Context is the single most important concept this week
1
Context is everything Claude can see right now. The model has no other memory between turns. If it isn't in context, it doesn't exist for Claude.
2
Context determines quality. Well-curated context produces careful answers. Cluttered context produces confused ones. Same model, different output.
3
Context is finite and it degrades. 200,000 tokens (about 150K words) fills faster than you'd think. As it fills, performance drops.
The mechanics of context, how it rots, and how to manage it come back in § III.
Day 1 · Why context10
Install
If you haven't installed yet, one command does it
macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash
Windows PowerShell
irm https://claude.ai/install.ps1 | iex
Verify it worked
claude --version
Self-contained installer; no Node.js needed. (Windows: install Git for Windows first.) R or Python: ask Claude Code to install them during § III.
Can't install locally? Use the Codespace fallback at socialscienceai.com/pre-workshop.
Day 1 · Install11
Pricing
Which Claude tier you need
Pro
$20/mo
Very limited Claude Code access. You may not finish all exercises this week.
Max (5x)
$100/mo
The realistic floor for this workshop. Best place to start if you can afford it.
Max (20x)
$200/mo
Best deal for heavy daily users. This is what I use.
API
Pertoken
No caps; requires an API key. Most expensive way to use Claude Code daily.
tiers as of May 2026
Day 1 · Pricing12
§ II
Workshop · Set up & first conversation
Setup
Open Claude Code, then have it download the materials
1
Make a project folder named workshop.
Macin FinderWindowsin File Explorer
2
Open the terminal.
Mac⌘ Space → type Terminal → EnterWindowsStart menu → type PowerShell → Enter
3
Type cd + space, drag your folder onto the terminal, press Enter.
4
Type claude, press Enter.
5
Paste this prompt, press Enter:
Download everything at socialscienceai.com/materials/session_1/ into this folder. When you're done, describe the data to me.
6
Say yes to any permission prompts Claude asks you.
Day 1 · SetupII.1
§ III
Context management
Context window
Everything Claude has seen this session is one long document
Limit: 200,000 tokens (about 150,000 words). A token is roughly three-quarters of a word. Fills faster than you'd think.
Day 1 · Context windowIII.1
Three sources
Context comes from three places. Claude can reach two of them.
Dictate it
When you know what to say but typing is slow. Wispr Flow, SuperWhisper, Aqua Voice.
Have Claude interview you
When you might not know what to say. Next slide.
Day 1 · Three sourcesIII.2
Workspace setup
Drop everything Claude might need into your project folder
# your project folder
your-project/
├── CLAUDE.md
├── notes/
│ ├── research_questions.md
│ ├── data_dictionary.md
│ └── meeting_notes.md
├── papers/
│ ├── johnson-2023.pdf
│ └── li-and-smith-2024.pdf
├── code/
│ ├── 01_clean.R
│ └── 02_analyze.R
├── data/
│ └── ipeds_panel.csv
└── drafts/
└── outline.md
CLAUDE.md is special
Read automatically at session start. Project conventions, your preferences, anything Claude should always know.
Other .md files
Research questions, data dictionaries, meeting notes, drafts. Claude reads them when relevant.
Source materials
PDFs of papers you're citing, prior code, scratch analyses, codebooks. Drop them in. Claude reads them on demand.
Principle
Richer directory, richer context. If Claude might need it, drop it in.
Day 1 · Workspace setupIII.3
Interview-me
Have Claude interview you before you ask for something hard
01
Vague task
plan my talk
what you typed
02
Interview
Q1. How long is the talk?
A. 20 minutes.
Q2. Who is the audience?
A. Business-school faculty.
Q3. What do you want them to remember?
A. My main result is robust to alternative IVs.
questions you may not have thought to specify
03
Sharp task
draft a 20-min talk for business-school faculty; landing point = main result is robust to alternative IVs
what plan mode now has to work with
Prompt template: interview me about [task]. ask one question at a time.
Day 1 · Interview-meIII.4
Context rot
Long sessions degrade. Fresh sessions don't.
The phenomenon
As context fills, Claude has more to keep track of and starts losing the thread. The signal-to-noise ratio drops.
