Context Loader

Context Loader Skill

Efficiently load and manage project context for Conductor's context-driven development workflow.

Trigger Conditions

Use this skill when:

• Starting work on a new track or feature
• User mentions: "load context", "project context", "get context"
• Beginning /orchestrator-supaconductor:implement workflow
• Need to understand project structure without reading all files

Token Optimization Protocol

1. Respect Ignore Files

Before scanning files, check for:

1. .claudeignore - Claude-specific ignores
2. .gitignore - Standard git ignores

# Check for ignore files
ls -la .claudeignore .gitignore 2>/dev/null

2. Efficient File Discovery

Use git for tracked files:

git ls-files --exclude-standard -co | head -100

For directory structure:

git ls-files --exclude-standard -co | xargs -n 1 dirname | sort -u

3. Priority Files (read_file First)

4. Large File Handling

For files over 1MB:

• read_file first 20 lines (header/imports)
• read_file last 20 lines (exports/summary)
• Skip middle content

Enforcement Rules (MANDATORY)

1. Check file size via run_shell_command ls -la before reading

   • 500KB: read_file first 20 + last 20 lines only

   • 1MB: Skip entirely, log as "skipped: too large"

2. Stop after Tier 1-3 files. Tier 4 (config files) only if task-specific.
3. Never load completed tracks — only active track spec.md + plan.md.
4. Maximum 15 files per context load. If more are needed, prioritize by tier.

Context Loading Workflow

1. Load CLAUDE.md (if exists)
2. Load conductor/product.md (project vision)
3. Load conductor/tech-stack.md (technical context)
4. Load conductor/tracks.md (completed work — prevents duplicate effort)
5. Load current track spec.md (requirements)
6. Load current track plan.md (tasks — check [x] vs [ ] status)

Evaluate-Loop Integration

This skill is used by multiple loop agents:

• loop-planner — loads context before creating a plan
• loop-executor — loads context before implementing tasks
• conductor-orchestrator — loads context to determine current loop step

Critical: Always load tracks.md and check plan.md task markers to prevent duplicate work across sessions.

Response Format

After loading context, summarize:

## Project Context Loaded

**Product**: [one-line summary]
**Tech Stack**: [key technologies]
**Current Track**: [track name/id]
**Active Phase**: [current phase]
**Pending Tasks**: [count of [ ] tasks]
**Completed Tasks**: [count of [x] tasks]
**Loop Step**: [current Evaluate-Loop step — Plan/Execute/Evaluate/Fix]