Writing for penguin.ooo

Posts, docs pages and landing copy are written by a model on the TokenDance gateway and shipped by agents. A post is written the way a person writes one: an outline first, then the whole post in one call, so its sections join; the review, the fixes, the cuts and the translation work on the whole post too, and a round of its own edits only the sentences at the headings. A long post or a docs page can fall back to section mode. This page is the contract; scripts/prose.mjs is the only way text reaches a draft, and prompts/ plus reference/ hold every word sent to the model. penguin-harness-dev/reference/authoring.md covers what is true of every post (the bilingual pair, frontmatter, where screenshots live).

1. Who writes what

The API model writes every word a reader sees: the outline and its plan, every section, the title and excerpt, every string inside a figure, alt text and captions, the second language, every docs sentence and landing string, and every later fix or cut. Each is its own API call through prose.mjs.

Agents never type a sentence into the text. They research, compile FACTS, write prompts and findings, put the model's text together, draw figures from the model's figure blocks, take screenshots, run the checks, review blind and ship. Putting the text into the manuscript is also the agent's job, and it is mechanical: write saves the draft, assemble joins section files in section mode, and strip publishes them. A wrong sentence goes back through a fix call with a numbered finding, never through an editor. The only hand edits are mechanical: frontmatter date / category, screenshot paths and their -en / -zh suffixes, whitespace, the boilerplate version note copied verbatim from an existing post, and copying landing strings into the dictionaries verbatim.

The key never enters the repo. It sits in a chmod-600 file outside the tree; the script reads TOKENDANCE_API_KEY or the file named by TOKENDANCE_KEY_FILE, never an argument, and prints neither. Raw responses (with the model's reasoning) are saved in a scratch directory and never committed.

The default model is seed-2.1-pro, a reasoning model: a call can take minutes, its reasoning counts against max_tokens (32000 by default), and finish_reason: length is a failure. The key has a concurrency cap; a post's calls run one after another, and the script's 429 backoff absorbs the rest. --model overrides the default per call.

2. Profiles

What differs by kind of text lives in one profile, reference/profile-<name>.md, chosen with --profile <name> (or --category, its shorthand; by default the input's frontmatter decides):

A blog category without a profile file needs --profile <file> until one is written; --profile <file> also tries a profile before it lands. The file names its profile by its name (profile-blog-practice.md, or profile-practice.md for a blog category) or by the --profile <name> its header says it is loaded for, and the script says which profile it loaded.

A profile holds the reader, the shape, the sizes and any rule that overrides the style guide for that kind of text, inside <!-- prompt --> markers (a <!-- prompt zh --> block, when present, replaces it for Chinese). The model receives it after the style guide, introduced as winning where the two differ. A profile does not repeat the style guide and states nothing about the product: facts come from FACTS only. Its ## Check settings are what the script enforces — sections, lead, tldr, exclamations, and the budgets words, lead-words, section-words, zh-chars, zh-lead-chars, zh-section-chars, tldr-words, zh-tldr-chars, each floor-cap, emphasis with its caps emphasis-words / zh-emphasis-chars (section 4), sources, the floor and cap of a post's reference list (section 7), body-lists and body-tables (none makes a list outside the TL;DR, or a Markdown table, an error, for a post written as an essay in paragraphs), and ui-labels: quoted in a post, where bold is only the emphasized sentence and a UI label goes in quotes ("…" / 「…」), bold in a docs page. A prompt rule with no matching setting is a rule check cannot see; change the two together.

3. Voice

reference/en-style.md and reference/zh-style.md hold the voice and the banned lists for every kind of text; the outline, section, translate, polish, fix and review calls all receive them.

One rule there asks the model to write for a reader with dyslexia or ADHD, a person who reads with effort and loses the thread easily: the point first, short sentences with one idea each, common words, no asides or "by the way" detours, no defensive hedging, no padding. Stating the reader that way measurably shortens the model's output and keeps it on its point. The one habit it does not remove is the "Yes, but…" / 「是的，但是……」 concession pivot, so that pivot is banned by name next to the "not X but Y" pivot, and check and review look for it. For the same reader a post opens with a TL;DR of its key points, and one sentence per section is marked in bold by emphasize; the writer bolds nothing else.

