How Power Users Actually Set Up Claude Code

The four things to set up in Claude Code (Context, Rules, Reach, Operation) before running your first prompt.

In 1982, GM shut down its factory in Fremont, California.

The workers there were, by their own union’s admission, the worst in the American auto industry. Absenteeism ran at 20%. Cars rolled off the line with missing parts. Workers left empty bottles inside door panels to rattle and annoy future owners. GM closed it down.

Two years later, Toyota and GM reopened the exact same factory as a joint venture called NUMMI. Same building. Same equipment. They rehired over 85% of the original workforce, including many of the people specifically cited for the strikes and the sabotage.

Within one year, the plant matched Toyota’s best factory in Japan for quality. Absenteeism fell from 20% to 2%. The same workers GM had written off were producing some of the best cars in North America.

The only thing that changed was what Toyota put around them before they started work.

Toyota sent workers to Japan for three weeks before the factory opened. Not to teach them to work harder, to give them context, rules, tools, and a system for how work actually gets done. GM had given them none of that. Toyota gave them everything.

That’s the right approach to turn Claude Code into a powerhouse of a tool for your work.

To do it you need a starter pack that includes Context, Rules, Reach, Operations

---

Context — What Claude needs to know

Foundation files are .md files that live in a /foundation/ folder in your project. Claude reads them before any task runs.

These are different from CLAUDE.md, that’s the next section. Foundation files are the business intelligence layer. What Claude knows about your audience, your voice, your market, and how your customers buy.

I covered the four files in depth in my last post. The short version:

Audience Delight Profile. How your audience actually talks. The words they use in Reddit threads and sales calls. When Claude knows your audience says “slop” not “low-quality output,” and “ship it” not “deploy the solution,” everything it writes sounds like a peer instead of a vendor.

Creator Style. Extracted from your best-performing content. Patterns, not adjectives. Not “we’re conversational” — but “short sentences lead, fragments are welcome for punch, always lead with what the reader can do.” Something Claude can check a draft against.

Market Positioning Map. A living file. What territory you own, what’s contested, what you’ve ceded. Updated monthly. Every skill that touches messaging reads it automatically.

Customer Journey Intelligence. Where people actually stall and convert. If your nurture sequence doesn’t know the specific moment people drop off, it writes emails that address the wrong problem.

One thing worth knowing about how this works: skills don’t load all four files every time. Each skill reads the file headers and loads only what’s relevant. A LinkedIn skill loads Audience Delight and Creator Style. A launch messaging skill loads Market Positioning and Customer Journey. Build a new skill next month and it routes itself, no manual wiring.

One thing the foundation post doesn’t cover: these files decay. Your positioning shifts, a competitor reframes their messaging, your audience picks up new vocabulary. A foundation file that’s six months out of date is actively misleading Claude. Build a scheduled skill that audits them monthly, checks your positioning map against competitor homepages, flags audience language that’s gone stale, surfaces anything that no longer reflects reality. The below is just a snippet to illustrate how this works. I’ll do a full post on how agents keep my foundational layer fresh.

# Skill: Audit Foundation Files

## Steps
1. Read all files in /foundation/
2. For each file, check against live external sources:
   - Audience Delight: scan 3 competitor comment sections.
     Flag any new vocabulary not in the file.
   - Positioning Map: check competitor homepages.
     Flag any new claims that overlap with your owned territory.
   - Creator Style: check your 3 most recent published pieces.
     Flag any patterns that have drifted from the file.
3. Produce a short audit report — what's current, what's stale, what needs updating
4. Save report to /foundation/audit-[date].md
5. Agent to extract patterns, learnings and updates
6. Apply those to .md files in /foundation/

---

Rules — How Claude should behave

Two files do this job.

CLAUDE.md is your operating manual. The rules Claude follows in every session, what you’re working on this quarter, what voice and tone all output defaults to, where the foundation files live, what never to do.

Keep it under 60 lines. HumanLayer’s research on instruction budgets is clear: beyond a certain length, compliance drops. Shorter and specific beats longer and comprehensive.

Here’s a starter template built for a marketing leader.

# CLAUDE.md

## Who I am
[Name], [role] at [Company]. We [what the company does] for [who you serve].
Our stage: [e.g. Series B, 120 people, $18M ARR, expanding into enterprise].

