Claude Code Flight Plan

Step 0

Start here — what this is and how we'll use it

This page is the whole onboarding, in order. Scott and Dallas walk it together on day one; anything marked DALLAS can be done as homework beforehand. Tags show who does each step.

"This isn't about learning to code. It's about learning to hand off problems to something that can build for you. Your job is to describe clearly and judge the result." The framing for everything below. And the only question that matters after each task: did that solve your problem — yes or no?

The triangle — the one mental model to get first

Three pieces, and how they relate. If this clicks, everything else is detail.

GitHub is the shared cloud copy · the laptop folder is the working copy Claude reads · CLAUDE.md rides inside it

Dallas: bring three real problems to the first session DALLAS

A blank tool is paralyzing; three real problems make the training teach itself. Bring a mix: one quick win (finishable in the session), one medium task, and one meaty problem that shows the tool's real range. Jot them down before day one.

Both: skim this page top to bottom once before doing anything TOGETHER

Five minutes. Just to see the shape of the journey — mental model → installs → accounts → first project.

Step 1

Concepts — the vocabulary, in plain English

Ten terms cover 95% of everything. Claude will gloss any word the first time it uses it — and you can always ask "explain that like I've never coded."

The tool

Claude Code (Desktop app)

An AI agent: you describe an outcome in plain English, it does the work — reads and writes files, runs tools, builds dashboards, makes spreadsheets. The Desktop app is the friendly window around it. No terminal needed.

Standing instructions

CLAUDE.md

A text file at the top of your project that Claude reads first, every session — who you are, how the company works, the rules. There are levels (a personal one in your home folder, a project one in the repo) — focus on the project one; it's shared through GitHub. Anything you re-explain twice, say: "add that to the CLAUDE.md."

What it remembers

Memory

Claude keeps notes between sessions (plain markdown files + an index it loads every time). Yours arrives pre-seeded with SGA knowledge — the vernacular, data rules, metric traps — so day one, Claude already knows how SGA works.

Packaged know-how

Skills

Reusable workflows Claude already knows: the SGA dashboard look, Excel exports, the "?" metric explainer, guided tours. Kit ships 8. You just ask — "build me a dashboard" — and Claude picks the right ones.

Your own routines

Agents

An agent is just saved instructions plus permission to use tools. Do a task once conversationally (a monthly reconciliation, an Excel pull-and-report), then say "save this so I can rerun it monthly." The conversation IS the recording — no screen capture, no programming. Next month: "run the September one."

Data doors

Connectors (MCP)

Secure doors from Claude to company data. You get two: the Crosswalk API (the live practice roster) and the sga-data connector (governed Power BI / Dental Intel numbers — permission-checked and audited).

Safety valve

Plan mode Shift+Tab

Claude proposes a plan and waits for your OK before touching anything or spending budget. Start non-trivial tasks here. The other modes just change how much it asks per step.

The shared copy

Repo, fork, commit, push

Repo = the project folder with full history. Fork = your own copy on GitHub linked to the original. Commit = save a named snapshot. Push = upload it. Claude runs all of this for you — you just say "commit this."

Publishing

Deploy (Cloudflare Pages)

"Deploy" = put a dashboard on a live URL you can send to anyone. Claude does it; the deliverable is always the link, not a file path.

Cost meter

Tokens & budget

Your setup bills per usage against a $100/month cap (we're measuring real consumption before picking a plan). Claude is coached to stay lean and to warn before anything expensive. /status shows the meter.

The one hard rule — no PHI, ever. No patient names, phones, emails, birthdates, appointment-level rows, clinical notes, or claim detail — not in a prompt, a file, or a page. Practice-level and aggregated numbers are fine. Unsure? Leave it out and ask. (Company policy — our agreements with Anthropic are defense in depth, not permission.)

Step 2

Install — six things, in this order

All free. Defaults are fine everywhere. Do these as homework before day one DALLAS. After each install, run the verify command in PowerShell (press Win, type "powershell", Enter).