The tip this rule comes from (the author's post, verbatim):

亲测超级好用😂 | 之前为了矫正大模型的说话问题（RL不是没有代价的，你就是代价……），我整过很多版本的工作要求，最极端的是输出前多加一步子Agent专门负责矫正Agent汇报……，然而，真的不如假装残疾人好使。方法就一行⬇️【直接在Agent.md里加入，假设用户是阅读障碍。】⬆️大模型就会说人话了，也不长篇大论了，也会捡重点的说了，也不防御性写作了，也不BTW了唯一扳不过来的是【Yes, But】这个有主意朋友可以跟我说说。现在 LLM 也很难低成本私有化部署和调试，各家又都是冲着把人干掉去训练的，结果现在工作就是个可悲的【人机打架】的L3阶段。而且就DeepSeek而言，我老觉得他有点文本代码化的倾向，就是完全没有脑子……比之前还没有那种，不做真正的思考，纯无脑执行那种。

The rule keeps the tip's one line, assume a reader with dyslexia, and turns the rest of its report into bans: its "BTW" detours and its defensive writing are the aside and hedging bans, and the "Yes, but" pivot the tip could not remove is banned by name.

The zh style guide keeps Harness and Agent Harness in English with that capitalisation, never 「夹具」「线束」「挽具」; a test fixture is 合成测试, 测试数据 or 测试配置; check makes 夹具 an error and a lowercase "harness" a warning.

4. The pipeline

All work happens in a scratch directory, one folder per post: <scratch>/articles/<slug>/ for inputs, the outline, the notes and the drafts, <scratch>/articles/raw/ for raw responses (--raw-dir). The folder name is the slug the figure ids carry; the .en. / .zh. token in a file name is how the script knows the language. A post is written as one document:

Flow. The post is one argument written in one sitting. The first sentence after every ## heading continues the last sentence before it; no section opens with a signpost ("In this section", "Next,", 「接下来」) or ends on a sentence that only announces the next heading. The plan's ends with column names the point a section lands on, not a sentence to write. write states this, review scores every join, transition mends the joins that fail without touching anything else, and check warns on the signposts the style guides list. A post's joins get a round of their own because a fix round aimed at the whole text leaves them broken: the author's rule is to write the whole post, revise it as a whole, then edit only the sentences at the headings.

Emphasis. A profile with emphasis: yes gets one marked key sentence, a whole one, per ## section and one in the lead: the claim a reader skimming the section must not miss. emphasize adds the ** pairs to a frozen draft and changes nothing else; --notes passes the author's notes as guidance on which sentence to mark. The script refuses an answer that differs from its input once the ** are removed, has no span or two in a section, holds a span over emphasis-words / zh-emphasis-chars or over part of a sentence (a span opens where its sentence opens and closes right before its final punctuation), or puts one in a heading, figure block, caption, version note, TL;DR, code span or link, or around a marker. The closing ** goes before the sentence's punctuation and markers, so the citation follows it there (**claim**[^1]., **句子**[^1]。), and the script refuses the other order: after Chinese punctuation a closing ** is not read as bold (**关键句。**后文 shows its asterisks). Bold in a post is only that sentence: the writing prompts put a UI label in quotes, and check makes any other bold an error. A text that changes after emphasize is assembled and emphasized again.

The ordering rule. The outline is frozen before the text is written; the text of a language is frozen before any figure of it is drawn; and no string inside a figure is ours. A figure that needs a string its block lacks goes back through a fix finding.

Section mode

A post well over a thousand words, or a docs page written task by task, can be written one section per call instead. section writes plan row n from the frozen outline, seeing the plan, the sections already written and only that row's facts, into sections/s<n>.<lang>.v<k>.md, and check --section follows every call (size alone → trim, shape → section --feedback, wording → fix). assemble joins the latest section files byte for byte, or the versions --pin s<n>=v<k> names. route splits merged findings by the section that quotes each one, and fix, translate, polish, review and trim then run per section, with the neighbours read-only. Their prompts aim at the lowest of the planned size, what the page has left and the profile's cap (section 5), and --page <assembled draft> tells them which version of each other section the page holds. From emphasize on, the flow is the same. Section mode trades the joins for control, so a post written this way needs its transitions checked by review on the assembled draft and mended by transition.

