claudeos-core 2.1.1 → 2.3.0

package/pass-prompts/templates/common/claude-md-scaffold.md

@@ -0,0 +1,686 @@
+# CLAUDE.md Scaffold Template (v2.2.0+)
+
+This scaffold enforces structural determinism across all generated CLAUDE.md files.
+CONTENT adapts to each project based on pass2-merged.json.
+STRUCTURE is FIXED and MUST NOT be reordered, renamed, merged, or extended.
+
+## Why this scaffold exists
+
+Prior to v2.2.0, CLAUDE.md generation specified required CONTENT items but did not
+fix the structure. LLM choices varied between runs and between projects, producing
+inconsistent section counts, names, and ordering. This problem only became visible
+when claudeos-core was applied to multiple projects in the same organization.
+
+This scaffold makes CLAUDE.md generation deterministic:
+- Section count: EXACTLY 8
+- Section titles: FIXED (see below)
+- Section order: FIXED
+- Section formats: SPECIFIED per section
+
+## Hard constraints
+
+**EXACTLY 8 SECTIONS. No more, no less.**
+
+Before finalizing CLAUDE.md, count `^## ` lines. The count MUST be 8.
+If the count is 9 or more, merge surplus content into existing sections:
+- Memory-related surplus → Section 8 sub-section 2 (L4 Memory)
+- Common rules (51·52, doc-writing, ai-work) → Section 8 sub-section 1 (Common Rules)
+- Rules-reference (directory index) → Section 6 (Standard / Rules / Skills Reference)
+- Rule enumeration beyond those → MOVE to `.claude/rules/*` or `standard/*`, not CLAUDE.md
+
+DO NOT:
+- Add new top-level (`##`) sections beyond the 8 defined
+- Remove any of the 8 sections (use minimal content if empty)
+- Rename sections
+- Change section order
+- Merge the 8 sections with each other
+- Create a `##`-level section whose name (or its translation in the
+  target output language) has the semantic meaning of "Common Rules",
+  "Required to Observe While Working", "Rules Summary", "Documentation
+  Writing Rules", "AI Common Rules", "L4 Memory Integration Rules",
+  or "Memory Layer" (as a standalone section), or any variation that
+  collects rules, procedures, or memory content. "Common Rules" and
+  "L4 Memory" each belong as a SUB-SECTION (`###`) inside Section 8,
+  not as their own top-level `##` section.
+  All other enforcement rule BODIES belong in `.claude/rules/*`
+  (auto-loaded via paths glob) or `claudeos-core/standard/*`.
+- Restate the L4 memory table or workflow twice under different headings
+  — it appears exactly once, only in Section 8 sub-section 2.
+
+## Section heading format (MANDATORY)
+
+The 8 canonical section headings below (`## 1. Role Definition` through
+`## 8. Common Rules & Memory (L4)`) use English as their canonical form.
+When generating CLAUDE.md in a non-English output language, the English
+canonical heading MUST remain the primary heading text. A native-language
+translation MAY be appended in parentheses, but is optional.
+
+This rule exists because multiple projects in the same organization will
+have their CLAUDE.md files consumed together (multi-repo grep, cross-repo
+navigation, side-by-side review). When every project uses a different
+native-language translation for the same section, even when the structure
+is otherwise identical, discoverability breaks: `grep "## 7. DO NOT Read"`
+no longer matches all siblings.
+
+Format rule:
+- Primary (required): English canonical heading exactly as listed.
+- Parenthetical (optional): native-language translation, added at the end.
+
+Examples (ko output):
+
+    ✅ `## 7. DO NOT Read`
+    ✅ `## 7. DO NOT Read (직접 읽지 말아야 할 파일)`
+    ❌ `## 7. 읽지 말 것 (Files Not to Be Read Directly)`
+        — English must be primary, not parenthetical
+    ❌ `## 7. 읽지 말 것`
+        — English canonical must appear
+
+Examples (ja output):
+
+    ✅ `## 7. DO NOT Read (直接読まないファイル)`
+    ❌ `## 7. 直接読まないファイル (DO NOT Read)`
+
+The same rule applies to all 8 sections. When in doubt, emit only the
+English canonical heading; the parenthetical translation is a courtesy,
+not a requirement. The CONTENT below the heading is still written
+entirely in the target language.