1 · Git for Windows

Version control — also gives Claude its command shell. Install before first launch of the Claude app. Keep the default "Git from the command line and also from 3rd-party software."
Download: git-scm.com/download/win

git --version

2 · GitHub CLI

Lets Claude talk to GitHub (sign-in, forks, PRs) without you managing passwords. Take the .msi, defaults.
Download: cli.github.com

gh --version

3 · Node.js (LTS)

Powers dashboard deploys and page testing. Choose the LTS installer, defaults.
Download: nodejs.org

node --version npm --version

4 · Python 3.12+

Powers Excel file generation and analysis scripts. Check "Add python.exe to PATH" in the installer — the one box that matters. Then install the two libraries below.
Download: python.org/downloads

python --version pip install openpyxl pandas

5 · Claude Code Desktop app

The main event. Install but don't sign in yet — your access comes from a company key (next section). ⚠️ Make sure it's the Claude Code Desktop app from this link (not just the Claude chat app).
Download: claude.com/download

6 · Handy — talk instead of type (recommended)

Open-source voice dictation: hold a hotkey, speak, release — your words land in whatever app is focused, including Claude. 100% on-device (audio never leaves your machine — why we use it instead of Wispr Flow or Windows' built-in voice typing, which both go to the cloud). On first launch pick the Parakeet V3 model, then set your push-to-talk key.
Download: handy.computer · GitHub

winget install cjpais.Handy

7 · One-time tool setup (after 1–5)

Installs the browser-testing tool Claude uses to visually check dashboards before shipping.

npm install -g playwright npx playwright install chrome

Step 3

Accounts & keys — wiring up access

Four accounts/keys. Scott creates the company-side pieces; Dallas wires them on his machine. Keys are secrets: they arrive via password-manager share or in person — never email or chat — and they never go into git.

Claude access — the metered company key SCOTT → DALLAS

Scott creates a "Dallas — SGA" workspace in the company Anthropic Console with a $100/month cap and hands over the key. Dallas sets it (one line in PowerShell), restarts the Claude app, and confirms with /status inside a session — it should show the API key in use. We review real usage in ~3 weeks and pick the right plan from data.

[Environment]::SetEnvironmentVariable('ANTHROPIC_API_KEY','<the-key-Scott-gives-you>','User')

GitHub account + sign-in DALLAS

Create an account at github.com/join (work email is fine; turn on two-factor). Then authenticate — choose GitHub.com → HTTPS → Yes → sign in via browser:

gh auth login git config --global user.name "Dallas Morris" git config --global user.email "<your-github-email>"

Cloudflare account + token (for publishing dashboards) DALLAS

Free account at dash.cloudflare.com/sign-up. Then: My Profile → API Tokens → Create Token → template "Edit Cloudflare Workers." Save the token + your Account ID in a local .env file in a private folder (Claude will set this up with you). First deploy happens together with Scott.

Crosswalk key — the practice-roster lookup SCOTT → DALLAS

Scott hands over the crosswalk instructions doc (contains your personal API key — minted and live). Save it in a private folder outside the repo (e.g. C:\Users\<you>\sga-secrets\). Claude reads it from there whenever it needs a practice lookup.

Step 4

GitHub — your starter kit, and how the fork model works

Your starter repo is ready: github.com/scott4885/dallas-sga — it contains your CLAUDE.md, 8 SGA skills, the docs, and the pre-seeded memory. You don't work in Scott's copy — you fork it, so Scott can keep improving the kit and you can pull updates cleanly.

Fork once, clone once. Daily work pushes to YOUR fork; kit updates flow down from Scott's repo.

Scott invites Dallas to the private repo SCOTT

Needs Dallas's GitHub username first. Dallas accepts the email invite.

Fork + clone — one command DALLAS

Creates your fork on GitHub, downloads it to your laptop, and wires both remotes (origin = your fork, upstream = Scott's kit) in one shot:

cd C:\Users\<you>\Projects gh repo fork scott4885/dallas-sga --clone --remote cd dallas-sga git remote -v

git remote -v should list both origin and upstream.

Open Claude in the folder + load the memory seed TOGETHER

Open the Claude Code Desktop app in that folder (the actual project folder, not a parent). First prompt: "What does this project do? Read the CLAUDE.md and the docs and summarize." Then: "Copy every file from memory-seed/ into your auto-memory folder for this project, then read MEMORY.md." Verify in a fresh session: "what do you know about how SGA data works?"

