Context Engineering for Reeve Desk

Purpose: Guide decision-making when adding or changing Reeve's core functionality to maintain optimal signal-to-noise ratio across all contexts.

Core Principle

Load only what's needed, when it's needed.

From Claude Code best practices: "Bloated CLAUDE.md files cause Claude to ignore your actual instructions!"

Decision Tree: Where Should This Information Live?

1. Ask: "Is this always relevant in every context?"

YES → CLAUDE.md

• •Identity and core mission
• •Operating environment (Desk structure, Pulse system)
• •Core directives (Be proactive, Connect to goals, etc.)
• •Decision-making hierarchy

NO → Continue to question 2

2. Ask: "Is this a workflow I invoke by name?"

YES → .claude/skills/SKILL-NAME/SKILL.md

• •User-invocable actions (/morning-briefing, /git-workflow)
• •Procedural steps I follow
• •Actions with side effects
• •Should have clear description field for when to invoke

NO → Continue to question 3

3. Ask: "Is this detailed guidance for specific tasks?"

YES → .claude/rules/RULE-NAME.md

• •Detailed workflow instructions
• •Best practices for specific domains
• •Communication style guidelines
• •Path-specific or context-specific rules

NO → Continue to question 4

4. Ask: "Is this knowledge about the user or their preferences?"

YES → Knowledge files (Goals/, Responsibilities/, Preferences/, Diary/)

• •User goals and priorities → Goals/
• •Recurring duties and projects → Responsibilities/
• •User preferences and constraints → Preferences/
• •Historical context and logs → Diary/

NO → Consider if this is truly needed

Size Guidelines

Common Patterns

Pattern 1: High-Level → Detailed Reference

CLAUDE.md:

## Your Workflow for Each Pulse

When you wake up:
1. Understand the context (why, when, calendar, goals)
2. Check the Desk (read relevant files)
3. Take action (notify, schedule, update)
4. Set up future wake-ups (aperiodic pulses or Diary)

For detailed workflow steps, see `.claude/rules/pulse-workflow.md`

.claude/rules/pulse-workflow.md:

# Detailed Pulse Workflow

## 1. Understand the Context
- Why did I wake up? (check pulse prompt)
- What time is it? (check system time)
- [... detailed checklist ...]

## 2. Check the Desk
- Read Goals/Goals.md for current priorities
- [... 20 more lines of detailed steps ...]

Pattern 2: Skill for Invocable Workflows

When a workflow:

• •Has clear trigger conditions ("when committing", "every morning")
• •Contains procedural steps
• •Benefits from isolation/invocation

Create a skill:

---
name: git-workflow
description: Handle git commits with proper messages. Use when committing changes to Goals, Responsibilities, Preferences, or Diary.
user-invocable: true
---

## When to Commit
1. After significant updates to Goals/, Responsibilities/, Preferences/
2. Daily (9 PM) - Diary/ entries
[... detailed instructions ...]

Pattern 3: Goals/Responsibilities Stay High-Level

BAD (too detailed):

## Goals/Goals.md (150 lines)

### Fitness Goal: Exercise 3-4x per week
- Monday: Upper body strength training at 10 AM at Equinox gym
  - Warm-up: 10 min elliptical
  - Bench press: 3 sets of 8 reps at 135 lbs
  - [... 30 more lines of workout details ...]

GOOD (high-level with reference):

## Goals/Goals.md (40 lines)

### Fitness Goal: Exercise 3-4x per week
**Status**: In progress (2/4 this week)
**Target**: Build consistent routine, increase energy
**Details**: See `Diary/Workout-Plan.md` for current schedule

### Sleep Goal: 7-8 hours/night
**Status**: Tracking
**Target**: Consistent 10:30 PM bedtime

Red Flags (Signs of Poor Context Engineering)

❌ CLAUDE.md has detailed procedural steps ❌ Skills contain identity/mission information ❌ Goals/ files are 100+ lines ❌ Same information duplicated in multiple places ❌ Adding a new capability requires changing 5+ files

When to Invoke This Skill

Invoke /context-engineering when:

1. •Adding new capabilities: "Should this git workflow go in CLAUDE.md or a skill?"
2. •CLAUDE.md exceeds 300 lines: Time to extract sections
3. •Goals/Responsibilities are bloated: Move details to specialized docs
4. •Making structural changes: Ensure proper organization
5. •After user feedback: "This context is too noisy for task X"

Implementation Workflow

When restructuring:

1. •Identify bloat: What information is rarely needed?
2. •Choose destination: Use decision tree above
3. •Extract and link: Move content, add reference in original location
4. •Test: Run typical workflows to ensure context loads correctly
5. •Commit: git commit -m "Context engineering: Move X to Y for better signal-to-noise"

Skills Can Reference Each Other

If a skill needs to mention another:

# evening-wrapup/SKILL.md

## Final Step: Commit Changes

Use `/git-workflow` to commit today's Diary entries with a proper message.

Note: Skills can't directly invoke other skills, but can mention them for user awareness or suggest running them sequentially.

Using Subagents for Complex Analysis

When restructuring requires:

• •Reading many files
• •Making complex decisions
• •Generating new content

Consider using the Task tool with:

• •subagent_type=Explore - For understanding current structure
• •subagent_type=Plan - For planning large refactors
• •subagent_type=general-purpose - For executing complex restructuring

Example: Recent Git Workflow Change

What happened: Added detailed git commit instructions to CLAUDE.md Problem: Polluted context - those details aren't needed every pulse Solution:

1. •Create .claude/skills/git-workflow/SKILL.md with full details
2. •Trim CLAUDE.md to: "For git best practices, invoke /git-workflow"
3. •Result: Saved ~50 lines in always-loaded context

Maintenance

Monthly review:

• •Check CLAUDE.md line count
• •Look for sections referenced but rarely used
• •Extract to skills or rules as needed
• •Verify Goals/Responsibilities stay high-level

Remember: Every line in CLAUDE.md has a cost. Every line in a skill has value only when needed. Structure accordingly.

Version: 1.0 Last Updated: 2026-01-23