5. Budgets

Sizes are set three times, so a post cannot grow past them unnoticed:

1. The plan. The outline's # Plan table gives every section a number, in English words or Chinese characters (Chinese-first posts plan in characters). plan rejects a table whose rows exceed the profile's section cap, whose lead leaves the lead range, whose sum exceeds the total cap, or whose section count leaves the profile's range.
2. The prompt. A whole-post write, translate, fix or polish states the profile's caps for the whole, the lead, each section and the TL;DR, and write takes the plan's row sizes as targets; each prints, after its call, every section against its cap and plan and the page against its cap. In section mode, every call that writes or changes a section (section, translate, polish, fix, trim) asks for the lowest of three sizes: the section's planned size (from a plan in the other language's unit, converted at the translation ratio), what the page has left once the other sections take theirs, and the profile's cap. The other sections count as --page <assembled draft> holds them, so a pinned older version counts and a newer file set aside does not; without --page, the newest file of each, or the planned size of one not yet written. It never offers check's 1.1 × tolerance, so a round of fixes cannot walk every section up to it and the page over its cap. polish, fix and trim also state the section's current size, and outline and a whole-piece write the profile's caps. Each states the counting rule check applies, in words and with an example sentence the script counts itself, so a cap is one number for the model and for check. The wording is the count-rule-<lang> fragment; change it with countBody.
3. check. A section above min(plan × 1.1, profile cap) is an error; with --plan-rows advisory, the mode for a post written as one document, a section above its plan row is a warning and only the profile's cap an error. A lead above its cap and a whole post above the total cap are errors either way. Without --outline, the profile caps apply. Below a floor is a warning: short is allowed, empty is not. On a section file it also reports the page total, counted the same way as the prompt's page budget (with --page too), names where the other sections' sizes came from, and warns when the total is over the page cap; every per-section API step prints the same line, so creep shows before assembly.

A TL;DR, in a profile that has one (tldr: yes), is not part of the lead: the **TL;DR** line and its bullets are counted apart and held to tldr-words / zh-tldr-chars, the plan's row 0 sizes the lead alone, and the lead's section call states the TL;DR budget separately.

Over budget. A whole draft over a cap goes through trim --in draft.<lang>.vN.md, which names the sections to cut and asks for just enough. A section over its profile cap aims at 95 % of that cap, never at its plan row. If the page is still over its cap, the rest of the cut, down to 95 % of the page cap, is shared among the other sections in proportion to their size; the lead is left alone. Each named section has a floor at 90 % of its target. The prompt asks for rephrasing inside complete sentences and keeps every fact: numbers with their units, approximations ("about", 约), attributions, and qualifiers as strong as they are ("many" never becomes "most"). trim refuses an answer that changes any other section, leaves a named section or the page over its cap, cuts a named section below its floor, or loses a number the text stated, so a telegraphic cut is rerun, not shipped. Figure blocks are not the model's to cut: a block the answer changed is put back from the draft before the guard runs. In section mode, a section whose only problem is its size goes through trim --in <section> --outline <outline> (or --cap <n>). Either way the call may only delete words or put a shorter phrase in place of a longer one. The script refuses the answer when it is still over the cap, or when a marker, a footnote reference, a figure block, a heading, the frontmatter or the version note changed: the answer is kept as <out>.rejected.md, the next version is not written, and the step exits 1. Retry once, then hand it to the author. With --page <assembled draft>, trim also cuts a section that is within its own cap to what the page has left when the page is over its cap; without it, the newest section files only produce a note. A section wrong in shape or content is rerun with section --feedback; rerunning it only to shorten it brings back errors already fixed, and so does writing a whole post again only to shorten it. Nobody cuts a section by hand.

Counting is the same everywhere: English words are whitespace tokens holding a letter or digit, Chinese characters are non-whitespace characters once the Markdown syntax is removed, so every punctuation mark and every letter of a Latin word counts; frontmatter, headings, figure blocks, markers and the references do not count. The numbers themselves live in each profile's ## Check settings.