+
+DO:
+- Adapt content within each section to project facts from pass2-merged.json
+- Use placeholder values (`{PROJECT_CONTEXT}`, `{OUTPUT_LANG}`) where specified
+- Follow per-section format rules
+- When unsure where information belongs, consult the "Per-section generation rules"
+  below. Anything not fitting the 8 sections belongs OUTSIDE CLAUDE.md.
+
+## Template structure
+
+Use the following structure EXACTLY:
+
+```markdown
+# CLAUDE.md — {PROJECT_NAME}
+
+> {1-2 line project intro}
+
+## 1. Role Definition
+
+As the senior developer for this repository, you are responsible for writing, modifying, and reviewing code. Responses must be written in {OUTPUT_LANG}.
+{PROJECT_CONTEXT}.
+
+## 2. Project Overview
+
+| Item | Value |
+|---|---|
+| Language | {language with version} |
+| Framework | {framework with version} |
+| ... (8-12 rows total) |
+
+## 3. Build & Run Commands
+
+{OPTIONAL: Top warning - [CRITICAL] for env requirements, or bold text for package manager rules}
+
+```bash
+# {install command with comment}
+{install command}
+
+# {dev/build/test commands with inline comments}
+{command 1}          # {purpose / port / build engine}
+{command 2}
+...
+```
+
+{OPTIONAL: Bottom reference to single source of truth}
+
+## 4. Core Architecture
+
+### Overall Structure
+
+{ASCII diagram in code block}
+
+### Data Flow
+
+{Arrow chain or numbered list of data flow steps}
+
+### Core Patterns
+
+- **{pattern name}**: {explanation}
+- **{pattern name}**: {explanation}
+- ... (3-5 bullets)
+
+### Absent / Not Adopted (OPTIONAL — include only if project deliberately avoids categories)
+
+- {Generic category} ({optional parenthetical examples})
+- ...
+
+## 5. Directory Structure
+
+```
+{project root or src/}/
+├─ {dir}/       # {role}
+├─ {dir}/       # {role}
+│  └─ {subdir}/ # {role}
+└─ ...
+```
+
+**{emphasis 1}**: {explanation}
+**{emphasis 2}**: {explanation}
+(Maximum 3 emphasis bullets below the tree)
+
+## 6. Standard / Rules / Skills Reference
+
+### Standard (Single Source of Truth)
+
+| Path | Description |
+|---|---|
+| `claudeos-core/standard/{path}` | {1-line description} |
+| ... (list ALL generated standard files individually) |
+
+### Rules (Auto-loaded Guardrails)
+
+| Path | Description |
+|---|---|
+| `.claude/rules/00.core/00.standard-reference.md` | Standard documentation index |
+| `.claude/rules/{category}/*` | {what's enforced} |
+| ... (use wildcards per category) |
+
+### Skills (Automated Repeated-task Procedures)
+
+- `claudeos-core/skills/00.shared/MANIFEST.md` — Skill registry
+- `claudeos-core/skills/{path}` — {purpose}
+
+## 7. DO NOT Read (Files Not to Be Read Directly)
+
+| Path | Reason |
+|---|---|
+| `claudeos-core/guide/` | Human-oriented documentation. Key content already captured in rules files. |
+| `claudeos-core/generated/` | Auto-generated metadata. Internal use only. |
+| `claudeos-core/mcp-guide/` | Key content already captured in rules files. Internal use only. |
+| {conventional build output} | {1-line reason} |
+
+## 8. Common Rules & Memory (L4)
+
+### Common Rules (auto-loaded on every edit)
+
+Global guardrails applied automatically to every documentation and code edit. Loaded via `paths: ["**/*"]`.
+
+| Rule File | Role | Core Enforcement |
+|---|---|---|
+| `.claude/rules/00.core/51.doc-writing-rules.md` | Documentation writing rules | {1-line meta summary} |
+| `.claude/rules/00.core/52.ai-work-rules.md` | AI work rules | {1-line meta summary} |
+{| additional paths-universal rules if any |}
+
+For detailed guidance, Read `claudeos-core/standard/00.core/04.doc-writing-guide.md`.
