Filesystem-Based Context Engineering

<mission_control> <objective>Manage filesystem context for unlimited capacity through dynamic discovery, offloading large context to files and persisting state between sessions</objective> <success_criteria>Context efficiently managed with write-once/read-selective patterns, reducing token usage while maintaining full data availability</success_criteria> </mission_control>

<trigger>When offloading large context to files, persisting state between sessions, or reducing token usage. Not for: Small temporary data or in-memory variables.</trigger>

<interaction_schema> WRITE_ONCE → READ_SELECTIVELY → DISCOVER_DYNAMICALLY </interaction_schema>

The filesystem provides unlimited context capacity through dynamic discovery. Instead of stuffing everything into the context window, agents write once and read selectively, pulling relevant context on demand.

Core Concept

Problem: Context windows are limited but tasks often require more information than fits

Solution: Use filesystem as persistent layer where agents:

1. Write once (large outputs, state, plans)
2. Read selectively (targeted retrieval via search)
3. Discover dynamically (find relevant files on-demand)

Benefit: Unlimited context capacity with natural progressive disclosure

Context Engineering Patterns

Pattern 1: Filesystem as Scratch Pad

Problem: Tool calls return massive outputs (10k+ tokens for web search, hundreds of rows for database queries). If this enters message history, it remains for entire conversation, bloating tokens and degrading attention.

Solution: Write large tool outputs to files instead of returning to context. Agent uses targeted retrieval to extract only relevant portions.

Implementation:

def handle_tool_output(output: str, threshold: int = 2000) -> str:
    if len(output) < threshold:
        return output  # Small output, return directly

    # Write to scratch pad
    file_path = f"scratch/{tool_name}_{timestamp}.txt"
    write_file(file_path, output)

    # Return reference with summary
    summary = extract_summary(output, max_tokens=200)
    return f"[Output written to {file_path}. Summary: {summary}]"

Usage:

• Web search results →

  scratch/web_search_20260126_143022.txt

• Database queries →

  scratch/db_query_users_active.txt

• API responses →

  scratch/api_response_20260126_143045.json

Benefits:

• Reduces token accumulation over long conversations
• Preserves full output for later reference
• Enables targeted retrieval instead of carrying everything
• Natural progressive disclosure

Pattern 2: Plan Persistence

Problem: Long-horizon tasks require plans. But as conversations extend, plans fall out of attention or get lost to summarization. Agent loses track of objectives.

Solution: Write plans to filesystem. Agent can re-read plan anytime to re-orient.

Implementation:

# scratch/current_plan.yaml
objective: "Refactor authentication module"
status: in_progress
steps:
  - id: 1
    description: "Audit current auth endpoints"
    status: completed
  - id: 2
    description: "Design new token validation flow"
    status: in_progress
  - id: 3
    description: "Implement and test changes"
    status: pending

progress:
  current_step: 2
  blockers: ["Waiting for security review"]
  next_action: "Complete token validation design"

Usage:

• Agent reads

  scratch/current_plan.yaml

  at start of each turn
• Updates progress as work completes
• Re-orients when context degrades

Benefits:

• Maintains objective visibility throughout long tasks
• Survives context compaction
• Enables "manipulating attention through recitation"

Pattern 3: Sub-Agent Communication via Filesystem

Problem: In multi-agent systems, sub-agents report to coordinator through message passing. This creates "telephone game" where information degrades through summarization at each hop.

Solution: Sub-agents write findings directly to filesystem. Coordinator reads files directly, bypassing intermediate passing.

Implementation:

workspace/
  agents/
    research_agent/
      findings.md        # Research agent writes here
      sources.jsonl      # Source tracking
    code_agent/
      changes.md         # Code agent writes here
      test_results.txt   # Test output
  coordinator/
    synthesis.md         # Coordinator reads outputs, writes synthesis

Usage:

• Research agent:

  workspace/agents/research_agent/findings.md

• Code agent:

  workspace/agents/code_agent/changes.md

• Coordinator: reads both, writes synthesis

Benefits:

• Preserves fidelity (no telephone game)
• Reduces coordinator context accumulation
• Enables asynchronous collaboration
• Natural audit trail

Pattern 4: Dynamic Skill Loading

Problem: Agents may have many skills/instructions, but most irrelevant to any given task. Stuffing all into system prompt wastes tokens and can confuse with contradictory guidance.