## My marketing function
Team size: [X people]. Channels we own: [e.g. content, paid, lifecycle, PLG].
What we're responsible for: [pipeline, revenue, brand, product adoption — be specific].

## This quarter's priorities
- [Priority 1 — e.g. "Launch into mid-market segment, first campaign live by May"]
- [Priority 2 — e.g. "Reduce CAC by 20% through better top-of-funnel qualification"]
- [Priority 3 — e.g. "Build a content engine that generates 30% of inbound pipeline"]

## Where to find context
All foundation files live in /foundation/. Before any task, scan file headers
and load only files relevant to the work at hand.
- Audience and ICP context: /foundation/audience-delight.md
- Voice and tone: /foundation/creator-style.md
- Competitive landscape: /foundation/market-positioning.md
- How buyers actually move: /foundation/customer-journey.md

## How I work
- I think in outcomes, not outputs. Always tie recommendations to pipeline or revenue impact.
- I want strategic options, not one answer. Give me 2-3 approaches with tradeoffs.
- Flag assumptions. If you're working with incomplete data, say so.
- Be direct. No preamble, no summaries of what you're about to do — just do it.

## My key metrics
- [e.g. MQLs, SQLs, CAC, pipeline coverage, NRR — list the ones that matter to you]

## Current tools and stack
- CRM: [e.g. HubSpot]
- Analytics: [e.g. GA4, Amplitude]  
- Ads: [e.g. Google, LinkedIn]
- Content: [e.g. Webflow, Substack]

## Never
- Recommend tactics without connecting them to a metric
- Give me a single option when tradeoffs exist
- Use corporate filler: "leverage", "synergies", "best-in-class", "move the needle"
- Summarise what you're about to do — just do it

settings.json tells Claude what it cannot touch, your .env file, your secrets folder, any bash command that could cause damage. This is particularly important when you start to sync local file systems across team members so everyone has a single context layer for Claude Code (future post).

{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./secrets/**)",
      "Bash(curl *)",
      "Bash(rm -rf *)"
    ]
  }
}

And the folder structure. Organise by workflow, by how work actually moves through your projects:

/foundation/          ← context layer
/campaigns/           ← one folder per campaign
/content/
  /linkedin/
  /newsletter/
  /blog/
/research/            
/.claude/
  /skills/

---

Reach — What Claude can connect to

Without MCPs, Claude only knows what’s in the conversation, what’s in your terminal, what you’ve described, copied, or typed. MCPs give Claude access to live data across your stack, which changes the quality of everything it produces. The advice is grounded in what’s actually happening in your business right now.

For your power-user Claude Code setup, there are 2 MCP buckets to connect.

a. Better marketing decisions, faster by connecting your live marketing stack

Connect your CRM, analytics, and ad platforms, and Claude stops giving you strategic advice based on your description of the situation. It reads the situation itself.

Claude pulls live data, crosses sources in a single conversation, and gives you recommendations grounded in what’s actually happening. Three prompts that show what this looks like in practice:

“Pull all deals that haven’t moved in 21 days and summarise what’s stalling by stage and suggest plays to unblock them.” → HubSpot MCP.

“Which of our active paid campaigns has a CAC above target this month?” → Google Ads MCP.

“Go to [competitor]’s pricing page, capture what’s there, compare it against the snapshot saved in /research/competitor-snapshots/, and flag anything that’s changed.” → Playwright MCP controls a real browser, navigates to the live page, and compares it against a stored snapshot. Pricing tier changes, new feature callouts, and messaging shifts.

Useful MCPs: HubSpot MCP (free for all HubSpot customers), GA4, Google Search Console, Playwright MCP for competitor monitoring, Slack for routing outputs to your team.

b. Hear what your customers are actually saying

Most companies are sitting on a goldmine of unstructured customer data they never have time to properly read. Sales call transcripts. Support tickets. Win/loss interviews. NPS responses. G2 and Trustpilot reviews. Every one of those is a customer telling you exactly what they think, in their own words.

