/interview-cheatsheet — long-form Chinese ML/LLM interview prep

Generate one comprehensive Chinese cheat sheet per invocation: formulas + derivations + from-scratch code + 25 高频题. Output receives fresh-agent same-family provisional math/code review before rendering. Detect-only by default: never auto-commits.

Inputs

• <topic> (required) — narrow enough for one 600-1000 line tutorial. Good: "RLHF / DPO / PPO", "MoE", "KV Cache + Speculative Decoding". Bad (too broad): "all of LLM training", "diffusion" (split into Forward Process / Sampling / CFG separately).
• --effort (default balanced) — balanced ≈ 600 lines, max ≈ 1000 lines with deeper proofs and more L3 questions.
• --byline (default "<Your Name>, <Affiliation>") — passed to /render-html --author.
• --commit (default false) — if false (default), stop after rendering; user reviews and commits. Never push without explicit user approval.

Style guide — STRICT (read docs/tutorials/attention_tutorial.md as canonical reference)

Section skeleton (12-14 sections)

## §0 TL;DR — callout intro line + numbered list of 5-7 takeaways
## §1 直觉 — why this matters; analogy; one-paragraph mental model
## §2 核心公式 — main formula + derivation (variance / scaling / boundary)
## §3 实现细节 — 50-80 line from-scratch PyTorch
## §4-7 变体 / 工程实践 / 常见 bug — variants, comparison tables, footguns
## §8 复杂度 / 资源 — time + memory complexity
## §9 与相关方法对比 — placement in the ecosystem
## §10 25 高频面试题 — L1 (10 必会) + L2 (10 进阶) + L3 (5 顶级 lab), all with <details><summary> collapsible answers
## §A 附录 (optional) — sanity-check output, reference list

Conventions — bake the established lessons in

Rule	Why	Example
Heading format ## §N Title with space after §N	Older versions had §0TL;DR glued	## §0 TL;DR Cheat Sheet
Math in table cells: use \lvert ... \rvert not |...|	| inside markdown table = cell separator → row break	$\text{score}_{ij} - m \cdot \lvert i-j \rvert$
Callouts with body list: split into callout intro line + separate list	Otherwise the list's first item is swallowed by the callout, then items 2..N restart numbering at 1	> 💡 **Sampler 选择** — 按 NFE/质量排序如下。 - Euler … - Heun …
Callout prefixes only: 💡 ⚠️ ✅ ❌ (others won't get class)	renderer maps these to callout-info/warn/good/bad	> ⚠️ **FP16 overflow** — 即使除了 √d_k …
Math: $...$ inline, $$...$$ display, $$\boxed{...}$$ for key boxes	MathJax CDN; literal in source	—
Code: ```python fences, real PyTorch that would run	reviewer will check executability	—
Personal-info banlist: owner's institution/lab/center names, degree-program affiliations, private server aliases, job-search context, /Users/... paths, specific lab/company names	reviewer flags as FAIL	byline goes via --author at render time, not in body
Language: Chinese primary, English technical terms in-place	matches established cheat-sheet style	"softmax 饱和", "vector field"

Eyebrow / subtitle / title naming

Field	Pattern
--eyebrow	Interview Prep · <Topic>
--subtitle	one Chinese sentence describing scope (e.g. 公式推导 + From-Scratch 代码 + 25 高频题（L1 必会 · L2 进阶 · L3 顶级 lab）)
--title	<Topic> 面试 Cheat Sheet or <Topic> Quick Reference
--lang	zh-CN

Slug

<topic> → kebab/snake-case <slug> for filenames. e.g. "RLHF / DPO / PPO" → rlhf_dpo_ppo.

Workflow

Step 1 — Plan structure (no files written)

Internally sketch:

• 12-14 section titles
• List of major formulas (with derivation outline for each)
• List of code blocks (skeleton + what it demonstrates)
• 25 interview questions sorted by L1 / L2 / L3 difficulty (each with one-line expected answer)
• Comparison table topics (e.g., "RLHF vs DPO vs IPO vs SimPO")

If the topic is too broad to fit in one cheat sheet, stop and ask the user to scope before drafting.

Step 2 — Draft MD

Write directly to docs/tutorials/<slug>_tutorial.md. Follow the style guide. Length target: 600 lines (balanced) or 1000 lines (max), ±20%.