Solution: Store skills as files. Include only skill names/brief descriptions in static context. Load relevant skill content when task requires it.

Implementation:

Available skills (load with read_file when relevant):

- database-optimization: Query tuning and indexing strategies
- api-design: REST/GraphQL best practices
- testing-strategies: Unit, integration, and e2e patterns
- security-review: OWASP Top 10, authentication patterns

Usage:

# Agent working on database task
skill_content = read_file("skills/database-optimization/SKILL.md")

# Agent working on API task
skill_content = read_file("skills/api-design/SKILL.md")

Benefits:

• Minimal static context
• On-demand skill activation
• No contradictory guidance
• Scales to hundreds of skills

Pattern 5: Terminal and Log Persistence

Problem: Terminal output from long-running processes accumulates rapidly. Copying/pasting into agent input is manual and inefficient.

Solution: Sync terminal output to files automatically. Agent greps for relevant sections without loading entire histories.

Implementation:

# Auto-sync terminal to file
script -c "npm run dev" scratch/terminal.log

# Agent searches for specific patterns
grep "ERROR" scratch/terminal.log
grep -A5 "failed" scratch/terminal.log

Benefits:

• Automatic log capture
• Targeted error finding
• No manual copy/paste
• Historical terminal access

Filesystem Navigation

Discovery Patterns

Find files by name:

Glob patterns:
- "**/*.yaml" - All YAML files
- "**/scratch/*" - Scratch pad directory
- "**/plans/*" - Plan files
- "**/logs/*" - Log files

Search file contents:

Grep patterns:
- "TODO|FIXME|BUG" - Find action items
- "ERROR|Exception" - Find errors
- "summary|conclusion" - Find summaries
- "^# .*" - Find headings

Targeted reading:

Read specific sections:
- First 50 lines: `read_file(path, limit=50)`
- Last 50 lines: `read_file(path, offset=-50)`
- Around pattern: `grep(pattern)`, then `read_file(path, offset=X, limit=Y)`

File Metadata Hints

File sizes suggest complexity:

• Small (<1KB): summaries, metadata
• Medium (1-10KB): full outputs, reports
• Large (>10KB): raw data, logs

Naming conventions:

• YYYYMMDD_HHMMSS_*

  - Timestamped files

• *_summary.*

  - Summarized outputs

• *_raw.*

  - Raw data

• current_*.*

  - Current state

Timestamps:

• Newer files likely more relevant
• Timestamps show activity patterns
• Enable time-based filtering

JSONL Append-Only Design

Pattern: All logs use JSONL (JSON Lines) format

Benefits:

• Agent-friendly parsing
• History preservation
• Pattern analysis capability
• Never-delete integrity

Example:

{"timestamp": "2026-01-26T14:30:00Z", "type": "post", "content": "...", "status": "published"}
{"timestamp": "2026-01-26T14:35:00Z", "type": "contact", "name": "Sarah", "updated": true}
{"timestamp": "2026-01-26T14:40:00Z", "type": "plan_update", "step": 2, "status": "completed"}

Reading JSONL:

# Read as list of dicts
logs = [json.loads(line) for line in open('logs.jsonl')]

# Filter by type
posts = [log for log in logs if log['type'] == 'post']

# Query by timestamp
recent = [log for log in logs if log['timestamp'] > '2026-01-26']

Progressive Disclosure in Filesystem

Level 1: Metadata

# Component Index

skills/my-skill/
├── overview.yaml # 200 tokens - auto-loaded
├── trigger_phrases.md # 100 tokens - auto-loaded
└── references/ # On-demand
├── examples/
└── patterns/

