How Auto-Documentation in Vibe Coding Reduced Bug Fix Time to 5 Minutes

TL;DR

Fixed a Discord attendance system bug in 5 minutes. The secret? Auto-documentation system. When working with AI, documentation is “a gift to your future self.” One Phase checkbox made the difference between 5 minutes and 5 hours.

Background: Vibe Coding and Auto-Documentation

I’ve been vibe coding for about 3 months now. Building CheckUS, an academy management system, with Claude Code.

Initially, I just focused on cranking out code quickly. But as the project grew, a problem emerged. AI can’t remember what it built before. When context is lost, you have to explain everything from scratch again.

So I introduced the Features tracking system.

checkus-docs/features/
├── INDEX.md                    # Full feature index
├── F052-weekly-schedule-attendance/
│   └── README.md               # Discord attendance system redesign docs
├── F072-workflow-commands/
│   └── README.md               # Workflow commands docs
└── ...

The key is the /finish command. Every time I complete a feature, Claude automatically updates the documentation:

What changed (code changes)
Why we decided this way (technical decisions)
What problems it solved

Honestly, I started this out of laziness, but it proved its worth today.

The Problem: “Not Connected” Despite Being on Discord?

A weird bug occurred during operations today.

“Student Lim Ji-hwan (7th grade) is in the Discord voice channel, but the study monitoring screen shows ‘Not Connected’”

At first, I thought it was a Discord bot issue. Checked the server logs.

WARN - Cannot find Discord user. Nickname: 중1 임지환

Discord nickname matching problem? But when I tested it myself, the bot worked fine. DMs came through too.

Checked the API response:

{
  "assigneeId": 40,
  "assigneeName": "임지환",
  "status": "ABSENT",
  "connectedActualStudyTimes": []  // Empty!
}

Direct DB check:

SELECT * FROM online_study_session WHERE student_id = 40;
-- id: 1078, student_id: 40, end_time: NULL (In progress!)

Session exists in DB but not in API response. Something’s off.

This is where documentation showed its power. I told Claude:

“This looks like the Discord attendance system migration that got stopped halfway. There should be related documentation. Go find it.”

Claude searched through the Features documents and found F052:

“Found F052 - Discord Attendance/Notification System Complete Redesign document. Checking… Phase 4 is incomplete.”

Claude read the document and discovered the problem:

## 📋 Overview
Complete redesign to fundamentally solve the complexity of the current Discord attendance check and notification system.

### Phase 4: Monitoring Service Migration
- [ ] StudyMonitoringService implementation
- [ ] API endpoint migration

Phase 4 was incomplete.

Claude analyzed the documentation further:

Old: Uses ActualStudyTime table
New: Uses OnlineStudySession table
Discord bot already writes data to OnlineStudySession
But monitoring service still queries ActualStudyTime

// The problematic code (before fix)
List<ActualStudyTime> unassigned = actualStudyTimeRepository
    .findByAssigneeIdsAndDateRange(...);  // Querying old table!

Fix: Done in 5 Minutes

Once I knew the cause, the fix was simple.

1. Add batch query methods to Repository

// Query sessions linked to WeeklySchedule
List<OnlineStudySession> findByWeeklyScheduleIdsAndPeriod(
    List<Long> weeklyScheduleIds,
    LocalDateTime startTime,
    LocalDateTime endTime);

// Query unscheduled sessions (weeklyScheduleId is NULL)
List<OnlineStudySession> findByStudentIdsAndPeriodUnscheduled(
    List<Long> studentIds,
    LocalDateTime startTime,
    LocalDateTime endTime);

2. Replace Repository in Service

// Before
private final ActualStudyTimeRepository actualStudyTimeRepository;

// After
private final OnlineStudySessionRepository onlineStudySessionRepository;

3. Update query logic

// Before
List<ActualStudyTime> unassigned = actualStudyTimeRepository
    .findByAssigneeIdsAndDateRangeAndAssignedStudyTimeIdIsNull(...);

// After
List<OnlineStudySession> unassignedSessions = onlineStudySessionRepository
    .findByStudentIdsAndPeriodUnscheduled(batch, startTime, endTime);

Commit and deploy. From problem discovery to resolution: 5 minutes.

Lessons Learned: Documentation is a Gift to Your Future Self

1. Documentation is Essential When Working with AI

AI can’t remember. When context is lost, you have to explain everything from scratch. But with documentation? Just say “There should be related docs, go find them”. Even if I don’t remember what I did or what number it was, the AI will find it and even identify the problem on its own.

2. Tracking Incomplete Status is Key

Phase 4 remained unchecked. Without this, I would have assumed “it must be complete” and searched in the wrong places.

### Phase 4: Monitoring Service Migration
- [ ] StudyMonitoringService implementation  ← This one checkbox made the difference between 5 minutes and 5 hours

3. Migration Isn’t “All or Nothing”

We split the ActualStudyTime → OnlineStudySession migration into 4 phases. Thanks to this:

The system kept running even in partial completion state
We could track how far we’d gotten

Conclusion

“Fixed it right away because I documented it”

That’s what I muttered to myself after fixing the bug today. Thanks to my past self 3 months ago who documented despite being lazy, my present self solved the problem in 5 minutes.

The core of vibe coding isn’t AI writing code quickly. It’s leaving documentation so AI can maintain context. That’s how you can answer “Why did we build it this way?” a month or a year later.

Reference: Our Project’s Documentation Workflow

Start new feature: Create features/F{number}-{name}/README.md
During work: Manage progress with phase checkboxes
Feature complete: Auto-commit + doc update with /finish command
On pause: Record current state with /pause command

If you’re curious about this pattern, check out the /blog command post.

TL;DR: If you automate documentation when vibe coding, you’ll fix bugs super fast later. For real.