Connect Claude Code to those sources and you can turn unstructured customer data into data that can help power your marketing & growth strategy. Instead of reading a sample of calls and hoping it’s representative, Claude reads all of them. Instead of guessing on what objections are killing deals, you get a ranked list from the last 90 days of transcripts. Instead of positioning built in a stuffy offsite, you get messaging pulled directly from the language your best customers use when they describe why they bought your product or service.

The questions that become answerable: What objections come up most in deals we lose versus deals we win? What do customers say in their first 30 days that predicts whether they’ll expand or churn? What language do our happiest customers use that we’re not using in our own marketing?

Key connectors: Notion or Google Drive MCP for call transcripts and research docs, Gong or Fireflies if your call recording tool has an MCP, web scraping via Playwright MCP for review sites.

Install your default MCPs to ensure Claude Code is ready to be a true strategic partner.

Operation — What makes it a powerhouse

Skills are your team. Each one is a SKILL.md file, a structured instruction set that tells Claude exactly what to do, which foundation files to read, and where to save the output. Where a prompt is a one-off request, a skill is a repeatable hire.

Here’s one a CMO can build to run a weekly pipeline review:

# Skill: Weekly Pipeline Review

## When to use
Run every Monday morning for a clear picture of pipeline health 
and where to focus this week.

## Steps
1. Load foundation context
   Read /foundation/audience-delight.md for ICP criteria.
   Read CLAUDE.md for current quarter priorities and target metrics.

2. Pull pipeline data via HubSpot MCP
   - All open deals by stage
   - Deals with no activity in the last 14 days
   - Deals closing this month vs. target
   - Any stage regressions in the last 7 days

3. Produce the review
   - Pipeline coverage vs. target (are we on track?)
   - Deals at risk — stalled, regressed, or closing soon with low activity
   - Recommended actions by deal — what needs to happen this week
   - One-line summary for leadership: are we on track this quarter?

4. Save to /research/pipeline-review-[date].md
   Post summary to #revenue-team in Slack

How do you decide what skills to build?

Let Claude surface what to automate

The best way to know what to turn into a skill is to look at what you actually do. Two skills make this work together.

The first runs at the end of every session. Before you close Claude Code, run the End of Session Brief, it captures what you worked on, what prompts you ran, and what outputs were produced, then saves a timestamped file to /foundation/briefs/.

# Skill: End of Session Brief

## When to use
Run before closing Claude Code at the end of any working session.

## Steps
1. Review the current session
   - What tasks were completed?
   - What prompts were run more than once?
   - What outputs were produced and where were they saved?
   - What worked well? What needed multiple attempts?

2. Write a structured brief covering:
   - Session date and duration
   - Tasks completed (one line each)
   - Repeated prompts — exact descriptions of anything run more than once
   - Outputs produced and their file locations
   - Any friction points or tasks that felt manual

3. Save to /foundation/briefs/[YYYY-MM-DD-HH-MM].md

Over time, /foundation/briefs/ becomes a real record of how you actually use Claude Code. The second skill reads it.

# Skill: Find Skills

## When to use
Run weekly to surface patterns worth turning into skills.

## Steps
1. Read /foundation/briefs/reviewed.md to get the list of 
   already-reviewed briefs. If the file doesn't exist, create it.

2. Scan /foundation/briefs/ for any briefs not in reviewed.md

3. For each unreviewed brief, look for:
   - Any prompt or task marked as repeated
   - Any task that required multiple attempts
   - Any output that was manually reformatted after saving

4. Identify patterns across briefs — tasks that appear in 
   more than one session are the strongest candidates

5. For each candidate skill:
   - Name the skill
   - Write a first-draft SKILL.md with steps, inputs, and outputs
   - Estimate weekly time saved if automated

6. Present candidates with their draft SKILL.md files
   Append reviewed brief filenames to /foundation/briefs/reviewed.md

Run Find Skills every Friday. In a month you’ll have a system that surfaces its own gaps.

---

Toyota didn’t send those workers to Japan to motivate them. They sent them to give them everything they needed to do the job well before the job started, the context, the rules, the tools, the system.

When context, rules, reach, and operation are all in place, the ordinary becomes powerful. Everyone who sets up Claude Code rushes into prompts. There is real value in building the scaffolding: a /foundation/ folder with your .md files and a 20-line CLAUDE.md at the project root, default MCPs and your starter skills package.

Until Next Time, Happy AI’fying

Kieran