Level 2: Instructions

# Full Skill (1500 tokens)

- Load when skill activated
- Contains all instructions
- Progressive disclosure enabled

Level 3: Data

# References/ (As needed)

- examples/ - Usage examples
- patterns/ - Implementation patterns
- scripts/ - Automation scripts
- data/ - Sample data

Best Practices

File Organization

1. Use descriptive names:

   auth_plan_20260126.yaml

   not

   plan1.yaml

2. Timestamp files:

   scratch/web_search_20260126_143022.txt

3. Group related files:

   evidence/

   ,

   scratch/

   ,

   context/

4. Use extensions:

   .yaml

   ,

   .jsonl

   ,

   .md

   ,

   .txt

Content Guidelines

1. Write summaries: Extract key info for quick reference
2. Preserve full data: Keep raw outputs for later analysis
3. Use structured formats: YAML, JSONL for machine-readability
4. Include metadata: Timestamps, types, status

Performance

1. Write once, read many: Optimize for read patterns
2. Use targeted reads: Line ranges, not full files
3. Search before read: Grep to find relevant sections
4. Cache frequently accessed: Keep metadata in memory

Example Workflow

Guidelines

1. Write once, read selectively - Don't return large outputs to context
2. Use descriptive names - Enable quick discovery
3. Timestamp files - Show temporal relationships
4. Preserve full data - Keep raw outputs for analysis
5. Structure formats - YAML/JSONL for machine-readability
6. Search before read - Grep to find relevant sections
7. Progressive disclosure - Load only what you need
8. Append-only logs - Preserve history for pattern analysis

References

Related Skills:

• context-fundamentals

  - Progressive disclosure principles

• evaluation

  - Multi-dimensional quality assessment

• iterative-retrieval

  - Progressive refinement for targeted context discovery

• filesystem-context

  - Context management patterns (this skill)

For Complex Discovery:

When basic search is insufficient, use iterative-retrieval for targeted context discovery:

# Basic search (grep/glob)
grep("pattern", "**/*.ts")

# Iterative retrieval with progressive refinement
/search "authentication patterns in TypeScript"

iterative-retrieval enhances filesystem-context by:

• 4-phase loop: DISPATCH → EVALUATE → REFINE → LOOP
• Relevance scoring: Identifies which files to read from filesystem
• Progressive refinement: Discovers relevant files systematically
• Termination conditions: Stops when sufficient high-relevance files found

Integration:

• Use iterative-retrieval to discover relevant files
• Use filesystem-context to persist discovered files for selective retrieval
• Combined: Discovery → Storage → Targeted retrieval

Key Principle: Filesystem provides unlimited context capacity through dynamic discovery. Write once, read selectively, discover on-demand.

---

<critical_constraint> MANDATORY: Write large outputs to files, not to context MANDATORY: Use descriptive, timestamped filenames for traceability MANDATORY: Search before reading (find relevant sections first) MANDATORY: Use structured formats (YAML/JSONL) for machine-readability MANDATORY: Never return full large outputs to context (use references) No exceptions. Filesystem context enables unlimited capacity. </critical_constraint>

---

Genetic Code

This component carries essential Seed System principles for context: fork isolation:

<critical_constraint> MANDATORY: All components MUST be self-contained (zero .claude/rules dependency) MANDATORY: Achieve 80-95% autonomy (0-5 AskUserQuestion rounds per session) MANDATORY: Description MUST use What-When-Not format in third person MANDATORY: No component references another component by name in description MANDATORY: Progressive disclosure - references/ for detailed content MANDATORY: Use XML for control (mission_control, critical_constraint), Markdown for data No exceptions. Portability invariant must be maintained. </critical_constraint>

Delta Standard: Good Component = Expert Knowledge − What Claude Already Knows

Recognition Questions:

• "Would Claude know this without being told?" → Delete (zero delta)
• "Can this work standalone?" → Fix if no (non-self-sufficient)
• "Did I read the actual file, or just see it in grep?" → Verify before claiming

---