A Couch & Associates Guide · v2.0Agentic AI Workflow

● The Build Manual

From AI chat to AI that builds.

Built for people who don't write code, and who don't need to. A step-by-step guide to working with agentic AI on your own machine. Set it up once, then let Claude read your files, run your tools, and build real things alongside you. Start simple — open any topic's ◇ Go deeper when you're ready for more.

A Couch & Associates field guide

Why this · Why now00

The shift

Stop chatting. Start building.

A browser chat window can only talk. An agent running on your computer can act — open files, run programs, write code, and create real things using your own hardware and software. That single shift is what this whole guide is about.

Browser chat

Talks · Forgets · Suggests

Agent on your machine

Acts · Remembers · Builds

The Build Manual

·MAP

Overview

What you'll do

Sixteen steps, one afternoon.

Value

By the end you'll have Claude installed, a tidy project folder, and a repeatable daily rhythm. Read the plain path first; reach for the advanced layers only when you want them.

Set up

01 Get a subscription
02 Install Claude Code
03 Verify it works
04 Create a project folder
05 Open a terminal & go in
06 Launch Claude

Ways of working

07 Keep the folder in view
08 Add Python
09 Work in Markdown
10 Tell Claude how to build
11 Spec before you build
12 Keep the agent sharp

Level up

13 Add Git version control
14 Write memories
15 Security & guardrails
16 Connect your tools

◇ Go deeper — advanced tips, per topic macOS / Windows — toggle, top-right

Build Manual · Overview

01/ 16

Set up

Step 01 · The ticket in

Get a subscription.

Value

A paid plan is the entry ticket — the free tier can't run the agent. Pro is plenty to start, and for most teams it's an easy business expense.

Go to claude.com and sign in or create an account.
Choose Claude Pro — about $20/month (or $17/month billed annually).
Keep the receipt — if your organization reimburses tools like this, it's a simple expense claim.

Heavy daily users can move up to a Max plan later. Start with Pro — it's enough to learn everything here.

Build Manual · Subscription

02/ 16

Set up

Step 02 · Install

Install Claude Code.

Value

The native installer puts a single claude command on your computer and keeps it updated automatically — no Node.js, no manual PATH setup.

Open your terminal, paste one line, press Enter.

PowerShell — Start ▸ type "PowerShell"

PS> irm https://claude.ai/install.ps1 | iex

Terminal — ⌘+Space ▸ type "Terminal"

$ curl -fsSL https://claude.ai/install.sh | bash

After installing, close and reopen your terminal so it sees the new command.

Build Manual · Install

◇DEEP DIVE

Install

◇ Advanced · Install

If the one-liner fights you

Install options & fixes.

Value

Most people only need the one command. These are the fallbacks and checks the pros reach for when something's off.

Already use Node.js? npm install -g @anthropic-ai/claude-code works on every OS (you update it manually).
Prefer a package manager? Homebrew on Mac, WinGet on Windows.
Run claude doctor — it reports your version, install method, and any fixes.
Windows: install Git for Windows so Claude gets a full shell instead of falling back to PowerShell.

Never use sudo with the npm install — it's the number-one cause of permission errors.

Build Manual · Install · advanced

03/ 16

Set up

Step 03 · Confidence check

Verify it works.

Value

Thirty seconds now saves an hour of "why won't this run" later. These two commands prove the install and self-diagnose problems.

your terminal

> claude --version
2.x.x  (Claude Code)

> claude doctor
# checks your setup and suggests fixes

If claude "isn't found", reopen the terminal first. If it still fails, claude doctor will point to the fix — usually a PATH refresh.

Build Manual · Verify

04/ 16

Set up

Step 04 · A home for the work

Create a project folder.

Value

One tidy home per project keeps Claude focused and your work backed up. Sync it to the cloud and backups are automatic.

Make a folder like C:\Projects\my-project. Tip: keep your Projects folder inside OneDrive so everything backs up automatically.

Make a folder like ~/Projects/my-project. Tip: keep Projects in iCloud Drive or OneDrive for automatic backup.

One folder = one project. Don't mix unrelated work — it keeps the agent and your backups clean.

Build Manual · Folder

05/ 16

Set up

Step 05 · Point it at the work

Open a terminal & go in.

Value

"cd" simply means "go into this folder." It's how you tell Claude exactly where to work.

PowerShell

# or: in the folder, Shift+Right-click ▸ "Open in Terminal"
PS> cd C:\Projects\my-project
PS> cd ..   # go up one level

Terminal

# or: Finder ▸ right-click folder ▸ New Terminal at Folder
$ cd ~/Projects/my-project
$ cd ..   # go up one level

Your prompt should now show the project folder's path. That's how you know you're "inside" it.

