SKILL.md

--- frontmatter

name: context-management
description: |
  Strategies for managing context/token budget in long conversations.
  Use when working on large codebases, extensive analysis, or multi-step tasks.
  Triggers: long tasks, large files, multiple searches, "running low on context", complex exploration

Context Management Strategies

Universal patterns for managing token budget efficiently in long conversations.

Core Principles

•Search before read - Find what you need, don't read everything
•Progressive disclosure - Overview → Sample → Targeted deep-dive
•Summarize as you go - Extract key points, discard verbose content
•Save large outputs to files - Keep conversation context lean
•Proactive awareness - Monitor usage, warn before limits

Smart Reading Strategy

For Any Large Content

code

Step 1: List/Search → Identify what exists
Step 2: Quick scan → Read headers/summaries only
Step 3: Relevance ranking → Prioritize what matters
Step 4: Targeted reading → Deep dive only on relevant sections
Step 5: Synthesize → Create concise summary, reference files for details

Progressive Disclosure Pattern

Phase	Context Cost	Action
Discovery	Low	List files, search keywords
Sampling	Medium	Read first N lines, headers only
Targeted	Medium-High	Read specific sections of interest
Full	High	Only if absolutely necessary

File-Based Output Strategy

When outputs would be large (>100 lines, >1000 rows):

markdown

✅ GOOD: Save to file, show summary
"Saved 1,500 results to output/analysis.csv
Summary: Top 10 items by revenue..."

❌ BAD: Dump everything into conversation
[1,500 rows of data filling context...]

When to Save to Files

•Query results > 100 rows
•Code generation > 100 lines
•Analysis with detailed breakdowns
•Multi-step outputs with intermediate results
•Reference material for later use

Context Checkpoints

Periodically assess context usage:

markdown

**Context Check**: ~60% used
- Explored: 15 files ✓
- Key findings: documented ✓
- Remaining work: 3 tasks

**Action**: Continuing normally

markdown

**Context Check**: ⚠️ ~85% used

**Saving state**:
- Analysis summary → saved to scratch/analysis_summary.md
- Query history → saved to scratch/queries.sql

**Options**:
A) Run `/compact` to continue with fresh context
B) Focus on specific remaining question
C) Wrap up with executive summary

Proactive Warnings

Before large operations, warn:

markdown

"This query will return ~50,000 rows.

**Context Management Plan**:
✓ Execute query
✓ Save full results to CSV
✓ Show summary statistics in chat
✓ Display top 20 rows as preview

This keeps our conversation efficient. Proceed?"

Best Practices Checklist

Before each major read:

• Will this return >100 lines? → Sample first
• Multiple files to read? → Search for relevant ones
• Already at 70% context? → Save next output to file

During exploration:

• Check context every 5-10 operations
• Save intermediate findings to files
• Aggregate instead of showing raw data

When context is low:

• Immediately save current state
• Summarize findings so far
• Suggest /compact or focused continuation
• Offer clear next steps

Recovery Patterns

When You Hit the Limit

markdown

⚠️ Context limit approaching

**Saved to files**:
- findings.md - Key discoveries
- next_steps.md - Remaining work

**Summary**: [Key points that survive context reset]

**To continue**: Run `/compact`, then reference saved files

Handoff Format

When context resets are needed, save:

markdown

# Session Handoff

## Completed
- [What was accomplished]

## Key Findings
- [Important discoveries]

## Remaining Work
- [What still needs to be done]

## Files to Reference
- path/to/relevant/files

Remember

Context is a precious resource. Be surgical: search to find, sample to assess, read only what matters. Save verbose outputs to files and keep the conversation focused on insights and decisions, not raw data.