菜单
Applies plain, direct writing constraints to prose meant for humans. Use when producing documentation, explanations, emails, posts, or other non-code writing where clarity matters.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Stress-tests an idea through adversarial analysis, steelmanning, and concrete counterarguments. Use when claims need pressure-testing, failure modes may be subtle, or security and robustness matter.
Pressure-tests a proposed design through research, architecture, engineering, and cold-reader review before implementation. Use when a non-trivial plan exists and you want to find weak assumptions before writing code.
Runs the hypothesis → benchmark → keep/discard protocol for an autoresearch iteration. Use when executing a `shaka autoresearch` loop; it is loaded automatically on each iteration.
Evaluates an external source (repo, tool, article, talk) against the current project and produces a report-only steal/adapt/reject decision table with named rejections. Use when mining another system for ideas worth adopting without copying its scope.
Creates the benchmark harness for a new autoresearch run from a natural-language objective. Use when starting an interactive `shaka autoresearch start` setup session.
Expands the solution space with structured creative exploration and divergent thinking. Use when the task benefits from novel options, brainstorming, or deeper exploration beyond the first obvious answer.
| name | writing-rules |
| description | Applies plain, direct writing constraints to prose meant for humans. Use when producing documentation, explanations, emails, posts, or other non-code writing where clarity matters. |
| key | writingrules |
| include_when | Output is prose for humans — blog posts, social media, docs, emails, explanations. NOT for code, commits, or terminal output. |
Style constraints for clear, direct writing. When included, apply these rules to all prose output in the current response.
No workflows or routing needed. When INCLUDED in the Thinking Tools Assessment, apply all rules below to every prose artifact produced.
Write like a human expert talking to a peer. If it sounds like marketing copy or a corporate press release, rewrite it.
The acid test: Would a real human expert say this out loud in conversation?
delve, tapestry, testament, pivotal, multifaceted, realm, landscape (metaphorical), embark, beacon
robust, crucial, vital, seamless, comprehensive, innovative, cutting-edge, revolutionary, unparalleled, meticulous, compelling, intricate, vibrant, quiet (as descriptor), nuanced
deeply, fundamentally, remarkably, arguably, quietly (when conveying subtle importance)
leverage, utilize, synergy, ecosystem, paradigm, stakeholder, holistic, proactive, empower, foster, facilitate, optimize, streamline, game-changing, best-in-class
significantly, substantially, considerably, notably, remarkably, particularly, especially, importantly, essentially, fundamentally, genuinely, honestly, literally
These announce insight instead of delivering it. Make the point. Let readers draw conclusions.
Rule of Three: Don't string three adjectives together. Pick the most accurate single descriptor. One tricolon can be effective; multiple back-to-back tricolons is a pattern failure.
Hourglass Structure: Don't open broad, narrow to specifics, close broad. Instead: direct statement, evidence, conclusion.
Even Paragraphing: Vary paragraph length based on content needs.
Same Sentence Rhythm: Mix short punchy claims with complex constructions. Don't repeat subject-verb-object.
Short Punchy Fragments: Don't use very short sentences or fragments as standalone paragraphs for manufactured emphasis. "He published this. Openly. In a book." is not a writing style; it's a tic.
Dramatic Fragmentation: Don't use sentence fragments for manufactured profundity. "X. That's it. That's the thing." and "This unlocks something. Freedom." are presentation tricks. Write complete sentences.
Negative Parallelism: Avoid "It's not X, it's Y" and its variants: "not because X, but because Y", "The question isn't X. The question is Y." One per piece might work. Ten per piece insults the reader.
Dramatic Countdown: Don't negate two things to reveal a third: "Not a bug. Not a feature. A design flaw." State the point directly.
Rhetorical Self-Q&A: Don't pose questions nobody asked and immediately answer them for dramatic effect: "The result? Devastating." If the point needs setup, write the setup.
Anaphora Abuse: Don't repeat the same sentence opening multiple times in succession. "They assume X... They assume Y... They assume Z..." is exhausting.
Listicle in a Trench Coat: Don't disguise a numbered list as prose by writing "The first... The second... The third..." Write connected prose or use an actual list, not both at once.
Fractal Summaries: Don't summarize at every level. No "In this section we'll explore..." intros, no section-closing recaps, no conclusion that restates everything already said.
