Context Window Management

What You'll Learn

• What a "context window" actually means
• How to check your context usage in Claude Code
• When to /recap and start fresh
• The 80% rule (and why it matters)
• Founder-specific context budgets for common tasks

---

Why This Matters (3 min read)

The Scenario:

You're validating a product idea. You ask Claude to:

1. Read 10 competitor websites
2. Analyze 20 customer interviews
3. Generate a product spec
4. Write marketing copy

After step 2, Claude says: "I'm approaching my context limit. Consider using /recap to free up space."

You're confused. Claude just... ran out of memory? You didn't even finish your task.

---

What Happened:

Claude Code has a context window — a limit on how much information it can "remember" at once.

Think of it like RAM on your computer:

• Small project: Uses 20% context → Works great
• Medium project: Uses 60% context → Still fine
• Large project: Uses 90% context → Claude struggles, starts forgetting things

When context is full, Claude can't read more files, can't write detailed outputs, can't help effectively.

You need to manage context like a budget.

---

What Is a Context Window?

Simple Definition:

The context window is everything Claude can "see" in the current session.

This includes:

• Your prompts (what you type)
• Files Claude has read
• Claude's responses
• Previous conversation history
• Your CLAUDE.md file (if present)

Size:

• Claude Code (Sonnet 4.5): 200,000 tokens (~150,000 words)
• That's roughly 500 pages of text

Sounds like a lot, right?

It is... until you start reading large files.

---

What Uses Context (Token Costs)

| Action | Approximate Token Cost | % of 200K Window | |--------|------------------------|------------------| | Your prompt (100 words) | ~130 tokens | 0.07% | | Claude's response (500 words) | ~650 tokens | 0.33% | | Read 1 product spec (2,000 words) | ~2,600 tokens | 1.3% | | Read 10 customer interviews (20,000 words) | ~26,000 tokens | 13% | | Read entire codebase (100 files) | ~50,000-100,000 tokens | 25-50% | | CLAUDE.md file (500 words) | ~650 tokens | 0.33% |

The Math:

Let's say you're doing customer research:

• Read 10 interviews: 26,000 tokens (13%)
• Ask Claude to synthesize: 130 tokens (0.07%)
• Claude writes analysis: 2,600 tokens (1.3%)
• You ask follow-up questions (5 rounds): 3,900 tokens (2%)
• Total: ~32,600 tokens (16.3%)

Still plenty of room!

---

But then you ask Claude to:

• Read competitor websites (10 sites, 50,000 tokens = 25%)
• Generate product spec (5,000 tokens = 2.5%)
• Write landing page copy (3,000 tokens = 1.5%)

New total: ~91,000 tokens (45.5%)

You're almost halfway done. And you haven't even started building yet.

---

How to Check Context Usage in Claude Code

Method 1: The /usage Command (Easiest)

In Claude Code, type:

/usage

Claude will show:

Context usage: 45,234 / 200,000 tokens (22.6%)

When to check:

• Before starting a big task
• After reading many files
• When Claude seems slow or forgetful

---

Method 2: Watch the Prompt Box Footer (VSCode Extension)

If using the VSCode extension, the prompt box footer shows:

📊 Context: 22.6%

Updates in real-time as you work.

---

Method 3: Look for Claude's Warnings

Claude will tell you when context is getting full:

⚠️ "I'm approaching my context limit. Consider using /recap to free up space."

This is your signal to act.

---

The 80% Rule (Critical)

Rule: When context hits 80% (160,000 / 200,000 tokens), stop and /recap.

Why 80% (not 100%)?

1. Quality degrades: At 80%+, Claude starts "forgetting" earlier parts of the conversation (the "lost in the middle" problem)
2. Slower responses: High context = slower processing
3. Risk of cutoff: At 100%, Claude CANNOT read more files or write long outputs
4. Safety margin: Leave room for Claude's response + your next prompt

Bad:

Context: 95% → Ask Claude to generate 5-page spec → Claude says "I can't, context full"

Good:

Context: 75% → /recap → Start fresh session → NOW ask for 5-page spec

---

When to /recap and Start Fresh

Use /recap When:

✅ Context hits 75-80% (safety margin) ✅ Switching to a new task (different project, different goal) ✅ Claude seems "confused" (forgetting earlier instructions) ✅ You've been working for 2+ hours (natural break point) ✅ Before a big read (about to load 50 customer interviews)