Build Manual · Navigate

06/ 16

Set up

Step 06 · Go

Launch Claude.

Value

This is the moment chat becomes an agent that can actually touch your files.

Inside your project folder, type claude and press Enter.
First time only: a browser opens — sign in to approve.
You're in. Type a request in plain English and watch it work.

Everything Claude makes lands in this folder. The folder is both the workspace and the boundary — see Step 15.

Build Manual · Launch

07/ 16

Working

Step 07 · Stay oriented

Keep the folder in view.

Value

Watching files appear in real time turns "magic" into something you can see, check, and trust.

Open the folder in File Explorer next to your terminal. Drag in any artifacts — documents, recordings, screenshots — you want Claude to work with.

Open the folder in Finder next to your terminal. Drop in any files you want Claude to work with.

Add the reference file to the folder first, then tell Claude it's there and what to do with it.

Build Manual · In view

◇DEEP DIVE

Context

◇ Advanced · Giving context

Point precisely

Feed it more than text.

Value

The more precisely you point Claude at something, the better the result — and you can point at pictures, not just words.

Paste or drop in a screenshot — a design, an error message, a chart — and ask Claude to work from it.
Use @filename to point at a specific file instead of letting Claude hunt for it.
Drop reference docs into the folder and say "read X before you start."

A picture often briefs Claude faster and more accurately than a paragraph of description.

Build Manual · Context · advanced

08/ 16

Working

Step 08 · Unlock real work

Add Python.

Value

Python is a programming language Claude can use to crunch data, process files, and automate tasks. Adding it noticeably improves your results.

Easiest: just ask — "Check if Python is installed, and install it if it isn't." Or do it yourself:

PowerShell

PS> winget install Python.Python.3.12
PS> python --version   # reopen terminal first

Terminal

$ brew install python
$ python3 --version

Success looks like a version number printing back. If it doesn't, ask Claude — it can usually fix it.

Build Manual · Python

◇DEEP DIVE

Verify

◇ Advanced · Self-checking

Less babysitting

Let Claude check its own work.

Value

An agent that can test its own output catches its own mistakes — so you supervise less and trust the result more.

Ask Claude to write and run tests, or to build and open the result and confirm it works.
For visuals, have it open the page and screenshot it to compare against what you asked for.
Tell it plainly: "Prove it works before you call it done."

This is the difference between a session you have to babysit and one you can walk away from.

Build Manual · Verify · advanced

09/ 16

Working

Step 09 · Shared memory

Work in Markdown.

Value

Markdown (.md) files are plain text that both you and the agent read easily. They become your project's shared plan and memory.

Ask Claude to capture plans, notes, and specs as .md files in the folder.
To read them nicely, install Obsidian (obsidian.md) and point it at your folder.
Use Obsidian just as a viewer — ignore its other features unless you want them.

Think of MD files as the living documents you and Claude both edit and trust throughout the project.

Build Manual · Markdown

10/ 16

Working

Step 10 · Standing instructions

Tell Claude how to build & store.

Value

A few standing instructions make Claude's output faster to use, easier to find, and simpler to keep.

"Cache all web research locally in this folder."

So findings are saved once, not fetched again and again.

"Use SQLite for any data storage."

A simple, file-based database that lives right in the folder.

"Build a single HTML file for this."

For quick visuals you can just double-click to open.

"Make a Node web app, plus a start file — start.bat on Windows, start.command on Mac."

For richer apps — one click launches it.

These are habits, not one-offs. Add them to your project's MD instructions so every session starts the same way.

Build Manual · Build & store

11/ 16

Working

Step 11 · The best-practice move

Spec before you build.

Value

The biggest quality gains happen before any code — in thinking the problem through together.

Draft the goal and requirements as MD files first.
Tell Claude: "Be critical. Challenge my assumptions. Ask clarifying questions often."
Only start building once the spec feels right.

Slow is smooth, and smooth is fast. Time on the spec is time you don't spend redoing the build.

Build Manual · Spec first

◇DEEP DIVE

Plan

◇ Advanced · Plan Mode

Think before it touches anything

Use Plan Mode.

Value

In Plan Mode Claude reads and reasons but can't change files — so you settle the approach before any code is written. It's the workflow Claude Code's own creator uses.

Press Shift+Tab to cycle modes; stop when you see "plan mode on."
Go back and forth until the plan is right, then approve and let it execute.
Push back hard: "Grill me on this plan and prove it works before building."