6. The FACTS file and the markers

facts.md has these sections, each a ## heading:

• ## FACTS — atomic, numbered facts, each ending with its source tags in priority order, the primary source first: - F12: … [S3] or [S3 S4]. Only the first public source of a fact is ever cited (section 7), so order the tags deliberately.
• ## SOURCES — one line per source: - S3: <title> · <url>, · private for a source that is never cited (the design repo), · nolink for a public source without a URL. A title is data: the page's own title, a repo file's path, PR #N: <PR title>, PenguinHarness docs: <page>, copied as it is, never reworded. Repo file links are pinned to the release the version note names (blob/v<version>/<path>, after git ls-tree confirms the file is there); blob/main is refused.
• ## OPINIONS — the author's thesis points, ours to assert: - O1: ….
• ## TERMS — spelling, the zh rendering and the UI labels, as a table whose first column is the term; a docs page's TERMS carries the labels exactly as the product's dictionaries spell them.
• ## NOT-FACTS — claims the model must not make.
• ## EXISTING-FIGURES — published screenshots the outline may reuse, by stem.
• ## NOCITE (optional) — facts the writer still marks but that cite nothing: - F12: <why>. Section 7 says what is cited; list here general background, a restatement under another source, and every repo file or docs page the post only refers to. Record the choices in the PR body.

The author's notes. A post may keep a notes.md beside facts.md for wording settled with the author, one line each: keep[ en|zh]: <exact string> (the answer must carry it verbatim), avoid[ en|zh]: <string> (the answer must never hold it) and note: <text> (a point the prompt states). A line tagged with a language applies to that language alone. write, translate, review and fix take the file with --notes; write, translate and fix refuse an answer that breaks a keep: or avoid: line, and review flags a sentence that breaks a note. Keep keep: lines to the few clauses the author settled; everything else is a note:.

The writer makes a factual claim only when FACTS holds it. A section call ends every factual sentence with its id ({F3}, several as {F3}{F7}) and every position with an OPINIONS id ({O1}); headings, figure blocks and frontmatter carry none. translate and polish keep the markers as they are and never add one; fix adds one only on a sentence a finding changed; the script prints any marker a call added or dropped. A fact that needs a citation needs a public source: a fact whose sources are all private may be used, with a warning, and gets no footnote.

7. Citations

The markers become paper-style citations when the post is published, and most sentences carry none: one footnote per paragraph at most, each source once per post, and about 13 sources per post as the ceiling (the profile's sources setting). Writers still mark every factual sentence, since the markers are the truth guard; the citer thins them. strip --facts facts.md reads each marker group (the markers closing one sentence): of its facts, in the writer's order, the first whose first public source is not cited yet is the group's offer, and the first group in a paragraph with an offer becomes the paragraph's one GFM footnote. The paragraph's other groups cite nothing, so a source they offered stays free for a later paragraph. A list item and a table row count as paragraphs. Opinions, private sources and facts under ## NOCITE cite nothing.

Cite what a reader would want to check: a third-party number, a verbatim quote, a named paper, standard or product claim by someone else, our own measurement, and the first mention of a product behaviour. List repo files and docs pages the post merely refers to under ## NOCITE, and name the file in a code span instead.

The TL;DR sums up the body, so its markers stay in the draft as anchors and cite nothing; the body's first use sets the numbers, and a citation in a published TL;DR is an error. The footnote goes before the punctuation that ends its clause, outside a closing quotation mark, bracket or bold (text[^1]., 文字[^1]。, "the slope"[^1]., **claim**[^1].); when the punctuation sits inside the closing mark (slope."), it follows the whole run and check warns. Footnotes are numbered by first use. The post ends with the list:

[^1]: [PenguinHarness docs: Agent loop](https://penguin.ooo/docs/agent-loop)
[^2]: A title without a link

A URL in the text is an autolink, <https://…>. A bare URL right before a citation is an error: strip writes the [^n] straight after it, and the site's renderer reads the reference as part of the link.

The post carries no references heading and no **Sources** line (the older form, now an error): the site labels the list from its own dictionary and renders the numbers as superscript [n] that jump to the list and back. The site numbers a footnote by its first use and builds its anchor from its label, so the labels must run 1, 2, 3 … in the order they first appear; check makes any other order an error.