DON'T /recap When:

❌ Context is under 50% (plenty of room) ❌ Mid-task (Claude is generating a spec, don't interrupt) ❌ Every 10 minutes (unnecessary resets hurt continuity)

---

What /recap Does

When you run /recap:

1. Claude summarizes the current session (key decisions, outputs, next steps)
2. Saves recap to a file (you can reference it later)
3. Starts fresh (context reset to ~0%)
4. You can resume by reading the recap file

Example:

textCopy
You: /recap

Claude:
## Session Recap (2026-02-08)

### What We Did
- Validated product idea: B2B team collaboration tool
- Read 10 customer interviews
- Identified top 3 pain points: [list]
- Generated product spec v1 (saved to specs/product-spec-v1.md)

### Key Decisions
- Target: Remote teams, 10-50 people
- Pricing: $49/user/month (validated with 5 interviewees)

### Next Steps
1. Write landing page copy (use specs/product-spec-v1.md as input)
2. Create competitive positioning doc
3. Draft email sequence for beta signups

Recap saved to: session-recaps/2026-02-08-product-validation.md

Now:

• Context reset to 0%
• Fresh session starts
• You can say: "Claude, read session-recaps/2026-02-08-product-validation.md and continue where we left off"

---

Founder-Specific Context Budgets

Here's how much context typical founder tasks use:

Task 1: Product Spec Generation (Low Context)

Steps:

1. You describe idea (200 words): ~260 tokens
2. Claude asks clarifying questions: ~400 tokens
3. You answer: ~300 tokens
4. Claude generates spec (2,000 words): ~2,600 tokens

Total: ~3,600 tokens (1.8%)

Verdict: ✅ Plenty of room. Can do 20+ specs before hitting 80%.

---

Task 2: Customer Research Synthesis (Medium Context)

Steps:

1. Read 10 interview transcripts (20,000 words): ~26,000 tokens
2. Ask Claude to find patterns: ~150 tokens
3. Claude analyzes + writes report (3,000 words): ~3,900 tokens
4. You ask 5 follow-up questions: ~2,000 tokens

Total: ~32,000 tokens (16%)

Verdict: ⚠️ Moderate. Can do 3-4 synthesis tasks before hitting 80%.

---

Task 3: Competitive Analysis (High Context)

Steps:

1. Read 10 competitor websites (scrape text, 50,000 words): ~65,000 tokens
2. Ask Claude to compare: ~200 tokens
3. Claude generates analysis (5,000 words): ~6,500 tokens
4. Follow-ups: ~3,000 tokens

Total: ~75,000 tokens (37.5%)

Verdict: ⚠️ Heavy. Only 2 analyses per session. /recap after each.

---

Task 4: Content Generation (Low-Medium Context)

Steps:

1. Read brand guide (1,000 words): ~1,300 tokens
2. Read product spec (2,000 words): ~2,600 tokens
3. Ask Claude to write landing page (500 words copy): ~2,000 tokens
4. Iterate 3 times: ~3,000 tokens

Total: ~9,000 tokens (4.5%)

Verdict: ✅ Light. Can generate 10+ pieces of content per session.

---

Pro Tips for Managing Context

Tip 1: Read Selectively (Don't Read Everything)

❌ Bad:

textCopy
You: "Claude, read everything in research/"
Claude: [reads 50 files, 100,000 tokens, context at 50%]

✅ Good:

textCopy
You: "Claude, read only research/interviews/2025-01-* (this month's interviews)"
Claude: [reads 10 files, 20,000 tokens, context at 10%]

Why: Only read what you need for THIS task.

---

Tip 2: Use .claudeignore to Exclude Junk

Create a .claudeignore file:

textCopy
# Don't read these
_archive/
drafts/old-*
*.pdf
*.jpg
*.png
node_modules/

Why: Claude skips these automatically. Saves context.

---

Tip 3: Summarize Long Files FIRST

If you have a 10,000-word research doc:

❌ Bad:

textCopy
You: "Claude, read research/mega-doc.md and summarize"
Claude: [reads entire 13,000 tokens]

✅ Good:

textCopy
You: "Claude, read research/mega-doc.md and write a 500-word summary to research/summary.md"
Claude: [reads 13,000, writes 650-token summary]

[NEW SESSION]
You: "Claude, read research/summary.md"
Claude: [reads 650 tokens instead of 13,000]

Savings: 95% context reduction on subsequent uses.