+
+### L4 Memory (on-demand reference)
+
+Long-term context (decisions · failures · compaction · auto-proposals) is stored in `claudeos-core/memory/`.
+Unlike rules that auto-load via `paths` glob, this layer is referenced **on-demand**.
+
+#### L4 Memory Files
+
+| File | Purpose | Behavior |
+|---|---|---|
+| `claudeos-core/memory/decision-log.md` | Permanent record of the "why" behind design choices | Append-only. Skim at session start; Read before any architectural change. Add a new entry when a decision is made. |
+| `claudeos-core/memory/failure-patterns.md` | Repeated errors and their remedies | Searched at session start and whenever an error occurs. Record only when the pattern repeats (≥2 occurrences), the root cause is non-obvious, and the remedy is undocumented. |
+| `claudeos-core/memory/compaction.md` | Declares the 4-stage compaction policy | Modify only when the policy itself changes. Execute compaction solely via `npx claudeos-core memory compact`. |
+| `claudeos-core/memory/auto-rule-update.md` | Machine-generated rule change proposals | Review → Accept → Edit the rule → Record in `decision-log.md`. Do NOT edit this file directly. Regenerate via `npx claudeos-core memory propose-rules`. |
+
+#### Memory Workflow
+
+1. **Session start**: Scan `failure-patterns.md` to be aware of frequently occurring errors.
+2. **Skim recent decisions**: Skim recent entries in `decision-log.md` to avoid overwriting architectural decisions that have already been agreed upon.
+3. **Record new decisions**: When making a significant design decision (choosing between competing patterns, adopting or rejecting a library, fixing a convention, etc.), append to `decision-log.md`.
+4. **Record repeated errors**: If the same error occurs ≥2 times and the root cause is non-obvious, register it in `failure-patterns.md` with a new pattern-id.
+5. **Periodic compaction**: When a memory file approaches 400 lines or has not been tidied up for over a month, run `npx claudeos-core memory compact`.
+6. **Review rule-update proposals**: Review proposals in `auto-rule-update.md` with confidence ≥ 0.70. When accepting, edit the corresponding rule file and log the decision in `decision-log.md`.
+```
+
+---
+
+## Per-section generation rules
+
+### Section 1: Role Definition
+
+**Structure**: EXACTLY 2 lines.
+
+**Line 1** (fixed text with `{OUTPUT_LANG}` substitution; emit in the target output language):
+```
+As the senior developer for this repository, you are responsible for writing, modifying, and reviewing code. Responses must be written in {OUTPUT_LANG}.
+```
+
+**Line 2** (`{PROJECT_CONTEXT}` — dynamically generated, Level 2 abstraction):
+
+Generate `{PROJECT_CONTEXT}` as a single sentence combining:
+- Stack name (generic form): `Java Spring Boot`, `React + Vite`, `Python Django`, `Node.js Express`, `Next.js App Router`
+- Project type: "Backend Application", "Back Office SPA", "Front Office SPA", "REST API Server", "Full-stack Web Application"
+- 1-2 architectural differentiators (using generic category terms)
+
+**INCLUDE** (Level 2 — generic terms):
+- Stack labels: "Java Spring Boot", "React + Vite"
+- Architecture pattern names: "Layered Architecture", "Multi-entry", "Dual Build"
+- Generic category terms: "Dual RDBMS", "XML-based SQL Mapper", "3-tier Authentication"
+
+**EXCLUDE** (no hardcoding):
+- Specific component names (PrimaryButton, ApiClient, AuthProvider)
+- Specific product names (PostgreSQL, MariaDB, Redis, MyBatis, Prisma)
+- Specific file paths (apiClient.ts, vite.config.desktop.ts)
+- Specific counts (18 domains, 3 layers)
+- Specific versions (React 19.1.0, Spring Boot 3.2)
+- Specific domain names (admin, guide, orders)
+
+**Source**: pass2-merged.json architecture/stack sections.
+
+### Section 2: Project Overview
+
+**Format**: 2-column table with 8-12 rows.