Each language is therefore numbered by its own first use. strip --numbering-from <the first language's post> keeps the first language's numbers when the second language cites the sources in the same order, and refuses with a clear message when it does not; strip that language without --numbering-from then. A source cited in only one of the two posts is an error either way (strip --numbering-from or pair reports it).

check on a published post proves every [^n] is defined and every definition used, the labels run 1 … N in first-use order, every link is in ## SOURCES, and no citation sits in a figure or a caption; a citation after the punctuation is an error. On a published post, and on a draft read through the same citer with --facts, a paragraph or a sentence with two citations is an error, and so is a post citing more sources than the profile's sources cap; fewer than its floor is a warning. pair proves both posts cite the same set of sources, and each section the same sources, whatever numbers each language gave them.

8. Source language and translation

English first by default. Write Chinese first when the title, a meme or the sources are Chinese: outline --lang zh plans in characters, and the English follows by translation. translate takes the whole post in one call, with TERMS, the notes (--notes) and the pinned title of the target language (--title); it keeps the markers at the ends of the same sentences and translates figure blocks string for string. With --reuse-figures <frozen target draft> --reuse-source <frozen source draft>, a figure whose source block did not change keeps its frozen translation, so it needs no redraw. If the translation still reads translated after one fix round, write that language afresh from the outline (write --lang <tgt>) instead of iterating on it. In section mode, translation runs per section, in order, each call seeing the plan and the sections already translated.

9. Figures

Kinds. Diagrams and charts we draw as inline SVG, real product screenshots in the light theme (captured with the e2e harness and its mock LLM, or reused from the published inventory), or a Markdown table when the comparison is words. No AI-generated illustrations, gradients, glows, shadows, emoji or icons, and no invented data: a conceptual comparison is drawn as states or flows, never as a chart with made-up numbers.

The words come from the model. The section that owns a figure writes its block:

```figure id=<stem>
alt: <one sentence: what is in the picture>
caption: <one sentence: what the reader should take from it>
title: <the heading drawn inside the picture, when it needs one>
text:
- <where>: <a string drawn in the figure>
```

<where> is a short lowercase English key (box, column, layer, step1, config2, …) telling the drawer where the string goes. A block with alt and caption only is a screenshot. A string is data: a placeholder such as <path> in it is escaped by the drawer, never read as HTML.

The manifest's kind column (or the outline's # Figures table) says which figures are screenshots, because a block cannot: one that gained text: lines reads as drawn. check with --figures or --outline makes any line but alt and caption in a screenshot's block an error.

Drawn figures are inline SVG in the post, one file per figure and language, <scratch>/articles/<slug>/figures/<id>.<lang>.html, holding exactly one <figure class="post-figure" data-figure="<id>"> with:

• <svg class="fig-svg" viewBox="0 0 560 <h>" role="img" aria-labelledby="<slug>-<id>-<lang>-title"> whose <title id="…"> is the block's alt;
• a <text> / <tspan> leaf for every text: string, at least once, and no leaf containing a letter (Latin or CJK) that is not a text: string or the block's title (numeric-only leaves — digits, %, ×, #, punctuation — are exempt, so axis ticks need no listing);
• then <ol class="fig-steps"> with one <li> per text: string in order (visually hidden, read by screen readers), then <figcaption> with the block's caption.

Marks carry the site's figure classes from packages/landing/src/styles.css (.post-figure, .fig-svg, the .fig-* text and structure classes, series .fig-s1–.fig-s3 whose meaning the post's legend assigns), never inline colours; dark mode comes from their html.dark variants. No blank line inside a figure (a blank line ends a CommonMark HTML block); <, > and & in strings are escaped; every id starts with <slug>-<id>-<lang>. A chart generated from data is drawn by a throwaway script in the scratch directory that reads the numbers and the figure block and refuses to emit a string not in the block; mirror pairs share one script and one scale. zh and en are drawn from their own blocks, on the same layout, wider where CJK needs it. Read the dataviz skill before any chart and use the site tokens, not its placeholder palette.

check --in <draft> --svg <figures dir> proves each drawn file against its block, in both directions, comparing text after unescaping; strip --figures-dir <figures dir> refuses a file that does not match its block and inlines the rest as one line; check --in <post> --draft <draft> re-runs the same assertions on the published post.

Screenshots stay in Prism-Shadow/penguin-harness-community under blog-assets/, referenced as /blog-assets/<name>-<lang>.png; a new one is merged there before the post's PR. strip writes them as an image line plus an italic caption, escaping <, > and & outside code spans.

