, include "cold outreach," "first touch," "sequence," "SDR email", not just the obvious phrases.
Cover different verb choices. "Write," "draft," "create," "build" can all mean the same thing to a user. Include them all.
How to Fix Overtriggering
If your skill is loading when it shouldn't:
Add negative conditions to the description: "Do NOT use for [X]."
Be more specific about what the task looks like. "Use when user is writing a LinkedIn post for B2B outreach, NOT for general content questions or research tasks."
Overtriggering wastes context and produces confused outputs. Treat it as seriously as undertriggering.
The GTM Context File Every Operator Needs
What It Is
Before building individual skills, create one foundational file: your GTM context.
This is a single source of truth about your product, your ICP, your positioning, and your standards. Every skill you build references it. Without it, skills produce technically correct but contextually wrong outputs.
Think of it as your marketing CLAUDE.md — the file that tells every other skill who you are, who you serve, and what good looks like.
What to Include
Product overview — what you do, who it's for, the core outcome you deliver
ICP definition — specific firmographics, pain points, triggers, what they're trying to accomplish, what they've already tried
Value proposition — how you're different, what makes you the logical choice, what you never say about competitors
Voice and tone — how you write, what phrases you use, what you avoid, examples of content that sounds like you
Banned phrases — words that should never appear in any output
Success metrics — what a good output actually achieves, specific not vague
Update this file every time your positioning shifts. Every skill that references it stays aligned automatically.
5 GTM Skills Worth Building First
The Priority Stack
Build these five first. Each addresses a task you're doing repeatedly and editing constantly.
1. LinkedIn post writer — encode your format, hook structure, CTA style, length limits, banned phrases, and tone. Input: topic or angle. Output: post ready to publish.
2. Cold outreach writer — encode your copy rules, personalization framework, message lengths by stage, CTAs that convert. Input: prospect context. Output: sequence that sounds like you.
3. ICP research summarizer — encode your scoring criteria, what signals matter, what information you actually need to qualify. Input: company or contact. Output: scored summary in your format.
4. Landing page CRO auditor — encode your conversion principles, friction checklist, what weak copy looks like. Input: page copy. Output: specific recommendations against your framework.
5. Content repurposing engine — encode how you transform long-form content into specific formats (LinkedIn posts, email hooks, short clips). Input: source content. Output: repurposed assets in your format.
The Build Order Logic
Start with the task you do most often and edit most heavily. That's where your skill investment pays back fastest.
One well-built skill used daily is worth more than ten loosely built skills you rarely invoke.
After the first five, look at where you're still re-explaining things to Claude. Each re-explanation is a skill waiting to be built.
Skills for Content and LinkedIn
What to Encode
The most valuable thing in a LinkedIn content skill isn't a general tone instruction. It's the structural rules that make your content yours:
- Which hook formats you use (and which you avoid)
- Your paragraph length and sentence structure rules
- How you handle the proof point section
- Your CTA patterns and which ones you rotate through
- Word and character limits per post type
- Phrases that immediately sound like AI slop and must never appear
How to Extract Your Own Rules
Take your 10 best-performing posts. Paste them into Claude and ask: "What are the structural rules and patterns that appear consistently across these posts? List them as explicit constraints I could encode in a skill."
Claude pulls the implicit rules out of the work you've already done. Review, correct, add anything it missed. That becomes your skill.
This works faster than articulating your style from scratch — and it's grounded in what actually performed, not what you think you do.
What Good Output Looks Like vs Bad
Include at least one complete example of output that's right and one that's wrong.
The "wrong" example is often more valuable than the "right" one. It shows Claude specifically what to avoid, the generic hook, the corporate CTA, the paragraph that could have been written by anyone.
If Claude can see the difference between your standard and the default, the skill does the filtering for you.
Skills for Cold Outreach
What to Encode
A cold outreach skill needs more than copy rules. It needs context:
- Your ICP's specific pain points and the language they use to describe them
- Your offer framed the way it lands, not the way it's written on your website
- Your sequence structure — what each touch does, timing between them
- What triggers which personalization angle (hiring signal, funding, job post, content engagement)
- Your CTA philosophy — what you ask for at each stage
- The exact phrases that make outreach sound templated, so Claude never uses them
Define What Bad Looks Like
Before writing the skill, define the failure modes:
- Template-feel openers ("I came across your profile and was impressed")
- Generic value props that could apply to anyone in the industry
- CTAs that ask for too much too early
- Messages that are three times longer than they should be
Include these in your skill as explicit anti-patterns. Claude needs to know what to avoid as clearly as it knows what to do.
Personalization Rules
The best outreach skills encode a tiered personalization system:
Tier 1 (high value) — fully custom first line using a specific signal (recent post, hiring move, funding announcement)
Tier 2 (mid value) — industry or role-level personalization, specific pain point reference
Tier 3 (volume) — ICP-level personalization, problem-first framing only
Encode which tier applies to which input context so Claude auto-calibrates without you specifying every time.
Skills for Landing Page and Conversion Work
Two Skills, Not One
Landing page work needs two separate skills:
CRO Auditor (Encoded Reference) — Claude reviews existing copy against your conversion framework. You encode your friction checklist, hierarchy principles, what weak CTAs look like, what objections must be addressed and where.
Copy Generator (Skill Booster) — Claude writes copy from scratch following your framework. You encode your structural rules, proof point requirements, above-fold principles, voice guidelines.
Building both means you can generate from scratch or audit what exists — both against your standards.
What to Encode in the Copy Generator
This is where most people underinvest. They describe the output in general terms instead of encoding the actual rules:
- Above-fold structure: what the headline does, what the subheadline adds, what the CTA says and doesn't say
- Proof point placement: where social proof appears and in what format
- Feature-to-benefit translation: how you convert capability descriptions into outcome language
- Length targets by section
- The specific voice — including words you never use (leverage, streamline, revolutionize, game-changing)
Input a product brief. Output: a complete page that doesn't need heavy editing.
What to Encode in the CRO Auditor
The auditor skill is more about your diagnostic framework than your output format:
- What friction looks like in copy (vague CTAs, missing proof, passive language)
- Your hierarchy of what to fix first (above fold > social proof > CTAs > feature sections)
- What specific objections must be handled and on which section of the page
- The difference between a credibility problem and a clarity problem
Input: page copy. Output: prioritised recommendations against your framework, not Claude's generic conversion advice.