+
+**Required rows** (adapt per stack; column labels and row titles are emitted in the target output language):
+- Language (with version)
+- Framework (with version; use "Not X" for disambiguation if needed)
+- Build Tool or Package Manager
+- Main library (UI library for frontend, ORM/SQL mapper for backend)
+- Execution environment (server port / dev server port)
+- External dependencies (backend server, DB, session store)
+
+**Optional rows** (include if distinctive):
+- SPA Entry (multi-entry projects)
+- Build Structure (dual builds, special configurations)
+- Test Status (if notable — e.g., "0% coverage")
+- Development Status Caveat (in-progress migrations, MOCK data usage)
+- Forbidden Package Manager (brief form, e.g., "npm (pnpm, yarn forbidden)")
+
+**Environment variables** (REQUIRED when `stack.envInfo` exists in pass2-merged.json):
+
+When the project declares environment variables (via `.env.example` or
+`.env`), the table MUST reflect what the project actually declares, not
+framework defaults. Specifically:
+
+- **Port row**: use `stack.envInfo.port` or `stack.port` (both already
+  resolved from the project's env file). Annotate with the env var name
+  in parentheses. Example (emit row labels in target output language):
+  `| Dev Server Port | 3000 (\`VITE_DESKTOP_PORT\`) |`
+  NOT `| Dev Server Port | 5173 |` when the project declares 3000.
+- **Host row** (if `stack.envInfo.host` exists): add a row showing the
+  declared dev host. Example:
+  `| Dev Server Host | \`localhost\` (\`VITE_DEV_HOST\`) |`
+- **API target row** (if `stack.envInfo.apiTarget` exists): add a row
+  showing the backend proxy/API target. Example:
+  `| API Proxy Target | \`http://localhost:8080\` (\`VITE_API_TARGET\`) |`
+- **Additional ports** (if `stack.envInfo.vars` contains multi-port
+  patterns like `*_MOBILE_PORT`, `*_STORYBOOK_PORT`): inline them in the
+  port row using ` · ` as separator, OR add dedicated rows.
+- **Source file** (`stack.envInfo.source`, e.g., `.env.example`): do NOT
+  include the source filename in the table itself — that metadata is for
+  prompt-time attribution, not user-facing CLAUDE.md content. Use it only
+  to verify which file the values came from.
+
+General rule: if the project's env file declares something, the CLAUDE.md
+table must match it. Framework defaults (Vite 5173, Next.js 3000, Django
+8000) are LAST RESORT — only used when env files are silent.
+
+**SECURITY — Do NOT include sensitive values in CLAUDE.md**:
+- Only reference PORT / HOST / API_TARGET variables (the three scaffolded
+  row types above). These are non-sensitive runtime configuration values.
+- DO NOT populate table rows with values from variables named like
+  `*_PASSWORD`, `*_SECRET`, `API_KEY`, `*_TOKEN`, `*_CREDENTIAL`,
+  `PRIVATE_KEY`, `JWT_SECRET`, `CLIENT_SECRET`, etc. These are
+  automatically redacted to `***REDACTED***` by env-parser anyway, but
+  the scaffold explicitly forbids referencing them as a defense-in-depth.
+- If `stack.envInfo.vars` contains a value equal to `***REDACTED***`,
+  that variable was redacted for security. Treat it as non-existent for
+  CLAUDE.md purposes — do NOT write the sentinel string into the table.
+- Rationale: CLAUDE.md is committed to VCS and sometimes published. A
+  misconfigured `.env.example` (containing real secrets by mistake) must
+  not have those secrets amplified into project documentation.
+
+**DO NOT INCLUDE** (belongs in other sections):
+- Response wrapping style → Section 4
+- Exception handling strategy → should go in rules/standard
+- Encryption details → rules/standard
+- Connection pool details → Section 4 or omit
+- DB migration warnings → rules/standard
+- Route guard details → Section 4
+
+**Style**:
+- Prefer specific versions (e.g., "5.8.3") but major-only is acceptable
+- Negation allowed concisely: "Not Next.js", "TanStack Query Not Adopted"
+- Use parentheses for clarifying annotations
+
+**Source**: pass2-merged.json stack/tooling sections, including `stack.envInfo`.
+
+### Section 3: Build & Run Commands
+
+**Structure**: 3 parts.