Step 3 — Fresh-agent math/code review (Codex GPT-6-Astra xhigh, same-family provisional)

Invoke spawn_agent with model: gpt-6-astra, reasoning_effort: xhigh, and a fresh thread. Do not reuse prior reviewer context.

Reviewer prompt:

You are reviewing a long-form Chinese interview-prep tutorial on <TOPIC> for math/code/factual correctness and style discipline.

## Files to read (READ-ONLY)
- Draft MD: <MD_PATH>
- Style reference: docs/tutorials/attention_tutorial.md
  (Read this only for STYLE — do NOT score the draft against the reference's content topic.)

## Return JSON with these 10 checks

1. formula_correctness — Independently re-derive each $$ display formula. Flag any error with file:line.
2. code_correctness — For each python block: would it run? Does it implement the stated math? Imports / shapes / device handling consistent?
3. interview_answer_correctness — Each L1/L2/L3 question's <details> answer. Specifically flag wrong year / wrong paper / wrong author / off-by-one indexing / inverted comparison.
4. historical_citations — Paper authors + year + venue. Flag wrong attributions (e.g., "DPO: Rafailov 2023 NeurIPS" must be checkable).
5. table_pipe_escape — Any markdown table cell containing `|x|` math (not `\lvert x \rvert`)? Cite line.
6. callout_list_collision — Any line matching the pattern `^> (?:💡|⚠️|✅|❌) \*\*[^*]+\*\* — (?:- |\d+\. )`? That swallows the list.
7. heading_consistency — All `## §N` and `### N.M` follow style guide (space after §N, no glued chars).
8. section_completeness — Sections §0..§10 (and §A if effort=max) present and non-trivial.
9. length_target — Within ±20% of target (600 for balanced, 1000 for max).
10. personal_info_leak — None of: the owner's institution / lab / center names, degree-program affiliations, private server aliases, job-search or recruitment context, absolute `/Users/...` paths. (Keep the concrete string banlist in local untracked notes — the public SKILL defines only the CATEGORIES; listing the real values here would itself be the leak.)

Return JSON:
{
  "verdict": "PASS | WARN | FAIL",
  "checks": {<check_name>: "pass|warn|fail with one-line note + file:line if applicable"},
  "blocking_issues": ["..."],
  "warnings": ["..."]
}

Verdict: PASS = all pass, WARN = at most cosmetic issues (length slight off / cosmetic style), FAIL = any math/code/factual error OR personal-info leak OR table-pipe / callout-list bug.

Step 4 — Fix and loop (no hard cap — judge by trajectory)

For each FAIL issue, edit the MD. Then re-invoke Codex with a fresh spawn_agent call (never continue with send_input). Stop when verdict = PASS or WARN with no FAIL items.

No hard round cap. Use these heuristics instead:

• ✅ Keep going if each round's FAIL items are shrinking, concrete, enumerable (e.g., citation year fixes, off-by-one, single-line code bugs). The reviewer is doing useful work — let it converge.
• ⛔ Stop and report if the same issue keeps coming back (loop detected), or if the FAIL items shift to architectural / scope concerns that need user input, or if the round count exceeds ~6 without convergence.

Most tutorials converge in 3-5 rounds. Going to 5-6 rounds is fine if substantive bugs are still being caught — the Video Generation tutorial (May 2026) went to 5 rounds and the final 2 rounds caught real citation errors and an over-attribution to Sora's patch size that would have shipped otherwise.

Step 5 — Render via /render-html

Call directly (do not invoke /render-html as a sub-skill; call its python script — gives clear control):

python3 skills/render-html/scripts/render_html.py docs/tutorials/<slug>_tutorial.md \
  --template academic \
  --out docs/tutorials/<slug>_tutorial.html \
  --title "<Topic> 面试 Cheat Sheet" \
  --subtitle "<one-line scope summary>" \
  --eyebrow "Interview Prep · <Topic>" \
  --author "<byline>" \
  --lang zh-CN

render_html.py runs its own 13-check codex review automatically. If that FAILs, fix the MD (often a table-pipe or callout-list issue the math/code reviewer missed) and re-render. Note that render_html.py itself writes <slug>_tutorial.review.json for the render-stage audit.