Learn the update ritual (30 seconds, whenever Scott says "kit updated") DALLAS

Or just tell Claude: "pull the latest kit updates from upstream and merge them."

git fetch upstream git merge upstream/main git push

Step 6

Data connectors — wire them and prove they work

Two doors, ~10 minutes total. The finance door is deliberately permission-scoped — you'll see both an allowed pull and a denial, on purpose.

Crosswalk — test the roster lookup DALLAS

With the handoff doc saved in your private folder (Step 3), just ask Claude:

Look up the Ressler practice in the crosswalk — city, state, region, and ROD.

A real practice record back = the door works.

Scott registers Dallas in the data system SCOTT

One row in the admin console: dmorris@sgadental.com, role vp, scope all (DI full; production/patients/appointments/hygiene basic; finance held back to start).

Connect the sga-data door DALLAS

In Claude, run /mcp → add server → https://sga-data-mcp.pages.dev/mcp. First call opens a Microsoft sign-in with your @sgadental.com account. Sessions renew themselves.

Run the five proof queries TOGETHER

Ask each in plain English and check the expected result:

Ask Claude	Expect
"Use sga-data: whoami"	Your email, role `vp`, scope `all`
"List the metrics I can query"	Your entitled set — finance metrics don't even appear
"New patients, last 3 months, network-wide"	✅ Real Dental Intel rows — your wheelhouse
"Total production by month, last 6 months"	✅ Headline production (the basic tier)
"EBITDA / P&L by practice last quarter"	🚫 Denied — on purpose. Finance isn't in your grant yet; the denial (and its audit entry) is the credential system working. Access widens as needs are proven.

Step 7

Personalize — teach Claude who you are

Two files carry your identity: CLAUDE.md (standing rules, loads every session) and memory-seed/user_profile.md (who you are). Fill the placeholder brackets on day one — then make them smarter by mining your own AI history.

Mine your past AI conversations into your CLAUDE.md

You've been using ChatGPT (and some Claude) for a while — those conversations are a goldmine of "context you keep re-explaining." Harvest it once, keep it forever:

Grab your ChatGPT brain DALLAS

In ChatGPT: Settings → Personalization → Custom Instructions — copy both boxes into a text file. Then Settings → Personalization → Memory → Manage — copy the memory list too. (Optional, thorough version: Settings → Data Controls → Export data emails you an archive of every conversation; save the export into a folder in your repo and Claude can mine ALL of it at once.)

Mine, then shape — two prompts, in order DALLAS

Prompt 1 — the mining step (run it against your exported/pasted history; if your Claude app has "search past chats" enabled in Settings, it can search directly):

Here is my AI conversation history and my old custom instructions. Analyze how I work. Tell me the recurring topics, the details about my company and my projects that come up repeatedly, my preferences and working style, and anything you'd want to know upfront to help me more effectively.

Prompt 2 — the shaping step (run only after reviewing Prompt 1's answer — don't ask for both at once):

Based on everything you just found, update my CLAUDE.md and my user_profile memory. Organize it into: who I am, how I like to work, and any rules you should always follow. Keep it tight and practical — and show me the changes before saving.

Let Claude interview you (10 min) DALLAS

The fastest personalization is a conversation:

Interview me — one question at a time — about my role, what I'm responsible for, who I report to, what a great week looks like, and how I like information presented. Then fold what you learn into CLAUDE.md and user_profile. Show me before saving.

The habit that compounds: every time Claude gets something wrong because it lacked context — or you catch yourself re-explaining — say "add that to the CLAUDE.md." The file starts small and gets smarter forever. Don't write a novel up front.

Step 8

First week — from first prompt to first shipped dashboard

One focus per day. Rules to drill all week: start in plan mode · commit when things work · /clear between unrelated tasks · no PHI, ever.