Most strong sessions start in Plan Mode. Plan first, build second.

Build Manual · Plan Mode · advanced

12/ 16

Working

Step 12 · The habit that matters most

Keep the agent sharp.

Value

Claude's memory within a session is finite. Keeping it clean is the single habit that most separates smooth sessions from frustrating ones.

Run /clear when you switch to an unrelated task — don't drag old context into new work.
Watch for drift: if Claude repeats itself, re-asks for a file it already saw, or gets vague, its memory is overloaded.
If it heads the wrong way, press Ctrl+C to stop and redirect — don't wait for a big wrong change.

A fresh start with a clear prompt almost always beats a long, cluttered session.

Build Manual · Stay sharp

◇DEEP DIVE

Steering

◇ Advanced · Steering & context

Rescue and refocus

Steering & context control.

Value

A few deliberate moves keep Claude focused on long jobs and rescue sessions that start to wander.

/compact Focus on … — summarizes the session but keeps the decisions you name.
Don't argue, rewind. Use /rewind (or Esc Esc) to undo a wrong turn — re-explaining keeps the mistake in context and pollutes the next try.
Decompose big asks into small, verified steps instead of one giant request.
"Use a subagent to research X" — it explores in a separate memory and reports back a summary, keeping your main session clean.

Compact at milestones on purpose; don't wait for it to happen automatically and lose the parts you cared about.

Build Manual · Steering · advanced

13/ 16

Level up

Step 13 · Real version control

Add Git.

Value

Git gives you a true undo across your whole project and a full history of changes — something file-sync can't do. It also gives Claude a record of what changed and why.

Ask Claude to install Git and set it up in your project.
Let Claude commit changes as you go — each commit is a save point.
Roll back anytime to a known-good version.

OneDrive backs up your files; Git versions your project. They do different jobs — use both.

Build Manual · Git

14/ 16

Level up

Step 14 · Clean handoffs

Write memories.

Value

Telling Claude you're done lets it save context for next time — so the next session starts informed, not from scratch.

When you finish, say: "We're done — consolidate what we did and write memories to the project folder for next time."
Claude updates its notes and MD files.
Next session, it picks up where you left off.

A clean handoff today is a fast start tomorrow — for you and the next session.

Build Manual · Memory

◇DEEP DIVE

CLAUDE.md

◇ Advanced · Project memory

Stop re-explaining your project

CLAUDE.md: the standing brief.

Value

Instead of re-explaining your project every session, write it down once in a CLAUDE.md file. Claude reads it at the start of every session.

Run /init and Claude drafts a starter CLAUDE.md from your folder.
Keep it under ~200 lines: build/run commands, conventions, "always do X" rules, what not to touch.
It also learns on its own (auto memory), and you can turn a prompt you repeat into a one-word /command.

If a conflict arises, CLAUDE.md wins over a one-off request. Keep it lean — every line costs attention.

Build Manual · CLAUDE.md · advanced

15/ 16

Level up

Step 15 · Stay safe

Security & guardrails.

Value

Claude stays inside your project folder unless you send it elsewhere — so the boundary is simple and in your control.

Keep one folder per project.
Don't put anything sensitive — passwords, private data, secrets — inside a project folder.
Review changes before approving anything that reaches outside the folder.

Rule of thumb: if you wouldn't want the agent to see it, don't keep it in the folder.

Build Manual · Security

◇DEEP DIVE

Control

◇ Advanced · Permissions

See it before it happens

Permissions & reviewing changes.

Value

You stay in control by seeing what Claude wants to do before it does it.

Claude asks before actions that change files or reach outside the folder — read the request before approving.
Use /diff to see exactly what changed before you keep it; pair it with /rewind to undo.
Manage what's allowed with /permissions.

When in doubt, approve narrowly. You can always grant more later.

Build Manual · Permissions · advanced

16/ 16

Level up

Step 16 · Beyond the folder

Connect your tools.

Value

Connectors (MCP) let Claude reach into your email, calendar, drive, or a database — so it can use real data, with your permission.

Open the connectors / tools menu and turn on what you need (e.g. Google Drive, Calendar).
Ask in plain English: "Find last quarter's report in my Drive and summarize it."
Approve the access prompt the first time.

Turn off connectors you're not using — each one Claude can see adds clutter and cost.

Build Manual · Connect

◇DEEP DIVE

Connect

◇ Advanced · Connectors

Wire it into your workflow

Setting up & managing connectors.

Value

A little setup turns Claude from a folder assistant into something wired into your actual day-to-day tools.