Step 6 — Combine audit trail

After both reviews pass, merge math/code review history + render review history into one docs/tutorials/<slug>_tutorial.review.json:

{
  "skill": "interview-cheatsheet",
  "source": "docs/tutorials/<slug>_tutorial.md",
  "source_sha256_prefix": "<16-char prefix>",
  "output": "docs/tutorials/<slug>_tutorial.html",
  "topic": "<TOPIC>",
  "effort": "balanced | max",
  "byline": "<author string>",
  "math_code_review": {
    "verdict": "PASS",
    "rounds": [
      {"run": 1, "verdict": "...", "thread_id": "...", "issue": "...", "fix": "..."},
      ...
    ]
  },
  "render_review": {
    "verdict": "PASS",
    "rounds": [...]
  },
  "summary": "<one-line: N-round math/code review + M-round render review settled at PASS>",
  "rendered_at": "<YYYY-MM-DD>"
}

Step 7 — Stop. Report to user.

Do NOT git add / git commit / git push. Report:

✅ /interview-cheatsheet "<TOPIC>" complete.

  Files:
    docs/tutorials/<slug>_tutorial.md          (<lines> lines, <bytes> bytes)
    docs/tutorials/<slug>_tutorial.html        (<bytes> bytes, <TOC> TOC entries)
    docs/tutorials/<slug>_tutorial.review.json

  Math/code review:  PASS after <N> rounds (<thread IDs>)
  Render review:     PASS after <M> rounds
  Length:            <actual> lines (target <effort>)

  Issues caught + fixed during review:
    - <one line per non-trivial fix>

  Suggested commit message:
    docs(tutorials): add <Topic> cheat sheet (rendered via /render-html)

  ⚠️ Did NOT auto-commit — user reviews and pushes manually.
  Also update docs/tutorials/README.md to add the new row.

Update the index

After the tutorial passes, optionally append a row to docs/tutorials/README.md:

| **<Topic> 面试 Cheat Sheet** | [`<slug>_tutorial.md`](<slug>_tutorial.md) | [`<slug>_tutorial.html`](https://wanshuiyin.github.io/Auto-claude-code-research-in-sleep/tutorials/<slug>_tutorial.html) | <one-line topic list> |

Suggest the row to the user but let them edit it in themselves if they want to curate.

Key invariants (the ARIS rules baked in)

Invariant	How it's enforced
Executor/reviewer family	Codex drafts; fresh gpt-6-astra reviews (math/code stage and render stage), same-family provisional
Fresh agent per reviewer call	Step 3 + render's own gate both use fresh spawn_agent calls, not send_input
Codex reasoning = xhigh	Hardcoded in Step 3 reviewer config
Personal info redaction	Both math/code reviewer and render reviewer check; banlist in style guide
Lessons-learned encoded	Table-pipe + callout-list collision rules in style guide AND review checks 5+6
No silent failure	If review FAILs and the FAIL set is no longer shrinking (loop) or hits ~6 rounds without convergence, stop and report — don't push

When NOT to use

• Topic too broad — split into smaller scopes first
• Topic outside ML/LLM core — this style guide assumes math + code + Chinese; for general topics use a different format or write directly
• Already have a draft you want to edit — use Edit directly, this skill is for greenfield generation
• Don't want HTML output — call /render-html separately or skip Step 5

Reference invocations

/interview-cheatsheet "RLHF / DPO / PPO"
/interview-cheatsheet "MoE (Mixture-of-Experts)" — effort: max
/interview-cheatsheet "KV Cache + Speculative Decoding"
/interview-cheatsheet "Long-context: RoPE / YaRN / NTK / MLA"
/interview-cheatsheet "Distributed Training (DDP / FSDP / ZeRO / TP / PP)"
/interview-cheatsheet "Quantization (GPTQ / AWQ / INT4 / FP8 / SmoothQuant)"

Reference style files

• Style canonical: docs/tutorials/attention_tutorial.md + .html
• Style secondary: docs/tutorials/flow_matching_tutorial.md + .html
• Review audit format: docs/tutorials/attention_tutorial.review.json

Provenance

Extracted from the two pilot tutorials (Attention + Flow Matching, May 2026). Both passed fresh-agent review; that review is same-family provisional in the base Codex mirror. The attention tutorial required 3 rounds and caught a table-pipe collision plus a callout-list collision.