---

Tip 4: Break Big Tasks into Sessions

❌ Bad (One Massive Session):

textCopy
Session 1 (context 0% → 95%):
- Read 10 competitor sites
- Read 20 customer interviews
- Generate product spec
- Write landing page copy
- Create email sequence
- Build competitive matrix
[CONTEXT FULL, CLAUDE STRUGGLES]

✅ Good (Multiple Focused Sessions):

textCopy
Session 1 (context 0% → 35%):
- Read 20 customer interviews
- Synthesize patterns
- Save output
- /recap

Session 2 (context 0% → 40%):
- Read 10 competitor sites
- Generate competitive matrix
- Save output
- /recap

Session 3 (context 0% → 20%):
- Read previous outputs
- Generate product spec
- Write landing page copy
- /recap

Why: Each session stays under 50%. Claude performs better.

---

Tip 5: Use CLAUDE.md for Persistent Context

Put recurring instructions in CLAUDE.md (doesn't count against session context):

markdownCopy
# Project: Product Validation

## Default Instructions
- Always save outputs to analysis/ folder
- Use founder-friendly language (not technical jargon)
- Keep specs under 2 pages
- Show your reasoning

## Common Tasks
- Product spec generation (use templates/product-spec-template.md)
- Customer research synthesis (save to analysis/customer-insights.md)

Why: Claude reads CLAUDE.md at START of every session. No need to repeat instructions, saves context.

---

Try This Now (Exercise — 10 Minutes)

Step 1: Check Your Current Context (1 min)

1. Open Claude Code
2. Type: /usage
3. Note the percentage

Step 2: Deliberately Fill Context (5 min)

1. Ask Claude: "Read all files in this project"
2. Check /usage again
3. See how much context that used

Step 3: Practice /recap (4 min)

1. Type: /recap
2. Claude summarizes the session
3. Note where recap file is saved
4. Start a NEW session
5. Say: "Claude, read [recap file path] and continue"

What You Learned:

• How much context your project uses
• What a recap looks like
• How to continue work after recap

---

Common Mistakes (And Fixes)

Mistake 1: Ignoring Context Warnings

Problem: Claude says "approaching context limit" but you keep asking for more work.

Fix: When Claude warns you, stop and /recap. Don't push to 100%.

---

Mistake 2: Reading Entire Codebases

Problem: "Claude, read my entire project folder" → 100,000 tokens → context blown.

Fix: Read ONLY the files relevant to your current task.

---

Mistake 3: Never Using /recap

Problem: Working for 4 hours straight, context at 95%, Claude "forgetting" earlier instructions.

Fix: /recap every 2 hours or at natural break points.

---

Mistake 4: Recap Too Often

Problem: /recap every 10 minutes → lose continuity → Claude "forgets" mid-task context.

Fix: Only recap when context hits 75-80% OR switching tasks.

---

Downloads

📥 Context Management Cheat Sheet (PDF) — Quick reference guide 📥 .claudeignore Template — Exclude common junk files 📥 Session Recap Template — Structure for manual recaps 📥 Founder Context Budget Calculator (Spreadsheet) — Estimate token usage

---

What's Next

Now you know HOW to manage Claude's memory. Next, let's talk about HOW TO TALK to Claude.

In the next lesson, we'll cover Prompting Fundamentals:

• The anatomy of a good prompt
• Before/after examples (bad → good prompts)
• The "Founder Prompt Formula"
• How to iterate when Claude gets it wrong

→ Next Lesson: Prompting Fundamentals for Founders

---

Key Takeaways

✅ Context window = Claude's working memory (200K tokens ≈ 500 pages) ✅ Check usage with /usage — know where you stand ✅ The 80% rule: Recap at 75-80%, don't push to 100% ✅ Read selectively: Only read files you need for THIS task ✅ Use .claudeignore to exclude junk automatically ✅ Break big tasks into sessions: Keep each under 50% context

---

Sources & Attribution:

This lesson curates content from:

• SparkCo.ai: Mastering Claude's Context Window (2025)
• Eesel.ai: Claude Code Context Window Size Guide
• Anthropic: Context Management Best Practices
• Anthropic: Context Windows Documentation

Additional insights and founder-specific adaptations by NerdSmith.

---

Version: 1.0 (2026-02-08) Module: 0 — Claude Code Bootcamp Lesson: 2 of 5