Browse available connectors and connect the ones you use.
Disable unused ones to keep Claude focused and fast.
Combine tools: "Check my calendar, then draft an email to everyone in tomorrow's 10am."

Only connect tools you trust with the data they can reach. Approvals are per-connector.

Build Manual · Connect · advanced

!FIXES

Troubleshoot

Week-one survival

Common mistakes & quick fixes.

Value

Almost every early stumble has a one-line fix. Keep this page nearby for your first week.

claude "isn't recognized"

Reopen your terminal. Still stuck? Run claude doctor.

Quality dropped mid-session

Run /clear and restate the task in a fresh, clear prompt.

It went down the wrong path

Use /rewind (or Esc Esc). Don't argue it back on track.

It keeps "forgetting" your project

Put the standing facts in a CLAUDE.md file.

It's editing the wrong thing

You're in the wrong folder — cd into the right one.

It asks permission constantly

That's by design. Approve, or adjust with /permissions.

When unsure: stop (Ctrl+C), clear, and start the step again with a sharper prompt.

Build Manual · Fixes

A–ZTERMS

Glossary

Shared vocabulary

Glossary.

Value

A shared vocabulary so nothing gets lost in the jargon.

TerminalThe text window where you type commands.

PATHThe list of places your computer looks for a command like claude.

Folder / projectOne home per piece of work; Claude works inside it.

Markdown (.md)Plain text with light formatting — your shared notes.

Git / repoA versioned history of your project you can roll back.

Context windowClaude's working memory for the current session.

CLAUDE.mdA file of standing instructions Claude reads every session.

MCP connectorA bridge that lets Claude use an outside tool (Drive, email…).

TokenThe unit your usage is measured in.

SubagentA helper that researches in its own separate memory.

Build Manual · Glossary

The bigger picture—

Beyond Claude

Go for gold.

Claude Code is one excellent option — OpenAI's Codex CLI and Google's Gemini CLI are strong too, and tools like Claude's Cowork sit between chat and full agent. The real shift isn't the brand. It's leaving browser chat behind for an agent that can use your computer to truly build.

1. Pick one.2. Learn it well.3. Then explore the rest.

Portability tip: an AGENTS.md file does the same job as CLAUDE.md but is read by Codex, Cursor, and others too — handy if your team mixes tools.

The Build Manual

★REF

Cheat sheet

Pin this page

The daily loop & commands.

Value

This single page is most of what you need day to day. Print it and keep it nearby.

The daily loop

Open a terminal
cd into your project folder
Run claude
Draft the spec in .md first
Build · review with /diff
Say "write memories" when done

Commands you'll use

Startclaude
Fresh start/clear
Free up memory/compact
Undo a wrong turn/rewind
See changes/diff
Check usage/usage
Diagnoseclaude doctor

Build Manual · Cheat sheet

◇DEEP DIVE

Reference

◇ Advanced · Full reference

Everything in one place

Command reference.

Value

The fuller set — type / in Claude any time to see the live list with descriptions.

/clear — wipe context, start fresh

/compact — summarize & free up memory

/context — see how full memory is

/rewind — roll back code or chat

/diff — review changes

/init — draft a CLAUDE.md

/memory — edit standing instructions

/permissions — control what's allowed

/usage — plan limits & usage

/cost — token cost (API billing)

/doctor — health-check the install

/help — list everything

@file — point at a specific file

!command — run a shell command

Shift+Tab — switch modes (incl. Plan)

Ctrl+C — stop and redirect

Esc Esc — open the rewind menu

Build Manual · Reference · advanced

✓CHECK

Checklist

Print this page

Your setup checklist.

Value

Tick these off as you go. Once they're all checked, you're up and running.

Set up

☐ Subscribed to Claude Pro ☐ Installed Claude Code ☐ Verified with claude --version ☐ Created a project folder (cloud-synced) ☐ Opened a terminal & cd'd in ☐ Launched claude & signed in

Working & level up

☐ Folder open in Explorer / Finder ☐ Python installed ☐ Tried specs in Markdown ☐ Practiced /clear between tasks ☐ Set up Git ☐ Wrote a CLAUDE.md ☐ Reviewed the security rules

Build Manual · Checklist

Your turnEnd

Go build

Your turn to build.

Set it up once this week. Start with something small and real. Ask Claude to challenge you — and let it do the building. When you want more on any topic, that's what ◇ Go deeper is for.

↗ claude.com ↗ docs.claude.com/claude-code ↗ obsidian.md

The Build Manual · v2.0