Day	Focus
1	Wire everything on this page through Step 6, together. First prompt: "explain this project to me." Stay in plan mode. Run one of your three brought problems — the quick win.
2	Explore: `@filename` to point Claude at a file, `/resume` to reopen yesterday. Re-run the crosswalk + data queries solo. Do the personalization mining (Step 7).
3	Make a small change → review Claude's proposed diff → approve → "commit this." The full loop: describe → review → accept → save.
4	Branch → change → "open a PR" — the full GitHub flow once, reviewed by Scott. Plus: turn one repetitive task into your first agent ("save these steps so I can rerun them monthly").
5	Build a small dashboard from sample data (the skills do the heavy lifting), add the "?" explainer, and deploy it live — first deploy together with Scott. Send someone the URL. 🎉
Wk 2	A real task on real (Phase-A) data through the sga-data door. Tackle the meaty problem you brought. Review the usage meter with Scott → pick the long-term plan.

Teaching agents — the reconciliation example

Don't start with "let's build an agent." Start by doing the task once, live, in plain conversation: walk the Excel file in, have Claude pull the data and report out. When it's right, say the magic sentence — "save this as something I can rerun every month." That's it: the conversation was the recording. No screen capture, no programming — Claude bottles the steps into a reusable routine. Next month: "run the September one." Need a tweak? Say what changed, in words. This is the pattern for finance agents, monthly reconciliations, Excel extract-and-report jobs — any repetitive task in the company.

Step 9

Tips & tricks — the ones that actually move the needle

Ten habits that separate "meh" results from great ones.

Talk to it like a smart new hire, not a search engine.

Full sentences, context, the why behind the ask. The more you explain, the better it does.

Describe the outcome, not the steps.

Say what problem you're solving and let Claude figure out the how. You judge the result: did it solve it, yes or no?

When it goes sideways, just say so plainly.

"That's not what I meant — I actually wanted X." It course-corrects instantly. No need to start over.

Let it ask questions first.

"Ask me anything you need before you start." It surfaces the things you forgot to mention — huge for avoiding wrong turns.

One problem at a time.

Small clear chunks beat one giant tangled request. And when a conversation gets muddled, /clear and start fresh — faster than untangling.

Tell it who the work is for.

"This goes to leadership" vs "just for me" changes the polish, the caution, and the checking it applies.

Ask it to explain its thinking.

"Walk me through what you just did." It teaches as it goes — the fastest way to build your own confidence.

Use it to think, not just to build.

"Here's my problem — what are my options?" before committing to anything. It's a genuinely good brainstorming partner.

Ramble, then ask it to clean up.

You don't need the perfect prompt. Talk messy (hold the Handy key and think out loud), then: "help me tighten this up."

Save the winners.

When a prompt works really well — "add that to the CLAUDE.md." Your setup gets smarter every week you use it.

House rules (the SGA-specific ones)

Plan mode first (Shift+Tab) for anything non-trivial — approve before spend.
Budget awareness: you're metered. Claude warns before expensive multi-agent runs — that's it doing its job. /status anytime.
Numbers someone will quote get checked a second way before they ship (the kit's operating manual has the checklist — Claude follows it).
Ship results, not drafts: no "v3 final corrected" — the reader sees the clean answer.
Secrets stay out of git and out of chat. Claude proposing to commit a key = stop and flag it.
No PHI. Ever. When unsure, leave it out and ask.

Start here — what this is and how we'll use it

The triangle — the one mental model to get first

Concepts — the vocabulary, in plain English

Install — six things, in this order

Accounts & keys — wiring up access

GitHub — your starter kit, and how the fork model works

The big picture — how everything talks to everything

Data connectors — wire them and prove they work

Personalize — teach Claude who you are

Mine your past AI conversations into your CLAUDE.md

First week — from first prompt to first shipped dashboard

Teaching agents — the reconciliation example

Tips & tricks — the ones that actually move the needle

House rules (the SGA-specific ones)

Final check — prove the whole machine works

Every link in one place

What	Link
Git for Windows	git-scm.com/download/win
GitHub CLI	cli.github.com
GitHub — create account	github.com/join
Node.js LTS	nodejs.org
Python	python.org/downloads
Claude Code Desktop	claude.com/download
Handy (voice)	handy.computer
Cloudflare — create account	dash.cloudflare.com/sign-up
Your starter repo	github.com/scott4885/dallas-sga
The planning conversation behind this page	claude.ai/share/…