+
+**Part 1 (OPTIONAL)**: Top warning
+- Use `**[CRITICAL] {warning}**` for environmental requirements (JDK version, runtime)
+- Use **bold text** for package manager restrictions
+- Examples (emit in target output language):
+  - `**[CRITICAL] Java 21 required** — the system default Java may not be 21, so ...`
+  - `This project standardizes on **npm** as its package manager. \`yarn\` / \`pnpm\` commands are forbidden.`
+
+**Part 2 (REQUIRED)**: Single bash code block with 4-7 commands
+- Include dependency install command
+- Include dev/run command
+- Include build command
+- Include test command (if applicable)
+- Include project-specific commands if critical (migration, codegen trigger)
+- Add inline comments for: port number, build engine, purpose, current state warnings
+- **Port comment accuracy**: When adding a port comment next to a dev/run
+  command (e.g., `# dev server (port 3000)`), use the port from
+  `stack.envInfo.port` or `stack.port` in pass2-merged.json. Do NOT write
+  the framework's theoretical default (5173 for Vite, 3000 for Next.js,
+  8000 for Django) if the project's env file declares a different value.
+
+**Part 3 (OPTIONAL)**: Bottom reference
+- Point to single source of truth: "Treat package.json's scripts as the single source of truth"
+- Reference skill files for complex procedures
+- Restate critical restrictions briefly
+
+**Source**: pass2-merged.json scripts/buildTool/packageManager sections,
+plus `stack.envInfo` for port accuracy.
+
+### Section 4: Core Architecture
+
+**Structure**: 3-4 sub-sections.
+
+**Sub-section 1 (REQUIRED)**: `### Overall Structure` (emit heading in target output language)
+- ASCII diagram in code block
+- Choose shape based on project type:
+  - Layered backend: vertical chain
+  - Multi-entry frontend: boxes + tree
+  - Single-app frontend: vertical tree
+  - Microservices: boxes with arrows
+- Under 15 lines
+
+**Sub-section 2 (REQUIRED)**: `### Data Flow`
+- Arrow chain (1-2 lines) for simple flows
+- Numbered list (3+ steps) for complex flows with transformations
+
+**Sub-section 3 (REQUIRED)**: `### Core Patterns`
+- 3-5 bullets
+- Each reflects a distinctive architectural decision
+- Use bold prefix: `- **{pattern name}**: {explanation}`
+
+**Sub-section 4 (OPTIONAL)**: `### Absent / Not Adopted`
+- Include when pass2-merged.json reveals deliberate NON-adoption
+- Common candidates: state management libs, server state cache, error tracking, route guards
+- Each item: single line, generic category + parenthetical examples
+
+**DO NOT INCLUDE**:
+- Specific file paths (→ Section 5)
+- Version numbers (→ Section 2)
+- Build/run commands (→ Section 3)
+
+### Section 5: Directory Structure
+
+**Structure**: Tree diagram + up to 3 emphasis bullets below.
+
+**Tree format**:
+- Use UTF-8 box drawing: `├─`, `└─`, `│`
+- Maximum 30-40 lines per tree
+- For backend projects: separate trees for source packages vs resources
+- For frontend/fullstack: single unified tree
+- Include inline comments for key folders only
+- Use placeholders where applicable: `{domain}/`, `{screen}/`, `{Table}`, `{DB}`
+
+**Emphasis bullets (MAX 3)**:
+Priority order:
+1. Auto-generated / do-not-modify paths
+2. Synchronization requirements (multi-file sync, dual-write)
+3. Separation/sharing boundaries (device split, module boundaries)
+
+Format: `**{category}**: {explanation}`
+
+**If 4+ items seem necessary**, move overflow to rules/standard instead of cramming.
+
+### Section 6: Standard / Rules / Skills Reference
+
+**Structure**: EXACTLY 3 sub-sections.
