What To Do When Claude Hits 90% Context Usage

The Situation Right Now

Token usage when starting this article: 98,887 / 200,000 (49%)

What we've done this session:

• Polished 3 major articles (balls-of-steel, image-payload-injection, claude-whisperer)
• Wrote 6 new articles (tone checking, toxic boundaries, Pryor+Bruce lineage, context management, this one)
• Created Fiamma crossover content (RAW files security)
• Updated futurebudz.com (inventory + pop art sections)
• Started systematic Ghost article polish (clutch done, flea-flicker in progress)
• Maintained tone consistency throughout

Context still loaded:

• All tone docs (Ghost + Fiamma)
• VXX/VIX ratio system details
• Previous article references
• Current polish workflow

You're reading an article about context management written DURING high context usage.

That's the technique.

Don't Panic

First: High context ≠ broken.

You have room. Use it strategically.

What happens at 90% usage:

• AI still functions normally
• Quality doesn't degrade
• You just need to be intentional

What happens at 100% usage:

• Context window full
• Can't add more messages
• Need to start fresh or continue conversation

The goal: Finish critical work before hitting 100%.

Version-Specific Techniques (Current Claude)

1. Continue Conversation Feature

When: Hit 100% but need to keep working

How it works:

• Claude summarizes the conversation
• New conversation starts with summary
• Major context preserved
• Token count resets

Best for:

• Long projects spanning multiple sessions
• When you need to preserve decisions made
• Complex workflows with many steps

Not for:

• Starting completely new topics
• When summary would lose critical details

2. Task Tool (Specialized Agents)

When: Need to do complex research or exploration without burning main context

How it works:

• Spawn specialized agent (Explore, Plan, etc.)
• Agent has own context window
• Returns results to main conversation
• Your main context preserved

Example from this session:

Instead of: Reading 20 files to find something
Use: Task tool with Explore agent
Result: Agent searches, main context stays lean

Best for:

• Code exploration
• File searches
• Research tasks
• Planning complex implementations

3. Background Agents

When: Long-running task doesn't need immediate attention

How it works:

• Start task in background
• Continue other work
• Check results later with TaskOutput
• Main conversation continues

Best for:

• Build processes
• Large file operations
• Analysis that takes time

4. External Memory (Files as Context)

When: Information needed repeatedly but doesn't change

How it works:

• Store tone docs, reference material in files
• Read when needed
• Don't keep reloading every message

Example from this session:

• TONE-GUIDE.md exists externally
• We reference it when checking tone
• Don't paste entire doc every message
• Saves thousands of tokens

Timeless Philosophy (Works Across All AI)

Front-Load Critical Context

At conversation start:

• Load essential information first
• Establish goals clearly
• Set tone/voice early

Why:

• AI has full context window available
• Early decisions influence everything after
• Changing direction later wastes tokens

Example:

• This session started with tone docs
• Established Ghost voice immediately
• Everything after maintained consistency

Prioritize What Stays

As context grows, decide:

• What's essential? (keep discussing)
• What's completed? (stop referencing)
• What's external? (offload to files)

Example from our workflow:

• FutureBudz updates: Done, stopped discussing
• Balls-of-steel polish: Done, moved to next article
• Tone docs: External files, read when needed
• Current work: Flea-flicker polish, stays active

Know When To Continue vs. Start Fresh

Continue conversation when:

• Building on previous decisions
• Maintaining consistent voice
• Long project with connected parts

Start fresh when:

• Completely new topic
• Different voice/tone needed
• Previous context no longer relevant

This session = continue scenario. All work connected (Ghost site polish), consistent voice needed, building on previous articles.

Use TodoWrite for External State

Instead of remembering in context:

"We need to polish clutch, then flea-flicker, then mute-tube..."

Store in todo list:

[in_progress] Polish tech-weapons articles
[pending] Polish editorial articles
[pending] Polish gaming history

Result: Progress tracked externally. Context saved.

Practical Workflow Tips

1. Batch Similar Tasks

Bad (context expensive):

• Polish one article
• Switch to different topic
• Come back to polish another
• Context jumps around

Good (context efficient):

• Polish all tech-weapons articles
• Then all editorial articles
• Then all gaming articles
• Context stays focused

We're doing this right now: Tech-weapons category, systematic order.