Visual check, on the test host only (a browser runs nowhere else): build the landing (--build "@prismshadow/penguin-landing..." with the remote runner), serve it with pnpm --filter @prismshadow/penguin-landing preview --port <p>, and drive a scratch Playwright script over the post in both languages, light and dark (the html.dark class, set the way the site's toggle sets it), at a desktop and a phone width. Assert scrollWidth <= innerWidth, no console or page errors, every .fig-steps hidden and every figure inside the prose column; save the screenshots and look at them (labels not clipped, CJK inside its boxes, strokes visible in dark, the references list and a jump to it). curl there may need --noproxy '*'.

10. Checks

• check (local, no network) exits 1 on a hard hit. For every kind: frontmatter, the profile's budgets, section range and TL;DR rule, no numbered headings, no emoji, raw HTML or leaked {{SLOT}}s (an autolink, <https://…> or <mailto:…>, is a link, not HTML), markers against FACTS and their sources (--facts), and the language rules — the patterns at the end of reference/en-style.md / reference/zh-style.md plus what the script counts (em dashes per section, exclamation marks against the profile, list and paragraph length, CJK–Latin spacing, ——, ASCII quotes and colons in Chinese). For one section (--section, or an s<n>.… file): its shape and the figures its plan row owns. For posts also: ## only, figure blocks complete and in manifest order (--figures), drawn figures (--svg, --draft), screenshot paths and files (--assets), footnotes on a published post, no **Sources** line, no in-page anchors, no list outside the TL;DR or no table where the profile says body-lists: none or body-tables: none, the emphasis rule where the profile has one (section 4; a draft before emphasize has no span, and one with any span needs one in every section), and the citation rules of section 7. In every profile, a ** that CommonMark cannot read as emphasis is an error: a closing run after punctuation and before a letter, an opening run after a letter and before punctuation. Sentences over 35 English words or 60 Chinese characters are warnings. Landing strings are checked as key: text lines. Warnings are for a person to judge.
• The step guards run when an answer arrives, before it is written. write refuses an answer whose version note or pinned title differs from the one given, that breaks a keep: or avoid: line of the notes, or that carries a footnote or, where bold is the emphasis, bold beyond the **TL;DR** line; its report lists the shape, the figures against the manifest and the frozen blocks, and the sizes. On a whole post, fix also refuses a changed heading no finding quotes, a changed title no finding names, and a section added or dropped; on an outline (outline*.md, or a text with # Plan) it validates the plan after the call instead. translate refuses another title than --title and a broken note. fix refuses an answer that brings in a fact or opinion id neither its input nor its findings name, adds, drops or reorders a figure, gives a screenshot's block a title or text: lines, or changes a figure block no finding names by its id or by quoting its text. trim, on a section or a whole draft, refuses what section 5 lists. fix, polish, translate and trim refuse an answer whose version note is not, byte for byte, the one they were given. Before any guard, an answer loses a lone fence line it echoed from the prompt (the steps send the text inside a four-backtick fence), and trim, transition and emphasize put back, by id, any figure block the answer changed; a dropped block is still refused. emphasize refuses what section 4 lists. transition compares the two texts block by block (paragraphs, headings, the TL;DR, figure and code blocks): it refuses an answer that adds, drops, splits or merges a block, changes any block but the paragraph right after a ## heading and the paragraph right before one, changes a sentence of those but the first after the heading and the last before it, changes a rewritten sentence's markers (set and order) or its **, changes the frontmatter, or breaks a note. A TL;DR, list, figure or caption next to a heading is no paragraph, so that side of the join stays as it is. A refused answer is kept as <out>.rejected.md and the step exits 1; a change the findings should allow goes back with a finding that names it.
• The fact diff (agent): every number, URL, code span, bold or 「」 label and proper noun resolves to a FACT or TERM; headings and figures match the outline.
• Review: the API review step (an editor's findings on naturalness, flow, concision, reading effort and figures, in the form fix consumes, never a rewrite; on the whole post, flow across sections is its main question, every join after a heading is scored, and the closing transitions: n of m line goes to <out>.flow.txt; per section in section mode) and the agents' blind reviews — a truth reviewer with FACTS and the sources, a native reader for the second language. The API review reads the text between fence lines, apart from its instruction. It expects the **TL;DR** line and the key sentences emphasize adds (check counts those), takes an objection stated and answered as part of the argument rather than a "Yes, but…" pivot, and never asks for a list where the profile says body-lists: none.