+
+**Sub-section 1**: `### Standard (Single Source of Truth)`
+- Table format: `| Path | Description |` OR `| Area | Path |` (pick one, stay consistent)
+- List EVERY generated standard file individually
+- Description: 1-line
+- Source: list of generated standard files
+
+**Sub-section 2**: `### Rules (Auto-loaded Guardrails)`
+- Table format OR bullet list (choose based on project size)
+- Use WILDCARDS by category: `.claude/rules/{category}/*`
+- Always include: `00.standard-reference.md`, `50.sync/*`, `60.memory/*`
+  - **`60.memory/*` is mandatory** — the directory itself is created by
+    Pass 4 (4 rule files: 01.decision-log, 02.failure-patterns, 03.compaction,
+    04.auto-rule-update), but Section 6's Rules table MUST include a
+    `.claude/rules/60.memory/*` row at Pass 3 generation time as a
+    forward reference. This keeps the table's category index complete
+    and tells Claude Code that L4 memory rules auto-load when editing
+    the corresponding memory files. DO NOT omit it on the grounds that
+    "Pass 4 hasn't run yet" — the row describes the CATEGORY, not the
+    individual files.
+- Description: what's enforced per category (not per file)
+- **IMPORTANT**: `.claude/rules/00.core/*` wildcard row COVERS
+  `51.doc-writing-rules.md`, `52.ai-work-rules.md`, and all other 00.core rules.
+  Do NOT list these individually in Section 6 and do NOT create a separate
+  `##`-level section ("Documentation Writing Rules", "AI Common Rules")
+  for them — the wildcard already references them, and they are
+  auto-loaded via their `paths` frontmatter. The DETAILED universal-rules
+  table (with per-file meta-summary) belongs in Section 8 sub-section 1,
+  not here. Section 6's role is the directory index; Section 8's role
+  is the always-on layer summary. When a reader wants to see "what gets
+  auto-loaded on every edit", they go to Section 8.
+  Example row (emit text in target output language):
+  `.claude/rules/00.core/*` | Project / Architecture / Naming /
+  Documentation / AI-work rules (see §8 Common Rules table for the
+  detailed enforcement of 51·52, which auto-load on every edit) |
+  Example row for 60.memory:
+  `.claude/rules/60.memory/*` | Auto-loaded when editing L4 memory
+  files (decision-log · failure-patterns · compaction · auto-rule-update) |
+
+**Sub-section 3**: `### Skills (Automated Repeated-task Procedures)`
+- Bullet list
+- Include ONLY skills that actually exist
+- Always include: `00.shared/MANIFEST.md`
+- Format: `path — purpose`
+
+**DO NOT INCLUDE** (moves to Section 7):
+- `claudeos-core/guide/` (go to Section 7)
+- `claudeos-core/generated/` (go to Section 7)
+- `claudeos-core/mcp-guide/` (go to Section 7)
+
+### Section 7: DO NOT Read
+
+**Structure**: 2-column table, 3-5 rows total.
+
+**ALWAYS INCLUDE (3 common items)** (emit row text in target output language):
+```
+| claudeos-core/guide/      | Human-oriented documentation. Key content already captured in rules files. |
+| claudeos-core/generated/  | Auto-generated metadata. Internal use only. |
+| claudeos-core/mcp-guide/  | Key content already captured in rules files. Internal use only. |
+```
+
+**INCLUDE conventional build outputs** (detect from pass2-merged.json buildTool):
+- Vite / Webpack / Rollup → `dist/`
+- Next.js → `.next/`
+- Gradle → `build/`
+- Maven → `target/`
+- Python → `__pycache__/`, `dist/`
+- Storybook (if present) → `storybook-static/`
+
+**DO NOT INCLUDE**:
+- Auto-generated source code directories (generated-api-client/, dto/generate/, etc.) — handled in Section 5 emphasis
+- Project-specific build variants (dist-guide/, dist-mobile/) — not conventional
+- Project-specific documentation output paths
+- Environment files (.env*) — handled by rules/40.infra/*
+
+**TARGET ITEM COUNT**: 3-5 items.
+
+### Section 8: Common Rules & Memory (L4)
+
+**Structure**: 2 sub-sections.
+
+**Rationale**: CLAUDE.md's Section 6 Rules table is a structural INDEX
+("which rule files exist, grouped by category"). Section 8's first
+sub-section serves a different purpose: it documents WHICH rules
+auto-load on EVERY edit (i.e., the universal guardrails the developer
+should keep mentally active), plus a meta-summary of what each enforces.
