Documentation Standards¶

These standards keep the repository consistent, accurate and safe to publish. Every contribution, human or AI-assisted, follows them. CI enforces the mechanical parts.

1. Accuracy and sourcing¶

Official sources first. Prefer, in order: the official exam guides, platform.claude.com/docs, code.claude.com/docs, modelcontextprotocol.io, support.claude.com, anthropic.com/engineering, then OWASP / NIST / other standards bodies. Community material is supporting evidence only.
Exam facts live in one place. Item counts, fees, pass marks, domains and weights come from 16-certification-tracks/registry.yaml. Any page that states an exam fact links to the registry or to exam-facts.md. tools/check_exam_claims.py enforces this.
Label anything unverified. If you cannot confirm a fact against an official source, write > **Unverified:** … and explain what is uncertain. Never invent exam details.
Date-sensitive facts carry a date. Model names, prices, context windows and feature availability change. Write them as "as of 2026-10-24" and link to the live page (models overview, pricing).
Current API rules (as of 2026-10-24) that examples must respect:
Model IDs: claude-opus-5-5 (default), claude-sonnet-5-5, claude-haiku-5-5, claude-fable-5-1. No date suffixes.
Do not send temperature, top_p or top_k. Claude Opus 5.5 rejects them; Sonnet 5.5 and Haiku 5.5 reject non-default values; and on API version 2026-10-21 every model rejects them. Tune prompts or effort instead.
No assistant-message prefill. Use structured outputs (output_config.format or client.messages.parse(...)) or instructions to shape output.
tool_choice of any or tool returns a 400 on Claude Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1 (Haiku 5.5 still accepts it). Use auto plus a clear instruction, strict: true tools, or structured outputs.
Thinking is adaptive (thinking={"type": "adaptive"} or omitted). Depth is tuned with output_config={"effort": "low" | "medium" | "high" | "xhigh" | "max"}. budget_tokens is legacy.
Prompt caching is a prefix match: stable content first, cache_control breakpoints, check usage.cache_read_input_tokens.
Always check stop_reason (end_turn, max_tokens, stop_sequence, tool_use, pause_turn, refusal, model_context_window_exceeded) before using content.

2. Original work and intellectual property¶

Write every explanation, example, diagram and question yourself. Do not copy tutorials, courses, repository docs, blog posts, diagrams or question banks, and do not lightly reword them.
Never reproduce, reconstruct or "remember" real exam questions. Do not reuse or paraphrase the sample questions printed in the official exam guides. Anthropic's exam policy treats exam content as confidential.
Quote sparingly (a sentence at most), in quotation marks, with attribution and a link.
Link to official resources rather than restating them at length.
Use fictional organisations (e.g. Northwind Lending, Contoso Health, Acme Logistics) and synthetic data. No real customer data, no real personal data, and no names of real employers.

3. Teaching structure for a topic page¶

Each concept page in modules 00–15 uses these headings, in this order. Merge or omit a heading only when it truly does not apply, and avoid repeating material covered elsewhere. Link to it instead.

# <Topic>
> **Level:** Beginner | Intermediate | Advanced · **Time:** ~N min · **Exam relevance:** CCAO-F D3, CCAR-P D2

## In plain English          <!-- beginner-friendly explanation, analogy welcome -->
## Why it matters
## How it works             <!-- technical explanation -->
## Example                  <!-- practical, runnable or copy-pasteable -->
## Enterprise use case      <!-- fictional organisation -->
## Common mistakes
## Security considerations
## Troubleshooting
## Check your understanding <!-- 3–5 questions, answers in a <details> block -->
## Further reading          <!-- official docs first -->

Reference pages that collect material rather than teach one topic (for example 12-ai-security/owasp-llm-mapping.md) are exempt from this template.

Each module folder has a README.md that gives the learning objectives, the page order, prerequisites, exam relevance (objective IDs) and a "next module" link.

4. Writing style¶

Plain, direct English. Spell out acronyms the first time they appear in a page.
Second person ("you"), active voice, short paragraphs.
Use tables for comparisons and Mermaid (```mermaid) for diagrams. GitHub renders both natively.
Code blocks always carry a language tag. Python is the primary language. Include TypeScript or cURL only where it adds something.
Relative links only for files inside the repository; link targets must exist (tools/check_links.py).
No marketing language, no emojis in body text, no fabricated statistics or quotes.

5. Code standards¶

Python 3.10+ (3.11+ for the optional MCP server in lab 09, because the mcp v2 SDK requires it). Labs import the shared helper 19-hands-on-labs/common/claude_lab.py, so they run offline in mock mode by default and against the real API with CLAUDE_LAB_MODE=live.
Never hardcode credentials. Read them from environment variables, and ship a .env.example with placeholder values. Real .env files are git-ignored.
Every runnable sample has a test in tests/ or is executed by tools/run_lab_smoke.py in mock mode.
Mark any step that costs money: > **Cost:** requires API credits (live mode).
Treat model output as untrusted input: validate before acting on it, never eval it, and escape it before rendering.

6. Practice questions¶

See 17-practice-questions/README.md. In short: JSON matching the schema, one file per domain, unique IDs, every distractor explained, every question mapped to an objective and a module, and every question marked unofficial.

7. Review checklist (copied into the PR template)¶

Original wording; no copied or reconstructed exam content
Facts checked against official sources; date-sensitive facts dated; unverified facts labelled
Links resolve (python tools/validate_all.py)
Code runs in mock mode; no secrets; costs flagged
Security considerations included where relevant