Features Tracking System for AI Coding Agents - Managing 29 Features - Yunjeong Luna Lee

TL;DR

When developing with AI tools like Claude Code, you lose work context. The Features tracking system documents every feature, saves 50x tokens with INDEX.md, and auto-generates blog drafts with the /blog command.

Tracking 29 features, 50x token savings, reduced blog writing barrier.

Background: Limitations of daily_work_summary

Initially, I only kept daily work logs in daily_work_summary/2025-10-15.md format.

# 2025-10-15
- Fixed login functionality
- Fixed search bug
- Started adding caching

Problems emerged days later:

No Feature-Level Tracking
- “How did I fix that search bug?” → digging through multiple date files
- “Where are caching-related commits?” → searching entire Git log
Claude Conversation Thread Management Failed + Parallel Work Chaos

With Claude Code, you need to run multiple features in parallel for efficiency. Can’t just sit and wait for one task to finish.

But this means I end up stopping work on F001, F003, F006 simultaneously at quitting time, and trying to remember multiple contexts overloads even me, a human.

Every morning, I wasted 30 minutes just figuring out “where did I leave off?”
AI Token Waste
- Explaining “what I did last week” → need to read 5 daily_log files
- 33 features → reading all 33 READMEs = 15,000 tokens consumed

→ Introduced Features system for feature-level tracking + token savings

Solution: Features System

Real-World Usage Flow

Scenario 1: When Pausing Work (Need to shut down computer)

(Working on F006 search feature...)

Me: /pause
AI: "Saved F006 progress"
    → Record current state in frontmatter
    → Update INDEX.md

(Computer shutdown, weekend passes...)

Scenario 2: When Resuming Work (Monday morning)

Friday quitting time, I was juggling F001, F003, F006 simultaneously.

Me: /resume
AI: "3 paused tasks:
    - F001 User Auth (90%, only tests remaining)
    - F003 Logging System (50%, configuring AOP)
    - F006 Search Optimization (30%, designing index)"

→ 30 minutes of confusion → 0 seconds ✅

Scenario 3: When Completing Work

Me: /finish
AI: Automatically handles:
    1. Update Features documentation (status: DONE)
    2. Write work log
    3. Git commit
    4. "Is this worth blogging about?" assessment

Result: AI remembers and manages my work context.

Core Idea

“Every feature = One document”

features/
├── INDEX.md                           # Cache for quick lookup
├── F001-classroom-management/         # Feature folders
│   ├── README.md                      # Work tracking
│   ├── blog.md                        # Blog draft (optional)
│   └── images/                        # Screenshots
│       ├── 01-before.png
│       └── 02-after.png
├── F002-controller-refactoring/
└── F029-logging-policy-aop/

1. State Management with Frontmatter

Why not use other approaches?

Many project management systems organize by status folders:

TODO/F001-user-auth.md
IN_PROGRESS/F001-user-auth.md
DONE/F001-user-auth.md

The problem: Moving files breaks Git history:

git log F001-user-auth.md → “Huh? This file was just created?”
Git sees it as “file deletion + new file creation”
GitHub blame doesn’t work properly

Solution: Keep files in place, change only metadata

---
id: F001
title: Classroom Management System
status: DONE                 # TODO → IN_PROGRESS → DONE
priority: HIGH
created: 2025-10-18 KST
completed: 2025-10-21 KST
elapsed_hours: 16
labels: [feature, backend, frontend]
---

# Classroom Management System

## ✅ Completed Tasks
- [x] Classroom CRUD API
- [x] Seat layout UI
- [x] Row/column width adjustment

## 🔗 Related Commits
- `8356a91` - feat: add classroom management feature

Benefits:

git log features/F001-classroom-management/README.md tracks the entire process
GitHub blame works normally
Only status changes, no file movement

2. Token Savings with INDEX.md

Reading 50 feature files wastes tens of thousands of tokens. Instead, read one summary.