Findings, from any reviewer, are numbered and quote the passage: N. [passage quoted verbatim] — problem — the fact (F#) or rule it violates — what the fix must achieve (not the wording). In section mode route relies on that quote to find the section; a passage it finds nowhere or twice goes to findings.unrouted.<lang>.vN.md for the orchestrator.

11. Assembly and publishing

• assemble (section mode) picks the latest version of each s<n>.<lang>.v<k>.md (or --pin s<n>=v<k>), requires sections 0 … N with frontmatter in section 0, refuses a figure owned by two sections, and prints each section's size against its plan and cap.
• Manual-paste fallback. When assemble cannot run, concatenate the latest section files in order (cat, one blank line between them) into draft.<lang>.vN.md, change nothing, run check --in … --outline …, and say in the PR body that assembly was manual. Typing a word into a section file is never the fallback.
• A post: strip --in <frozen draft> --out packages/landing/content/blog/<slug>.<lang>.md --facts facts.md --figures-dir <figures dir> [--assets <community>/blog-assets], the first language first, then the second with --numbering-from the first's post (or without it when its citations come in another order, see section 7). It prints the model's notes (<!-- omitted: … -->, <!-- not applied: … -->, <!-- figure changed: … -->) so they are read. Then check --in <post> --draft <draft> --facts facts.md --pair <other post>. packages/landing/test/blog.test.ts pins post order per category, so a new post edits that list. Add a changelog pair (Type: process, Scope: landing) per penguin-harness-dev/reference/changelog.md.
• A published post back into a draft: unstrip --in packages/landing/content/blog/<slug>.<lang>.md --out-dir <scratch>/articles/<slug>/ writes recovered.<lang>.md (each inline figure and each screenshot back as its figure block, each [^n] back as a fact marker after its punctuation), figures/<id>.<lang>.html (each inline figure as published, so check --svg and strip --figures-dir work without a redraw) and facts.<lang>.seed.md (## SOURCES from the reference list, verbatim; ## FACTS one line per citation, the sentence that carried it). The second language takes --pair <the first's recovered draft> --facts <its seed>, so both drafts carry the same source and fact ids section by section. Its guard proves every figure against its block and strips the draft with the seed: figure and image lines must come back byte for byte, every other line once citations are removed; it then reports how many citations the current citer keeps. The seed is a starting point: merge it into facts.md and verify each fact again.
• A docs page: write (one call) to a draft, strip to packages/docs/content/<slug>.<lang>.md; a new page is also added to src/lib/nav.ts, which the docs package's content test keeps in sync.
• Landing strings: write with --profile landing; each key: text line is copied verbatim into strings-en.ts or the zh dictionary in strings.ts.

12. Verification

Markdown is skipped by format:check; the text is covered by check, by the landing package's blog tests (content integrity, anchors, links) or the docs package's content tests, and by reading. A post PR runs the landing tests, format:check and node .agents/skills/penguin-harness-dev/reference/ai-slop/scan.mjs packages/landing/content/blog. A post is finished when, on the test host, those pass with git diff --check; check --in <post> --draft <draft> --facts facts.md --pair <other post> and pair report 0 errors, check --svg passes, both reference lists are within the profile's sources cap and the last review scores m of m joins; and the built landing, driven in 8 combinations (en and zh, light and dark, 1280 and 390 px wide), shows every citation as [n] or [n, m] with no two groups touching and none at a line start, jumps from every [n] to its note below the sticky nav and back with no new tab or reload, opens the external-link dialog for a reference link, never scrolls sideways, hides every .fig-steps, holds exactly one bold span in the lead and in each section, carries no 「夹具」 on a zh page and logs no console or page error. A change to this skill runs selftest and format:check on the test host, git diff --check, and --dry-run of transition and write. When the author asks for the suite on another machine, penguin-harness-dev/reference/verification.md has the remote runner and its traps. After changing the script, a contract, a profile's settings or the check patterns, run node .agents/skills/penguin-harness-writing/scripts/prose.mjs selftest (no network): it runs the local steps on built-in fixtures, and the API steps' guards against a stand-in gateway on a loopback port with a dummy key. A case that guards a fixed bug goes there.

13. Files