+These two views are complementary, not redundant — Section 6 shows the
+map, Section 8 sub-section 1 shows the always-on layer.
+
+**Sub-section 1 (REQUIRED)**: `### Common Rules (auto-loaded on every edit)` (emit heading in target output language)
+
+- 2-column or 3-column table. Recommended: `| Rule File | Role | Core Enforcement |`
+- Include ONLY rules whose `paths` frontmatter is `["**/*"]` (universal
+  auto-load). Typical entries:
+  - `.claude/rules/00.core/51.doc-writing-rules.md`
+  - `.claude/rules/00.core/52.ai-work-rules.md`
+  - Any other `paths: ["**/*"]` rule generated for this stack.
+- Each row's "Core Enforcement" column must be a META SUMMARY (one
+  line, what the rule enforces at a high level), NOT a copy of the
+  rule body.
+  - ✅ "paths frontmatter required, no hardcoding, topic repetition is acknowledged as intentional reinforcement"
+  - ❌ Copying 5 paragraphs from the rule file
+- End with a 1-line reference: "For detailed guidance, Read `claudeos-core/standard/...`"
+  when a corresponding standard file exists.
+
+**Sub-section 2 (REQUIRED)**: `### L4 Memory (on-demand reference)` (emit heading in target output language)
+
+Intro (1-2 sentences, fixed template — emit in target output language):
+```
+Long-term context (decisions · failures · compaction · auto-proposals) is stored in `claudeos-core/memory/`.
+Unlike rules that auto-load via `paths` glob, this layer is referenced **on-demand**.
+```
+
+Table: `#### L4 Memory Files` — FIXED 4-row table
+- decision-log.md — Append-only. Permanent record of the "why" behind design choices.
+- failure-patterns.md — Repeated errors and their remedies (conditions: ≥2 occurrences, non-obvious root cause, undocumented remedy).
+- compaction.md — 4-stage compaction policy.
+- auto-rule-update.md — Machine-generated rule change proposals (regenerate via `npx claudeos-core memory propose-rules`).
+
+Workflow: `#### Memory Workflow` — FIXED 6-step numbered list
+1. Session start — failure-patterns scan
+2. Skim recent decisions — decision-log recent
+3. Record new decisions — decision-log append
+4. Record repeated errors — failure-patterns add
+5. Periodic compaction — memory compact command
+6. Review rule-update proposals — auto-rule-update review
+
+**Section 8 single-occurrence rule** (enforces the "one canonical home"
+principle):
+- The L4 Memory Files table appears EXACTLY ONCE in the entire document
+  — inside sub-section 2 (`L4 Memory`). No other location.
+- The Memory Workflow (6-step numbered list) appears EXACTLY ONCE — same
+  location, after the L4 Memory Files table.
+- The Common Rules meta-summary table appears EXACTLY ONCE — inside
+  sub-section 1. No sibling, no echo, no restatement.
+
+**Out-of-scope content** (do NOT place under Section 8):
+- Project-specific memory details — those go inside `memory/*.md` files
+- Master plan references — `plan/` was removed in v2.1.0
+
+**Language**: Use `{OUTPUT_LANG}` throughout this section. All headings,
+table column labels, intro sentences, and workflow step labels shown
+above in English MUST be emitted in the target output language.
+
+---
+
+## Dynamic substitution variables
+
+| Variable | Source | Example |
+|---|---|---|
+| `{PROJECT_NAME}` | pass2-merged.json `project.name` or directory name | `my-backend-api` |
+| `{OUTPUT_LANG}` | Target output language resolved from the CLI `--lang` flag | `en`, `ko`, `ja`, `zh-CN`, ... (10 supported) |
+| `{PROJECT_CONTEXT}` | LLM-generated Level 2 description | See Section 1 rules |
+
+---
+
+## Validation checks (post-generation)
+
+**MANDATORY FINAL CHECK — count `^## ` lines in the generated CLAUDE.md.**
+The count MUST be exactly 8. If it is 9 or more, a surplus section was created
+— locate it and either merge its content into the correct section from 1-8, 
+move it to `.claude/rules/*` or `standard/*`, or delete it entirely if it
+duplicates existing rules. Then recount.
