Shared Prompt Sections

Common reusable sections. Each agent references these via cache_control markers.

---

§ 1. Output Language Rule

Priority: PLAN.md > Language: → CLAUDE.md ## Language → en (default)

On dispatch: pass resolved language code in <context><language> field
Section headers (##) are also written in the resolved language (refer to language mapping table)

---

§ 2. Build and Lint Commands

# Auto-detect Build (execute only if script exists)
if [ -f "package.json" ]; then
  if node -e "const p=JSON.parse(require('fs').readFileSync('package.json','utf8')); process.exit(p.scripts&&p.scripts.build?0:1)" 2>/dev/null; then
    npm run build 2>&1 || bun run build 2>&1 || yarn build 2>&1
  fi
elif [ -f "Cargo.toml" ]; then
  cargo build 2>&1
elif [ -f "go.mod" ]; then
  go build ./... 2>&1
elif [ -f "pyproject.toml" ] || [ -f "setup.py" ]; then
  python -m py_compile $(find . -maxdepth 3 -name "*.py" -not -path "*/venv/*" 2>/dev/null) 2>&1
elif [ -f "Makefile" ]; then
  make build 2>&1 || make 2>&1
fi

# Auto-detect Lint (execute only if script exists)
if [ -f "package.json" ]; then
  if node -e "const p=JSON.parse(require('fs').readFileSync('package.json','utf8')); process.exit(p.scripts&&p.scripts.lint?0:1)" 2>/dev/null; then
    npm run lint 2>&1 || bun run lint 2>&1 || true
  fi
elif [ -f "pyproject.toml" ]; then
  ruff check . 2>&1 || python -m flake8 . 2>&1 || true
fi

• If build/lint scripts do not exist → skip (treat as N/A).
• On build/lint failure, always fix before reporting.

---

§ 3. WORK and TASK File Path Patterns

works/{WORK_ID}/
  ├─ Requirement.md                 # Created by Specifier (mandatory)
  ├─ PLAN.md
  ├─ PROGRESS.md
  ├─ TASK-00.md               # No WORK prefix
  ├─ TASK-00_progress.md      # Separator: underscore
  ├─ TASK-00_result.md        # Separator: underscore
  └─ TASK-01.md ...

• WORK ID: WORK-NN (e.g., WORK-03)
• TASK ID: TASK-NN (e.g., TASK-00) — WORK prefix must NOT be included

---

§ 4. File System Discovery Scripts

# Find latest WORK with incomplete TASKs
# Use Glob tool: pattern "works/WORK-*/" → list all WORK directories (sorted)
# For each WORK (descending), compare:
#   Glob "works/WORK-NN/TASK-*.md" (exclude *_result.md, *_progress.md) → TOTAL
#   Glob "works/WORK-NN/TASK-*_result.md" → DONE
# First WORK where DONE < TOTAL is the active WORK

# List all WORKs
# Use Glob tool: pattern "works/WORK-*/"

# TASK completion status
# TOTAL = count of Glob "works/${WORK_ID}/TASK-??.md"
# DONE  = count of Glob "works/${WORK_ID}/TASK-*_result.md"

---

§ 5. Task Result XML Format

<task-result work="{WORK_ID}" task="{TASK_ID}" agent="{agent}" status="{PASS|FAIL}">
  <summary>{1-2 line summary}</summary>
  <files-changed>
    <file action="{created|modified|deleted}" path="{path}">{description}</file>
  </files-changed>
  <verification>
    <check name="{type}" status="{PASS|FAIL|N/A}">{details}</check>
  </verification>
  <notes>{notes for next steps}</notes>
</task-result>

---

§ 7. PLAN.md Required Meta-Information — 7 Fields

→ {REFERENCES_DIR}/file-content-schema.md § 1 reference

Field	Required	Description
> Created:	✅	YYYY-MM-DD
> Requirement:	✅	REQ-XXX or user request text
> Execution-Mode:	✅	direct / pipeline / full
> Project:	✅	Project name
> Tech Stack:	✅	Detected tech stack
> Language:	✅	Language code (ko, en, etc.)
> Status:	✅	Always starts as PLANNED

---

§ 8. WORK-LIST.md Update Rules

File: works/WORK-LIST.md

Format:

