Agentic Workflow

Phase 1 - Setup (zero or near-zero tokens)
  - CLAUDE.md loaded automatically (cached on first message)
  - Session scope defined (one task, one area)
  - .claudeignore excludes noise directories

Phase 2 - Execution (targeted token use)
  - Agent uses Grep + targeted Read (not broad cat)
  - Tool calls return minimal, relevant results
  - You provide trimmed error messages, not full logs
  - No irrelevant topic switches in this session

Phase 3 - Handoff (minimal output)
  - Ask for diff not full rewrite
  - Review change, confirm correctness
  - Start new session for next task

# Template for a focused first message:
"Task: [one sentence]
File: [path, lines if known]
Error: [exact error or failing test name]
Constraint: [e.g. must not break existing tests]
Output: [diff / new function / explanation]"

# Example:
"Task: Fix empty-string validation bypass in user email field.
File: app/models/UserModel.php, create() method ~line 45
Error: UserTest::testCreateWithEmptyEmail passes when it should fail
Constraint: do not change the method signature
Output: diff of the changed validation lines only"

{
    "hooks": {
        "PreToolUse": [
            {
                "matcher": "Bash",
                "hooks": [
                    {
                        "type": "command",
                        "command": "echo 'REMINDER: pipe long output to | tail -30'"
                    }
                ]
            }
        ],
        "PostToolUse": [
            {
                "matcher": "Edit",
                "hooks": [
                    {
                        "type": "command",
                        "command": "php -l \"$CLAUDE_TOOL_OUTPUT_FILE\" 2>&1 | tail -5"
                    }
                ]
            }
        ]
    }
}

# Tell Claude Code to remember:
"Remember: always show diffs not full file rewrites"
"Remember: use Grep tool not bash grep"
"Remember: pipe all bash output to | tail -30"
"Remember: I prefer 3-column layouts in all PHP lesson files"
"Remember: no em dash in content, always hyphen"

# These get stored in ~/.claude/projects/[path]/memory/
# and are loaded as context at session start
# Cost: ~500-1000 tokens added to system prompt (cheap vs. repeated corrections)

Morning planning (no agent, zero tokens):
  1. List today's tasks in a text file
  2. Group by feature area (auth, billing, frontend)
  3. Order by dependency (models before controllers)

Per-task session:
  1. Open new Claude Code session
  2. State the task clearly in first message (use template)
  3. Let agent work - guide tool calls if needed
  4. Review and confirm output
  5. Close session

Batch similar tasks:
  - All model changes -> one session (shared context valid)
  - All test writes   -> one session
  - All CSS fixes     -> one session
  (Don't mix concerns - context stays focused)

End-of-day review (one short session):
  "Show me a git diff summary of today's changes.
   List only files changed, not the full diff."
  -> 1-2 messages, ~500 tokens