---
last_updated: 2025-11-01 20:00 KST
auto_generated: true
---

## 📊 Summary
- **TODO**: 8
- **IN_PROGRESS**: 2
- **DONE**: 21

## 🚀 IN_PROGRESS

### [F006](F006-teacher-schedule-management/) Teacher Weekly Schedule
- **Started**: 2025-10-30 05:00 KST
- **Progress**: 30%

## 📋 TODO

### [F023](F023-current-user-annotation/) @CurrentUser Annotation
- **Estimated**: 2-3 hours
- **Topic**: Automate controller auth checks

Impact:

Reading INDEX.md: ~300 tokens
Reading all files: ~15,000 tokens
50x savings!

3. Automated Blog Drafting

Feature complete → /blog command → draft + image recommendations auto-generated.

features/F001-classroom-management/
├── README.md
└── blog.md          # Auto-generated

blog.md example:

---
title: "Implementing Drag-and-Drop Seat Layout in React"
date: 2025-10-21
tags: [React, DnD, TypeScript]
draft: true
---

## Background
We needed a seat layout feature in the academy management system...

## Implementation
Used HTML5 Drag & Drop API directly instead of react-dnd...

+ Image recommendations:

### 📸 Recommended Screenshots

1. **Seat dragging screen**
   - Filename: `01-drag-interaction.png`
   - Content: Capture while dragging a seat
   - How: Classroom management > Seat layout mode > Drag

📁 Save location: `features/F001-classroom-management/images/`

⚠️ **Important**: Capture now or you won't be able to later when UI changes!

→ Forces immediate capture after development

4. Full Automation with Slash Commands

`/finish` - 8-Step Automated Process

Check changes
Clean code (remove debug logs)
Update Features system ⭐
   → frontmatter status: DONE
   → completed: [current time]
   → Update INDEX.md
Write work log
Sync requirement.md
Git commit
Record commit hash in Features file ⭐
Final report

`/blog` - Blog Draft + Images

Analyze recent work
Judge blog-worthiness
User confirmation
Write blog draft
Recommend images/screenshots ⭐
Guide save location

`/pause` & `/resume` - Suspend/Resume Work

/pause
→ Save state of current IN_PROGRESS work
→ Update INDEX.md

(Next day)

/resume
→ Read only INDEX.md to check suspended work (save tokens)
→ When user selects, read that feature file
→ Resume with TodoWrite

Results

Quantitative Impact

29 features tracked (8 TODO, 2 IN_PROGRESS, 21 DONE)
50x token savings (thanks to INDEX.md cache)
4,469 lines organized (api/, refactoring/, analysis/, etc.)
7 TODOs auto-generated (CONTROLLER_IMPROVEMENTS.md → F023-F029)

Qualitative Impact

1. Context Preservation

Start new conversation:
"Continue F006 work"

AI reads INDEX.md → reads F006 file → immediately resumes

2. Perfect Git History Preservation

$ git log features/F001-classroom-management/README.md
→ All 12 commits over 3 days trackable

3. Lower Blog Writing Barrier

Before: "I'll write it later..." → Never writes
After: blog.md + image recommendations → Draft complete immediately

4. Industry Comparison

Tool	Features	Features System
Claude Task Master	MCP-based, 36 tools	✅ Simpler with Markdown
GitHub Issues	Project-level	✅ Feature-level granularity
Notion	Manual management	✅ AI automation
-	-	✅ Blog integration (unique)
-	-	✅ Image timing (innovative)

Lessons Learned

1. “Conversation Context > Git Log”

Problem: When developing multiple features simultaneously

git log --since="1 hour ago"
→ F001, F002, F003, F032 commits mixed
→ AI asks "What were you working on?"

Solution: If in same conversation thread, AI already knows!

User: (Working on F032...)
User: /finish

AI:
1. Check conversation context: "Worked on F032 in this session" ✅
2. Immediately: "Should I complete F032?"