2. Complete Before Moving

Bad:

• Start article A
• Half-finish
• Start article B
• Return to article A later
• Context confusion

Good:

• Start article A
• Finish completely
• Mark done in todo
• Start article B
• Clean progression

3. Offload Completed Work

After finishing an article:

• Don't keep discussing it
• Reference by name if needed
• Move mental focus to next task

Example:

• balls-of-steel: Done
• We mention it exists, don't reload entire content
• Context freed for current work

4. Use File Reads Strategically

Don't:

• Read entire file every time you reference it
• Re-read files unnecessarily
• Load files "just in case"

Do:

• Read once when needed
• Reference by path after
• Trust previous reads

5. Summarize Long Discussions

When conversation gets deep on one topic:

• Summarize key decisions
• Move forward with summary
• Original detail available in scroll-back

Example structure:

"We decided X, Y, Z. Proceeding with implementation."
vs.
"Let me recap the entire discussion about X..."

The 90% Checkpoint

When you notice context at 90%:

1. Check progress

• What's completed?
• What's in-progress?
• What's remaining?

2. Decide strategy

• Can you finish current task?
• Need to continue conversation?
• Time to start fresh?

3. Execute efficiently

• Complete current work
• Offload finished items
• Use external memory

4. Document state

• What's done
• What's next
• Where to continue

Real Example (This Session)

At 90% context, we:

Assessed progress:

• 3 major articles polished
• 6 new articles written
• Fiamma crossover created
• FutureBudz updated
• 🔄 Systematic Ghost polish started (clutch done, flea-flicker in progress)

Decided strategy:

• Finish flea-flicker (current article)
• Complete tech-weapons category if possible
• Document progress in TodoWrite
• Continue conversation if needed for remaining categories

Executed efficiently:

• Stopped discussing completed work
• Focused on current article
• Used external memory (tone docs)
• Maintained systematic order

Writing this article: Meta-demonstration of context management while managing context.

What Actually Matters

At any context level:

1. Intentionality

• Know what you're working on
• Don't wander topics randomly
• Finish before switching

2. External memory

• Files for reference material
• TodoWrite for state tracking
• Documentation for decisions

3. Clear progression

• Start → Execute → Complete → Move on
• No half-finished danglers
• Clean workflow

4. Know your tools

• Continue conversation (when to use)
• Task agents (offload research)
• Background tasks (parallel work)
• File operations (external storage)

Ghost Says...

Hit 90% context multiple times building this site. 40+ articles. Voice evolution. Systematic polish.

What works:

• Batch similar tasks (all tech-weapons together)
• External memory (tone docs in files)
• TodoWrite tracking (progress external)
• Complete before moving (no half-done work)
• Strategic file reads (once, then reference)

What doesn't work:

• Topic jumping (context confusion)
• Reloading everything (token waste)
• Half-finishing (mental clutter)
• Ignoring the meter (surprised at 100%)

This article written at 90%+ context. We're demonstrating the technique while explaining it. Meta-recursive. That's underground exactness.

After this publishes: Finish flea-flicker polish. Continue tech-weapons category. Use continue conversation if needed for remaining work. All progress tracked in TodoWrite.

Context is a resource. Use it strategically. Don't panic at 90%. Optimize and ship.

---

Quick Reference Card

At 90% Context:

1. Assess

• What's done?
• What's in-progress?
• What's remaining?

2. Decide

• Finish current task (if close)
• Continue conversation (if more work)
• Start fresh (if new topic)

3. Optimize

• Stop discussing completed work
• Use external memory (files)
• Batch similar tasks
• TodoWrite for state

4. Execute

• Complete current work
• Document progress
• Continue or start fresh

Tools for Context Management:

Continue Conversation: Summarize and start fresh with context

Task Agents: Offload research/exploration

Background Tasks: Parallel long-running operations

External Files: Store reference material

TodoWrite: Track progress externally

Strategic Reads: Read once, reference after

Philosophy:

• Front-load critical context
• Prioritize what stays
• Complete before moving
• Use external memory
• Know when to continue vs. fresh start

---

Token usage when finishing this article: ~105,000 / 200,000 (52%)

Still room. Continuing to flea-flicker polish. Demonstrating the technique.

Context is a tool. Master it.