Around turn 20–30 of a heavy session, performance is noticeably worse than turn 3.
What it looks like
· Claude misremembers earlier decisions
· Forgets constraints you set early in the session
· Goes in circles on a problem it would solve fresh
· Uses old patterns after you've corrected it
"The model has more to keep track of, and it starts losing the thread. This is the single most important thing to understand about working with AI coding tools."
— Paul Goldsmith-Pinkham
Day 1 · Context rotIII.5
Auto-compaction
When context fills, Claude summarizes older turns into a memory block
Auto-compact works fine for most research workflows when state is in files. Manual /compact gives you control over when and what to keep.
Day 1 · Auto-compactionIII.6
Managing context
Three habits that keep context useful
01
Write state to files
Have Claude write progress, plans, and decisions to .md files in your project. Files persist across sessions; conversation context doesn't.
02
Keep sessions short
5–10 focused turns per session, then start fresh. Marathon sessions are where the rot happens.
03
Compact intentionally
Type /compact before context fills, with a hint about what to keep. Better than waiting for auto-compact to lose detail.
three rules via Paul Goldsmith-Pinkham
Day 1 · Managing contextIII.7
Plan mode
Plan mode is where you and Claude agree on the approach before any code runs.
Before Claude touches a file, it states what it intends to do. You read the steps, push back on the wrong ones, refine. By the time anything executes, the two of you have agreed on the approach.
Without plan mode
Claude assumes
Claude reads your prompt and acts. It picks variable names, file paths, methods, and sequencing on its own.
By the time you see what it did, you are unwinding decisions that were never yours.
With plan mode
you negotiate
Claude proposes the steps it would take, in order, before any of them run.
You approve, reject, or amend. The assumptions become visible and editable before they execute.
Day 1 · Plan modeIII.8
Workflow rhythm
Every agentic task moves through three phases
Naming the rhythm matters. Once you've practiced Plan, Execute, Clear once, you can recognize which phase you're in mid-task. That recognition is the difference between using Claude Code well and getting lost in it.
01
Plan
Claude proposes what it would do, before it does anything. You read the plan, approve, reject, or amend.
Shift+Tab Tab to enter
02
Execute
Claude runs the plan: edits files, runs commands, asks permission as needed. You watch, and intervene if it drifts.
ESC to interrupt
03
Clear
When the task is done or context fills, compact or clear. Files persist; conversation context doesn't.
/compact or /clear
Same three phases every time, every day this week. Vocabulary we'll use through Days 2–4.
framing via Steve Pocock
Day 1 · Plan / Execute / ClearIII.9
Plan-mode anatomy
What plan mode looks like in your terminal
Plan mode
1. Read codebook.md to confirm variable names
2. Load data/ipeds_panel.csv with read_csv
3. Profile: nrow, summary, missingness rates
4. Save profile to results/profile.txt
5. Report top findings inline
This plan requires approval
Do you want to proceed?
> 1. Yes, run the plan
2. No, let me give feedback to refine the plan
3. Cancel
Esc to cancel · Tab to amend · Shift+Tab+Tab toggles plan mode
What’s happening
You hit Shift+Tab+Tab to enter plan mode. Claude proposes the steps it would take, before doing any of them.
What you see
A numbered list of discrete actions. Each step is something you could approve or strike on its own.
What you do
1 approves the plan and exits plan mode into execute. 2 sends Claude back with your feedback. 3 cancels.
When to use it
Any task with more than one step. Cleaning a dataset, multi-file refactors, anything where the wrong first move sends Claude in a wrong direction.
Day 1 · Plan-mode anatomyIII.10
§ IV
Permissions
Permission model
Claude Code asks before doing anything that changes your system
Reads
File contents, project structure, git status. Looks at things without changing them.
Runs silently. Reads do not change anything, so no prompt.
Edits
Creating, editing, or deleting files. Changes things on your computer.
Asks each time. Shows the proposed change first. Approve, decline, or always-approve for this project.
Shell
Runs terminal commands like Rscript, git, rm. (A shell is the program that runs your terminal commands.)
Asks before every command, including destructive ones like rm or git reset. Hit Ctrl+E to have Claude explain a command before you approve.