Information Priority:

Features/README.md (highest) - Already structured context
daily_work_summary (supplement) - Details, trial & error
Git diff (code examples) - Actual changes

→ Git log used as fallback only in new threads!

2. “AI is a Teammate, Not a Tool”

Don’t just use AI as a coding assistant, design a workflow AI can follow.

❌ "AI, do this" (explain every time)
✅ "AI, start F023 work" (system auto-loads context)

3. “File Movement is Git’s Enemy”

Moving files todo/ → in-progress/ → done/ when changing state:

Needs git log --follow
Breaks GitHub blame
Hard to track history

→ Solved with Frontmatter metadata!

4. “Tokens are Money”

Reading 50 files every time:

Slow
Expensive
Context fills up

→ 50x savings with INDEX.md cache

5. “Screenshots are About Timing”

Capture immediately after development, not when writing blog.

→ Built image recommendations into /blog command

6. “Standardize Across All Projects”

Established Features system as required standard for all projects, not just CheckUS.

# C:/Users/YJL/.claude/CLAUDE.md (Global settings)

## 📋 Features Tracking System (Standard for All Projects)

### Core Principle
**Every task is tracked as a Feature.**
- ✅ New feature development → Feature
- ✅ Bug fix → Feature
- ✅ Refactoring → Feature
- ✅ Documentation → Feature
- ❌ "Work not tracked in Features" doesn't exist

Auto-setup prompt:

# When running /finish in a new project

AI:
1. Check for docs/features/ folder
2. If not found:

┌────────────────────────────────────────┐
│ ⚠️  No Features tracking system.       │
│                                        │
│ Set it up now?                         │
│                                        │
│ Will create:                           │
│ - docs/features/INDEX.md               │
│ - docs/features/F001-[current work]/   │
│                                        │
│ [Y/n]                                  │
└────────────────────────────────────────┘

Considerations

1. INDEX.md is a “Cache”

Individual files are source of truth
INDEX.md can be regenerated if corrupted
Auto-update is essential

Problem: Creating new Feature for every small bug fix?

F032: Implement /blog command
F033: Fix /blog web search bug
F034: Fix /blog encoding error
F035: Improve /blog tone

→ Too many Features!

Solution: Auto-search related Features when running /finish

User: "Fixed /blog web search error. /finish"

AI:
1. Extract keywords from conversation context: "blog", "web search"
2. Search INDEX.md → Find F032
3. Offer user choice:

┌─────────────────────────────────────────┐
│ Looks related to F032.                  │
│                                         │
│ A) Continue F032                        │
│    - status: DONE → IN_PROGRESS         │
│    - Add "🐛 Additional Changes (2nd)"  │
│                                         │
│ B) Create new Feature (F033)            │
│                                         │
│ Choice: [A/B]                           │
└─────────────────────────────────────────┘

If A selected: Resume Feature

---
id: F032
status: IN_PROGRESS  # DONE → IN_PROGRESS
resumed: 2025-10-31 12:00 KST
completed: 2025-10-31 10:30 KST  # Keep first completion time
---

## 🐛 Additional Changes (2nd - 2025-10-31)

### Background
Found web search encoding bug while working on F032...

### Changes
- Specify UTF-8 encoding
- Improve search result parsing logic

### Related Commits
- `abc1234` - fix: web search encoding error

→ Related work consolidated in one place!

Conclusion

The Features tracking system:

Context Preservation - AI doesn’t lose work context
Token Savings - 50x efficiency with INDEX.md
Blog Automation - Auto-generate draft + image recommendations
Git History Preservation - State management with Frontmatter
Global Standardization - Applied to all projects

Ultimately managed 29 features systematically and lowered blog writing barrier.

If you use AI coding tools, build a workflow AI can follow first.

Next in series: /blog Command - How I Reduced Blog Writing Time by 83%

Features Tracking System for AI Coding Agents - Managing 29 Features