Filesystem Context Manager

A skill for maintaining context across Claude Code sessions using filesystem-based persistence. Implements the six patterns from Context Engineering to prevent context loss during long or multi-session tasks.

When This Skill Activates

This skill automatically activates when you:

• Start a new session and need to resume previous work
• Work on tasks spanning multiple conversations
• Need to preserve decisions, context, or learnings
• Want to maintain continuity across sessions
• Are working on complex multi-step implementations

Keywords: save context, resume work, continue session, persist context, remember this, don't forget, maintain context, session continuity, pick up where we left off

Core Concepts

The Context Problem

Issue	Impact	Solution
Context window fills up	Early context forgotten	Write to files, read selectively
Session ends mid-task	Progress lost	Persist plans and state
Complex decisions made	Reasoning forgotten	Document decisions in files
Patterns discovered	Must re-learn each session	Store in memory files

File-Based Context Hierarchy

.claude/
├── context/              # Current session state
│   ├── current-task.md   # Active task description
│   ├── decisions.md      # Decisions made this session
│   └── blockers.md       # Current blockers
├── memory/               # Persistent learnings (survives sessions)
│   ├── patterns.md       # Codebase patterns discovered
│   ├── gotchas.md        # Things to remember
│   └── preferences.md    # User preferences
├── scratch/              # Temporary working files
│   └── tool-outputs/     # Large command outputs
├── plans/                # Implementation plans
│   └── *.md              # Individual plan files
└── skills/               # Skill definitions

Pattern 1: Scratch Pad for Large Outputs

When tool outputs exceed ~2000 tokens, write to file instead of keeping in context.

Implementation

# Instead of showing full output in chat
npm run build 2>&1 | tee .claude/scratch/build-output.txt

# Then reference specific parts
grep -n "error" .claude/scratch/build-output.txt

When to Use

• Build outputs with many warnings
• Large database query results
• Extensive grep/search results
• API responses with large payloads

Example Workflow

1. Run command, pipe to scratch file
2. Return summary to chat: "Build completed with 3 errors, see .claude/scratch/build-output.txt"
3. When debugging, grep specific errors from file
4. Delete scratch file when done

Pattern 2: Plan Persistence

Store implementation plans in structured files for continuity.

Plan File Template

# Plan: [Feature/Task Name]

## Metadata
- **Status**: PLANNING | IN_PROGRESS | BLOCKED | COMPLETE
- **Created**: YYYY-MM-DD
- **Updated**: YYYY-MM-DD HH:MM
- **Priority**: HIGH | MEDIUM | LOW

## Objective
[Clear statement of what we're trying to achieve]

## Context
[Background information and why this matters]

## Implementation Steps
- [ ] Step 1: Description
  - Files: `path/to/file.ts`
  - Notes: Implementation details
- [ ] Step 2: Description
- [ ] Step 3: Description

## Files to Modify
| File | Purpose | Status |
|------|---------|--------|
| `app/api/route.ts` | Main endpoint | Pending |
| `lib/service.ts` | Business logic | Done |

## Decisions Made
1. **[Decision]**: [Rationale]
2. **[Decision]**: [Rationale]

## Blockers
- [ ] Blocker 1: Description
- [ ] Blocker 2: Description

## Notes
[Any additional context]

Usage

# Create new plan
echo "# Plan: Feature X" > .claude/plans/feature-x.md

# Resume work - read plan first
cat .claude/plans/feature-x.md

# Update progress
# Edit the plan file as steps complete

Pattern 3: Session Context Persistence

Maintain current session state for continuity.

Current Task File

# Current Task

## What I'm Working On
[Brief description of current focus]

## Recent Actions
- [Action 1] - [Result]
- [Action 2] - [Result]

## Next Steps
1. [Next immediate action]
2. [Following action]

## Files Currently Open/Relevant
- `path/to/file1.ts` - [Why relevant]
- `path/to/file2.ts` - [Why relevant]

## Session Notes
[Any important context for this session]

Decisions File

# Session Decisions

## [Date]

### Decision: [Title]
- **Context**: [Why this came up]
- **Options Considered**:
  - Option A: [Description]
  - Option B: [Description]
- **Chosen**: Option A
- **Rationale**: [Why this was chosen]
- **Impact**: [What this affects]

Pattern 4: Memory Persistence

Store learnings that should survive across all sessions.