By default Claude Code asks before almost every action. Use /permissions to pre-approve commands as you build trust.
Day 1 · Permission modelIV.1
Anatomy
What a permission prompt looks like in your terminal
A terminal is the text-based interface to your computer. Claude Code prints prompts like this one when it needs your approval.
Bash command
Rscript code/01_clean.R
Run the data cleaning script
This command requires approval
Do you want to proceed?
> 1. Yes
2. Yes, and don't ask again for: Rscript *
3. No
Esc to cancel · Tab to amend · ctrl+e to explain
What's happening
Claude paused before running a Bash command and is asking for approval.
What you see
The exact command, then a one-line description of what it's for. Read both before approving.
What you do
1 runs once. 2 auto-allows the pattern (any Rscript *) for this project. 3 declines.
Shortcuts
Esc cancels. Tab lets you edit the command before approving. Ctrl+E asks Claude to explain what the command does.
Day 1 · AnatomyIV.2
Graduated path
Three layers from least to most permissive. Pick the lowest one that fixes the friction.
The annoyance of repeated prompts is a real cost. The right fix is to remove the prompts you do not need, not to remove all the gates.
Layer 1. Allowlist specific commands.permissions.allow
Stops prompts only for commands you list. Read-only or scoped enough that you would always approve: Read, Grep, Bash(git status), Bash(Rscript *). Risky commands (rm, git push --force, git reset --hard) still prompt.
Edit and Write tools no longer prompt. Bash and other tools still prompt. Useful when edit prompts are the main annoyance and you trust your /rewind ability. Lint and Stop hooks still fire.
Layer 3. Auto mode.defaultMode: "auto"recommended starting point
Built-in classifier allows operations it judges safe and prompts on operations it judges risky. The classifier is decent but not perfect. Strictly better than skipping all gates, because it still gates the actually-dangerous calls.
Layer 4. Skip everything.--dangerously-skip-permissionsnot for routine work
Removes every gate: any shell command (rm, git push --force, sudo), any file write, and any instruction injected into a document Claude reads. Reserve for a worktree or sandbox where the blast radius is contained.
Day 1 · Graduated pathIV.3
§ V
Slash commands
Why slash commands
Plain language goes to the model. Slash commands stop at the harness.
Three things only slash commands can do: clear context, show your cost, edit permissions.
Day 1 · Why slash commandsV.1
Slash commands
The most important slash commands
/help
Show help and available commands.
/clear
Start a new conversation. Past sessions stay in /resume.
/compact
Summarize the conversation to free context.
/resume
Resume a previous conversation by name or picker.
/usage
Show session cost, plan limits, activity stats. /cost is an alias.
/model
Switch models for the current session.
/permissions
Manage allow / ask / deny rules for tool permissions.
/init
Initialize the project with a starter CLAUDE.md.
/exit
End the session.
+ advanced (later this week): /rewind · /loop · /rc
Day 1 · Slash commandsV.2
§ VI
Tooling recommendations
Language tools
R and Python work better with Claude than Stata or SAS
R / Python
Plays well with Claude.
Training data
An enormous amount of R and Python code is on the public web, so Claude writes both fluently.
Terminal access
Both run from simple terminal commands. The agent can launch them, watch the output, and iterate on its own.
Lots of training data. Easy to drive from the terminal.
Stata · SAS · SPSS
Works, but less smoothly.
Training data
Less of this code lives on the public web, so Claude’s output is rougher and needs more correction.
Terminal access
These tools are built for clicking through menus. The agent can run a script but cannot drive the GUI.
Less training data. Harder to drive from the terminal.
You can still use Stata, SAS, or SPSS this week. Just expect rougher code and more cleanup work landing on you.
Day 1 · Language toolsVI.1
Companion tools
Two non-agentic tools worth installing this week
Dictation
Talk, don’t type
Hold a key, speak, release. Text drops into your prompt. About 3x faster for long instructions.
Built-in dictationfree; built into the OS. Raw transcript.
Terminal
A better shell
A modern terminal adds quality-of-life features: tabs for multiple sessions, autocomplete on commands, split panes for side-by-side work, clickable links and file paths, and cleaner handling of long Claude Code output. The built-in Terminal.app still works.
Claude.ai
