Daily Workflow

Daily Workflow

What to expect day-to-day when using ai-coding-ok in your project. How PDCA runs, how to confirm it's working, and when to intervene.

---

What happens automatically

Once ai-coding-ok is installed, every coding task follows this flow:

You say: "Add a search endpoint for products"
                    │
                    ▼
          ┌─────────────────┐
          │ Mode B: Plan    │  ← AI reads 7 files (~30 seconds)
          │ Load context    │
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │ Do: Code + Test │  ← AI writes implementation
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │ Check: Verify   │  ← AI runs tests
          └────────┬────────┘
                   │
                   ▼
          ┌─────────────────┐
          │ Mode C: Act     │  ← AI updates memory
          │ Write to memory │
          └────────┬────────┘
                   │
                   ▼
          Response includes:
          ## Memory Updates
          ✅ task-history.md — TASK-042: Add product search endpoint

---

How to confirm PDCA ran

1. Look for the Memory Updates section

Every AI response should end with:

## Memory Updates

- ✅ task-history.md — [TASK-XXX] summary
- ✅ decisions-log.md — ADR-XXX (if applicable)
- ⬜ project-memory.md — no changes needed

If this section is missing, PDCA did not complete. Remind the AI:

Run the Act phase: update task-history.md with this task's summary.

2. Check task-history.md

After a few tasks, your task-history.md should have entries:

### [TASK-042] Add product search endpoint
- **Date**: 2026-07-14
- **Type**: feat
- **Summary**: Added GET /api/products/search...

3. Watch the terminal (Claude Code)

In Claude Code, the Stop hook will show activity when the session ends. If you see a message about memory updates, it's working.

---

When to manually remind the AI

PDCA is ~95% reliable. The 5% of cases where you might need to nudge:

Situation	What to say
AI finished coding but no Memory Updates section	"Run the Act phase now."
AI skipped reading memory files	"Read the memory files first per AGENTS.md."
AI made an architecture decision without recording it	"Add this decision to decisions-log.md as a new ADR."
Very long session (50+ rounds)	"Check if project-memory.md needs updating."

---

Reading memory files yourself

You can read the memory files anytime to see what the AI sees:

# What the AI reads before each task
cat .github/agent/memory/project-memory.md
cat .github/agent/memory/decisions-log.md
cat .github/agent/memory/task-history.md

Or in Claude Code:

Show me the current project memory state.

---

Maintaining memory quality

Weekly: Quick scan

Once a week, scan the last few entries in task-history.md. Verify:

• Entries are accurate (what was actually done)
• No duplicate entries
• Format is consistent

Monthly: Deep review

Once a month:

1. Review project-memory.md — any stale facts? Remove them.
2. Review decisions-log.md — any decisions that should be deprecated?
3. Archive old task-history.md entries (keep last 30):

   # Move entries 31+ to archive
   head -n 30 task-history.md > task-history.md.new
   # Append old entries to docs/task-history-archive-2026-Q3.md

On architecture changes

When you make a significant architecture change:

1. Write the ADR yourself — don't just rely on the AI. You understand the trade-offs better.
2. Update project-memory.md — if modules, tech stack, or constraints changed.
3. Verify the AI's entry — check that the AI's task-history entry matches your understanding.

---

Working across multiple sessions

Session 1: Start a feature

You: "Add a search endpoint for products"
AI: [Plan → Do → Check → Act]
    → TASK-042 recorded

Session 2: Continue the feature

You: "Continue TASK-042 — add pagination to the search endpoint"
AI: [Plan: reads task-history, sees TASK-042]
    → Knows search endpoint was added, what files were changed
    → Continues from where you left off

Session 3: Fix a related bug

You: "Search returns empty results for Chinese characters"
AI: [Plan: reads project-memory, sees SQLite FTS5]
    → Reads decisions-log, sees ADR-003 about FTS5
    → Knows to check tokenizer configuration
    → Fix + Act: TASK-044 recorded

The key: No session starts from zero. Every session inherits the full context from all previous sessions.

---

CI integration

The memory-check.yml workflow runs on PRs and checks:

• task-history.md was updated (new entry exists)
• No unfilled placeholders
• Version markers are consistent

If it fails, the PR gets a comment reminding you to update memory.

See CI Integration for setup details.

---

Common workflow patterns

Pattern 1: Quick fix

You: "Fix the typo in the README"
AI: Plan (reads memory) → Do (fixes typo) → Check (n/a) → Act (TASK-XXX: Fix typo)

Even trivial changes get a task-history entry. This keeps the record complete.

Pattern 2: Multi-step feature

Session 1: "Add user authentication" → TASK-050
Session 2: "Add login page" → TASK-051
Session 3: "Add password reset" → TASK-052

Each step is a separate task with its own memory entry. The AI in Session 3 knows about Sessions 1 and 2.

Pattern 3: Research then implement

1. "Research the best search library for our use case"
   → AI reads memory, researches, records findings
   → Act: ADR-005 in decisions-log.md

2. "Implement the search using the library from ADR-005"
   → AI reads memory (including ADR-005), implements
   → Act: TASK-060 in task-history.md

---

Next steps

• Customization Guide — tailor ai-coding-ok to your project
• Upgrade ai-coding-ok — keep the framework up to date
• Team Best Practices — working with ai-coding-ok in a team