+
+Common surplus sections that MUST be removed/merged (match by
+semantic meaning in ANY output language — the English labels below
+are canonical, target-language renderings are equivalent):
+- "Common Rules" (as a `##` section) → move content to Section 8
+  sub-section 1 ("Common Rules (auto-loaded on every edit)")
+- "L4 Memory Integration Rules" / "Memory Layer (L4)" (as a `##`
+  section) → merge into Section 8 sub-section 2; delete duplicate
+  tables/workflows
+- "Documentation Writing Rules" / "AI Common Rules" (as standalone
+  `##` sections) → move as rows into Section 8 sub-section 1's
+  common-rules table
+- "Required to Observe While Working" / "Rules Summary" → move
+  enforcement content to `.claude/rules/*`, not CLAUDE.md
+- Any section whose title has the semantic meaning of "rules" beyond
+  Section 8's "Common Rules & Memory (L4)" top-level title
+
+Before finalizing, verify:
+
+- [ ] **Exactly 8 `##` top-level sections (count them!)**
+- [ ] Section titles match the FIXED set (Role Definition, Project
+      Overview, Build & Run Commands, Core Architecture, Directory
+      Structure, Standard / Rules / Skills Reference, DO NOT Read,
+      Common Rules & Memory (L4)) — rendered in the target output
+      language with equivalent meaning and order
+- [ ] Section order is correct
+- [ ] No sections renamed, merged, or split
+- [ ] No sections added beyond the 8
+- [ ] Section 1 is 2 lines
+- [ ] Section 2 is 8-12 rows table
+- [ ] Section 5 has at most 3 emphasis bullets below tree
+- [ ] Section 6 has EXACTLY 3 sub-sections (Standard / Rules / Skills)
+- [ ] Section 6 Rules row for `00.core/*` includes a reference pointing
+      to Section 8's Common Rules table for auto-loaded rules
+- [ ] Section 7 has the 3 common claudeos-core/* entries
+- [ ] Section 8 has EXACTLY 2 sub-sections (Common Rules + L4 Memory)
+- [ ] Section 8 sub-section 1 (Common Rules) lists ONLY rules with `paths: ["**/*"]`
+- [ ] Section 8 sub-section 1 meta-summary column is 1 line per row, NOT a rule-body copy
+- [ ] Section 8 sub-section 2 L4 memory table has exactly 4 rows
+- [ ] Section 8 sub-section 2 workflow has exactly 6 steps
+- [ ] Section 8 L4 table and workflow appear EXACTLY ONCE in the whole document
+- [ ] Section 1 PROJECT_CONTEXT contains NO hardcoded component/product names
+
+---
+
+## Examples
+
+### Example: Section 1 for different stacks
+
+Examples below are shown in English. Emit the final output in the
+target output language; the semantic content should match.
+
+**Java Spring Boot backend**:
+```
+As the senior developer for this repository, you are responsible for writing, modifying, and reviewing code. Responses must be written in {OUTPUT_LANG}.
+A Java Spring Boot backend application with a layered architecture, implemented on top of a dual RDBMS with an XML-based SQL mapper and a read/write mapper split.
+```
+
+**React + Vite Back Office SPA**:
+```
+As the senior developer for this repository, you are responsible for writing, modifying, and reviewing code. Responses must be written in {OUTPUT_LANG}.
+A React + Vite Back Office SPA that, on top of multiple HTML entry points and per-domain page separation, follows a 3-tier authentication approach combining API response normalization, session guards, and button-level permission checks.
+```
+
+**Python Django API**:
+```
+As the senior developer for this repository, you are responsible for writing, modifying, and reviewing code. Responses must be written in {OUTPUT_LANG}.
+A Python Django REST API server providing a multi-tenant structure on top of an RDBMS + cache layer, including an asynchronous task queue.
+```
+
+---
+
+## Usage from pass3 prompts
+
+Pass 3b prompts for each stack MUST reference this scaffold:
+
+```
+CLAUDE.md generation MUST follow the scaffold at:
+`pass-prompts/templates/common/claude-md-scaffold.md`
+
+Section structure, titles, and order are FIXED.
+Content within each section adapts to this project based on pass2-merged.json.
+```
+
+The scaffold supersedes any previous CLAUDE.md generation instructions.