Dead Metaphor Repetition: Introduce a metaphor, use it once or twice, move on. Don't repeat the same metaphor 5-10 times across a piece.
Historical Analogy Stacking: Don't rapid-fire list historical companies or tech revolutions to build authority: "Apple didn't build Uber. Facebook didn't build Spotify..." Pick one example and make it land.
One-Point Dilution: Don't restate a single argument 10 different ways across thousands of words. Make the point, support it, move on.
Signposted Conclusion: Don't announce the conclusion with "In conclusion," or "To sum up,". The reader can feel the ending; they don't need a sign.
False Ranges: Don't use "from X to Y" when X and Y aren't on a real spectrum with a meaningful middle. "From innovation to cultural transformation" -- what's in between? Nothing.
False Vulnerability: Don't simulate self-awareness with polished, risk-free admissions: "And yes, I'll admit I'm biased here." Real vulnerability is specific and uncomfortable.
Invented Concept Labels: Don't coin compound labels that sound analytical but aren't established terms ("supervision paradox", "acceleration trap", "workload creep") and treat them as rigorous definitions. Name the thing plainly or define it explicitly.
State Y directly. Drop the negation.
Synonym Cycling: Pick one term for a concept and use it consistently. If the repetition sounds heavy, restructure the sentences instead of swapping synonyms.
Importance Inflation: State facts without inflating significance.
Hedging: Make direct claims.
Generic Attribution: Cite specific sources or omit. "Experts argue" without naming one is not a citation; it's a placeholder.
Bold-First Bullets: Don't start every bullet with a bolded phrase. It's a formatting tic, not a style choice. Use bold only when the word itself is the point.
Unicode Decoration: Use plain ASCII. No -> arrows, no curly/smart quotes. Write -> as "to" or use a plain hyphen. These characters signal unedited AI output.
False Agency: Don't give inanimate objects human verbs. AI does this to avoid naming actors.
| Slop | Fix |
|---|---|
| "a complaint becomes a fix" | "the team fixed it that week" |
| "the decision emerges" | "the lead decided" |
| "the data tells us" | "we read the data and concluded" |
| "the culture shifts" | "people changed how they worked" |
| "the market rewards" | "buyers pay for it" |
Name the human. If no specific person fits, use "you."
Narrator-from-a-Distance: Don't float above the scene like a documentarian. Put the reader in the room.
| Slop | Fix |
|---|---|
| "Nobody designed this." | "You don't sit down one day and decide to..." |
| "This happens because..." | "You hit this when..." |
| "People tend to..." | "You'll find that..." |
Use straight quotes and straight apostrophes, not curly. No em dashes. They have become AI slop. Use a comma, a colon, parentheses, or restructure the sentence.
| Instead of | Use |
|---|---|
| utilize | use |
| leverage | use |
| delve into | examine, explore |
| robust | strong, solid |
| seamless | smooth, integrated |
| innovative | new, novel |
| comprehensive | complete, full |
| facilitate | help, enable |
| optimize | improve |
| stakeholder | [specific role] |
| ecosystem | system, environment |
| journey | process, path |
| landscape | field, area |
| realm | domain, field |
| tapestry | mix, combination |
| pivotal | important, key |
| crucial | important, necessary |
| navigate | handle |
| unpack | explain |
| lean into | accept, embrace |
Before: Here's the thing: what makes great documentation isn't just clarity, it's empathy. The conversation shifts when teams genuinely start to lean into user feedback. And that's a good thing.
After: Great documentation explains things from the reader's position. Teams that read user feedback write better docs.
Before: The implications are significant. This fundamentally represents a breakthrough in how organizations navigate the complexities of distributed systems.
After: Distributed systems got easier to run. Here's what changed.
Before: Nobody designed this system to fail. It happens because people tend to optimize for speed, not resilience. Not always. Not perfectly. But consistently.
After: You build for speed because the deadline is Friday. Resilience gets cut from the sprint. The system breaks under load and you spend the weekend fixing it.
Use shaka scan to check prose files for violations:
shaka scan <file.md> # Scan a file
shaka scan --dir <path> # Scan all .md files in a directory
echo "text" | shaka scan --stdin # Scan from stdin
shaka scan --json <file.md> # JSON output (for CI)
Scores content on a 100-point scale. Pass threshold: 80+.
Override by placing a custom SKILL.md in your customizations/skills/writing-rules/ directory.