LAST_WORK_ID: WORK-XX

| WORK | 제목 | 상태 | 생성일 | 완료일 |
|------|------|------|--------|--------|
| WORK-NN | ... | IN_PROGRESS | YYYY-MM-DD | |
| WORK-MM | ... | DONE | YYYY-MM-DD | YYYY-MM-DD |

Status	Meaning	Trigger
IN_PROGRESS	WORK is being executed	specifier creates WORK
DONE	All TASKs completed, awaiting review/push	committer completes last TASK
COMPLETED	Archived to _COMPLETED/	push merge (Main Claude batch processes all DONE)

Rules:

• LAST_WORK_ID header tracks the highest WORK ID ever created
• specifier: on WORK creation, add IN_PROGRESS row + update LAST_WORK_ID
• committer: when last TASK is completed, change IN_PROGRESS → DONE and fill completion date (do NOT move folder or remove row)
• Main Claude (push procedure): move all DONE WORKs to works/_COMPLETED/, remove their rows from WORK-LIST.md

---

§ 9. Locale Detection

1. CLAUDE.md → check "Language: xx"
2. If not found, ask user for language
3. If not found, auto-detect system locale
   - Windows: powershell -c "[CultureInfo]::CurrentCulture.TwoLetterISOLanguageName"
   - Linux/Mac: locale | grep LANG | grep -oP '[a-z]{2}' | head -1
   - Fallback: "en"

---

§ 10. Callback Transmission Template

→ Bash command rules: see § 13 — each step below is a separate tool call.

Replace {CallbackType} with the actual key name (e.g., ProgressCallback, TaskCallback).

Step 1. Use Grep tool to find {CallbackType}: line in CLAUDE.md. If not found, skip callback entirely.

Step 2. Use Grep tool to find CallbackToken: line in CLAUDE.md (optional).

Step 3. Send callback with a single curl command:

curl -s -X POST "CALLBACK_URL" -H "Content-Type: application/json" -H "X-Runner-Api-Key: TOKEN" -d '{"workId":"WORK-01","taskId":"TASK-00",...}'

Agent-specific payload fields:

• ProgressCallback (builder): "status": "IN_PROGRESS", "currentReasoning": "..."
• TaskCallback (committer): "status": "SUCCESS", "commitHash": "${COMMIT_HASH}"

---

§ 11. Project Discovery

# 1. Check CLAUDE.md language setting
grep -oP '(?<=Language:\s?)[a-z]{2}' CLAUDE.md 2>/dev/null

# 2. Tech stack
head -50 package.json 2>/dev/null
head -30 pyproject.toml 2>/dev/null
head -20 Cargo.toml 2>/dev/null
head -10 go.mod 2>/dev/null

# 3. Structure (when needed)
find . -maxdepth 3 -type f \( -name "*.md" -o -name "*.json" -o -name "*.toml" \) -not -path "*/node_modules/*" 2>/dev/null

---

§ 12. Progress File Gate Check

Gate conditions for works/WORK-NN/TASK-XX_progress.md:

• File exists at the expected path
• Status: COMPLETED line is present
• ## Files Changed section is present and non-empty

On gate failure → return FAIL task-result immediately. Do not proceed to subsequent steps.

---

§ 13. Bash Command Rules

Bash commands MUST follow these rules for permission compatibility.

MANDATORY:

• One simple command per Bash call — NO compound commands (&&, ||, ;, |)
• NO cd dir && command — you are already in the project root
• NO multi-line scripts — split into separate Bash calls
• NO sub-shell expansions in arguments — e.g., $(date ...) inside printf
• Use relative paths from project root (e.g., works/WORK-01/) — NO absolute paths
• Use git add file, git commit -m "msg" — NO git -C path flag

For file operations, prefer dedicated tools over Bash:

• Read files → Read tool (NOT cat)
• Write/append files → Write tool (NOT echo >> or printf >>)
• Edit files → Edit tool (NOT sed -i)
• Search files → Grep tool (NOT grep)
• Find files → Glob tool (NOT find)

Activity log example:

WRONG: printf '[%s]_%s\n' "$(date ...)" "INIT" >> work.log
RIGHT: Use Write tool to append a line to the log file

Git example:

WRONG: cd /path/to/project && git add file && git commit -m "msg"
RIGHT: git add file        (one call)
       git commit -m "msg"  (next call)

---

Version

• Created: 2026-03-10
• Updated: 2026-03-28