Patterns File

# Codebase Patterns

## Authentication
- Always use `createClient` from `@/lib/supabase/server` for API routes
- Check BOTH Authorization header AND cookies (see commit ac642e8)

## API Routes
- Next.js 15 requires async params: `context: { params: Promise<{ id: string }> }`
- Always await params before use

## Database
- Use service role client for admin operations
- RLS policies apply to anon key only

## Testing
- Run `npm run type-check:memory` to avoid heap errors
- Use `npm run dev:memory` for development server

Gotchas File

# Gotchas & Lessons Learned

## [Category]

### [Issue Title]
- **Symptom**: [What you see]
- **Cause**: [Why it happens]
- **Fix**: [How to resolve]
- **Prevention**: [How to avoid]
- **Reference**: [Commit/PR/Doc]

Preferences File

# User Preferences

## Code Style
- Prefer explicit types over inference
- Use early returns for guard clauses
- Keep functions under 50 lines

## Workflow
- Always run type-check before committing
- Use conventional commits (feat:, fix:, etc.)
- Test in staging before production

## Communication
- Be concise, skip unnecessary explanations
- Show code diffs for changes
- Summarize at end of complex tasks

Pattern 5: Sub-Agent Communication

When using multiple agents, use filesystem for state sharing.

Agent Workspace Structure

.claude/
└── agents/
    ├── research/         # Research agent outputs
    │   └── findings.md
    ├── implementation/   # Implementation agent state
    │   └── progress.md
    └── review/           # Review agent feedback
        └── issues.md

Handoff Protocol

# Agent Handoff: [From Agent] → [To Agent]

## Completed Work
[What was accomplished]

## Key Findings
1. [Finding 1]
2. [Finding 2]

## Files Created/Modified
- `path/to/file.ts` - [What was done]

## Recommendations for Next Agent
1. [Recommendation]
2. [Recommendation]

## Open Questions
- [Question needing resolution]

Pattern 6: Self-Updating Memory

Capture learnings automatically during sessions.

Auto-Capture Triggers

When these occur, update memory files:

• Error resolved after debugging → Add to gotchas.md
• Pattern discovered in codebase → Add to patterns.md
• User states preference → Add to preferences.md
• Workaround found → Add to gotchas.md

Update Protocol

# Append new learning to appropriate file
echo "
### [New Learning Title]
- **Context**: [When this applies]
- **Details**: [The learning]
- **Added**: $(date +%Y-%m-%d)
" >> .claude/memory/patterns.md

Quick Start Commands

Start New Session

# 1. Check for existing context
cat .claude/context/current-task.md 2>/dev/null || echo "No active task"

# 2. Check active plans
ls -la .claude/plans/*.md 2>/dev/null || echo "No active plans"

# 3. Review memory
cat .claude/memory/gotchas.md 2>/dev/null | head -50

Save Session Context

# Before ending session, update current-task.md with:
# - What was accomplished
# - What's next
# - Any blockers

Resume Previous Work

# 1. Read the plan
cat .claude/plans/[plan-name].md

# 2. Read current context
cat .claude/context/current-task.md

# 3. Check recent decisions
cat .claude/context/decisions.md | tail -30

Integration with Existing Skills

Skill	Integration
context-manager	Use scratch/ for large context analyzer outputs
project-sync	Store sync status in context/sync-status.md
bug-fixing	Document debugging journey in context/debug-log.md
session-manager	Named sessions map to plan files

Best Practices

1. Update incrementally - Don't wait until session end
2. Be specific - Vague notes don't help future sessions
3. Clean up scratch/ - Delete temporary files when done
4. Review memory/ - Periodically prune outdated information
5. Use consistent format - Templates make parsing easier
6. Reference file paths - Future sessions need to find relevant code
7. Date entries - Know when context was captured
8. Keep plans updated - Stale plans cause confusion

Cleanup Commands

# Clean old scratch files (older than 7 days)
find .claude/scratch -type f -mtime +7 -delete

# Archive completed plans
mv .claude/plans/completed-*.md .claude/plans/archive/

# Prune old context
# Keep only last 5 days of context files

---

Version: 1.0.0 Last Updated: 2025-01-08 Maintained By: CircleTel Development Team Based On: https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering