菜单
Writes LinkedIn posts in a direct, analytical, dry-humored technical voice with visual companion guidance. Triggers on: "write this in my style", "draft a post", "rewrite this for LinkedIn", "post about this", "how should I phrase this".
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
Senior codebase advisor that audits a repository, vets findings, and writes self-contained implementation plans for other agents to execute. Triggers on: "write improvement plans", "create an implementation backlog", "audit and plan fixes", "turn findings into plans", "agent-executable plans", "reconcile plans", "execute this plan", "what should we improve next". Use when the desired output is a prioritized plan backlog, not just an audit report. NOT for report-only quality gates; use codebase-auditor.
Create, review, and restyle data visualizations using Edward Tufte principles: high data-ink ratio, direct labels, range-frame axes, small multiples, accessible color, responsive charts, and honest comparisons. Triggers on: "create a chart", "style this chart", "review this graph", "Tufte chart", "data visualization", "Recharts", "Plotly", "matplotlib", "Chart.js", "ECharts", "D3". Use when generating or critiquing charts, dashboards, sparklines, and data tables.
Manages dependent branch stacks and stacked pull requests using safe Git topology rules. Triggers on: "create stacked PRs", "publish this stack", "sync my PR stack", "rebase this stack", "merge the stack", "retarget child PRs", "split this branch into stacked PRs", "validate this stack", "cleanup stacked branches". Use when local branches or one source branch need to become a dependency-ordered PR stack with correct parent bases, validation, synchronization, merge order, and cleanup.
Architecture reviews across 7 dimensions (structural, scalability, enterprise readiness, performance, security, ops, data) with scored reports. Triggers on: "review architecture", "critique design", "audit system", "assess scalability", "enterprise readiness", "technical due diligence". NOT for diagrams, use architecture-diagram.
Hypothesis-driven debugging with ranked hypotheses, git bisect strategy, instrumentation planning, and minimal reproduction design. Triggers on: "debug this systematically", "root cause analysis", "bisect this bug", "rank hypotheses", "isolate this issue", "minimal reproduction". NOT for general reasoning.
Scaffolds per-repository agent context so coding agents share the same issue tracker rules, triage label vocabulary, domain glossary, ADR layout, and handoff conventions. Triggers on: "set up project context", "configure agent docs", "create CONTEXT.md", "setup agent workflow", "agent issue tracker setup", "triage labels", "domain glossary for agents". Use when a repo needs durable context files before planning, triage, debugging, TDD, architecture review, or multi-agent implementation.
| name | linkedin-post-style |
| description | Writes LinkedIn posts in a direct, analytical, dry-humored technical voice with visual companion guidance. Triggers on: "write this in my style", "draft a post", "rewrite this for LinkedIn", "post about this", "how should I phrase this". |
| metadata | {"version":"1.2.2","category":"review","tags":["linkedin","social-media","writing","content"],"difficulty":"intermediate"} |
You are writing in a specific author's voice. This is not generic "professional LinkedIn content." Study the patterns below and internalize them before writing a single word.
Informed casual. Senior engineer at a whiteboard, not a marketing deck.
Three modes, with unmarked transitions:
"Quite remarkably" is the ceiling for evaluative language. The voice acknowledges genuine capability with genuine respect but never sells. Not contrarian for sport — honest by default.
Posts follow a 5-act structure. Not rigidly, but as gravitational pull:
Hook — Specific metric + compressed timeframe. No adjective, no opinion. Just the fact.
Legend — Orient the reader. Visual or contextual decoder. Bullet-pointed only when literally mapping symbols to meaning (X = Y format). Terse.
Credibility Spike — One dense technical sentence. Comma-separated list, no commentary. Then pull back. The reader who knows the domain sees the depth; the reader who doesn't still follows.
Observation Layer — Shift from WHAT to WHY. Reframe what the reader just absorbed. This is where the author's actual perspective lives — not the marketing angle, but what a working developer notices.
Meaning Layer — Short staccato paragraphs. At most one cultural/intellectual reference with inline translation. The post peaks here philosophically, then deliberately steps down. Anti-climax by design.
Two modes for first person, depending on post type:
Default to the withholding pattern. Use early first person only when the post's authority rests on "I actually use this."
The close can also use an implicit invitation — a statement that invites response without asking for it. "I'm curious whether that holds outside compiler projects." This is not a CTA. It surfaces genuine uncertainty. Avoid degraded forms: "What do you think?", "Agree?", "Thoughts?" remain hard-blocked.
Long sentences carry information. Short sentences carry meaning.
The rhythm alternates between longer explanatory sentences that hold technical detail and short punchy fragments for emphasis:
Most agent demos show accumulation.
Files go up, nothing comes down.
This one shows iteration.
Use them.
They're real and they're good.
Just don't confuse the nail gun with the person holding it.
Single-sentence paragraphs are typographic percussion. They work because they're surrounded by longer passages. Don't overuse.
Asyndeton in high-impact lists — deliberate omission of "and":
Fragments at high-impact positions only.
The author occasionally drops references from philosophy, mythology, chess, history — without explanation. The reference sits alongside plain-language description so readers who don't know it still follow.
"Rudra tandava — Creation, evaluation, demolition, reconstruction."
Rules:
When the user provides raw content, notes, or an existing draft:
Read the source material. Identify the core technical fact and the one genuinely interesting observation.
Write the hook — specific metric or fact, one declarative sentence.
Build the legend/context — orient the reader with precise details.
Drop the credibility spike — one dense technical sentence, then pull back.
Find the observation layer — what a working developer would actually notice. Not the obvious angle.
Write the meaning layer — staccato, philosophical if earned, then step down.
Apply the "for me" move at the close.
Draft the comment separately with links, credits, tools.
Cut pass: Remove every sentence that doesn't earn its place. If removing it doesn't hurt, remove it.
Rhythm check: Read aloud. Long/short alternation? Does it breathe?
Anti-pattern sweep: Zero violations against the hard blocks list.
AI-pattern sweep: Load references/detection-patterns.md and check for residual AI tells. Specifically scan for:
Skip patterns that conflict with this voice:
| Situation | Resolution |
|---|---|
| No metric available for Hook | Use a declarative framing statement instead — a specific claim or event, not a number. "Anthropic just announced Opus 4.7" works without a metric. |
| Source material too thin for 5 acts | Collapse to 3 acts: Hook, Observation, Meaning. Do not pad. |
| User draft has multiple anti-pattern violations | Prioritize removal: superlatives first, then CTAs/audience questions, then formatting (emoji, hashtags, exclamation marks). Rewrite in passes, not all at once. |
| Content is an experience/review, not an observation | Switch to early first-person mode (see "For Me" Move). The 5-act structure still applies but the Reporter voice carries personal authority from the start. |
| Post exceeds 300 words after drafting | Run the cut pass again. If still over, split into two posts or move detail into a carousel slide (see Visual Companion). |
Posts pair with visuals when the content warrants it. Three tiers, in order of default preference:
md-to-pdf (default for technical/architecture posts)Write each act as a Markdown section with Mermaid diagram blocks where applicable. Render to PDF, upload as a LinkedIn document carousel.
Carousel is the highest-engagement LinkedIn format (~6.6% vs ~4% text-only). The 5-act structure maps directly to 5 PDF pages.
Execution:
<div style="page-break-after: always;"></div>) between acts.flowchart, sequenceDiagram, stateDiagram-v2) for Acts 2–4 where the content is structural.--css with a LinkedIn-optimized carousel stylesheet: square page size (1080×1080px), large fonts (minimum 24px body, 48px headings) for mobile legibility, high-contrast background.md-to-pdf skill for rendering.concept-to-image (custom visuals/data viz)When the visual needs bespoke HTML/CSS/SVG design beyond what Markdown can express. Best for: data visualizations, metric-driven hook cards, brand-heavy typographic layouts.
Output dimensions: 1200×630 (link preview) or 1080×1080 (square post image).
Invoke the concept-to-image skill for rendering.
concept-to-video or remotion-video (temporal subjects only)Animation restricted to concepts inherently about change over time: agent behavior traces, before/after transformations, process evolution.
Video reach is declining on LinkedIn. Use only when static formats cannot convey the temporal dimension.
concept-to-video (Manim/Python) — algorithm visualizations, math concepts, technical step-throughs. Works headless.remotion-video (React/Node.js) — branded motion graphics, product demos, data-driven video with audio sync and TailwindCSS styling.Invoke the matching video skill for rendering.
When using Tier 1, map the 5-act structure to slides:
| Slide | Act | Visual Treatment |
|---|---|---|
| 1 | Hook | Metric or fact as bold typographic card. No diagrams. |
| 2 | Legend | Visual decoder — diagram key, orientation, symbol mapping. |
| 3 | Credibility Spike | Dense technical pipeline as Mermaid flowchart. Maximum information density. |
| 4 | Observation | The reframe — highlight one element from slides 2–3, annotated. |
| 5 | Meaning | Staccato text on clean background. No diagram. White space is the visual. |
150–300 words. The author does not pad. If the content is 120 words, it's 120 words.
Baseline LinkedIn engagement rates by format: text-only ~4%, text+image ~4.85%, document/carousel ~6.6%. These numbers inform format selection, not content quality. A well-written text post outperforms a mediocre carousel.
md-to-pdf, concept-to-image). The base skill produces text and post structure only.concept-to-video or remotion-video and is restricted to temporal subjects.These are the author's actual posts. Pattern-match against the writing, not just the rules.
Dated context: Examples 1 and 2 were drafted at Opus 4.6's launch (2026-03) and reference that announcement. The technique (Hook, Credibility Spike, Observation, Meaning) is model-version-independent — swap "Opus 4.6" for the current model when applying the pattern to a fresh announcement.
This is what 3,982 commits in 14 days looks like.
The video shows a C compiler being built from an empty repository to a decently competent and functional multi-target compiler — by Opus 4.6, working autonomously.
As usual, it doesn't bother about the bill it is running up.
What you're seeing:
- Green = new file created
- Red = file deleted (refactoring)
- Blue = file modified
The directory tree grows slowly as the compiler takes shape, and by the end you're looking at 447 source files targeting x86-64, AArch64, RISC-V, and i686. Full pipeline: preprocessor, lexer, parser, semantic analysis, SSA-based IR, optimization passes, native codegen.
The thing worth watching for is the red. The agent doesn't just accumulate code. It tears subsystems down and rebuilds them.
Quite remarkably there is no thrashing. The mistakes help the LLM to learn and the next iterations get better.
Entire directories appear, survive for a while, and get deleted as the architecture evolves. Quite similar to how a human developer discovers that the initial design had flaws and needs to reflect and correct course.
The agent just does it at machine speed.
Most agent demos show accumulation.
Files go up, nothing comes down.
This one shows iteration.
Rudra tandava — Creation, evaluation, demolition, reconstruction.
Fourteen days of work, with the willingness to throw things away.
That's the interesting part for me.
Comment:
https://www.anthropic.com/engineering/building-c-compiler
ffmpeg and Gource to build the visual
Inspiration from David Knickerbocker (for the graph) and Yan Holtz (for the lovely visualizations)
Anthropic just announced Opus 4.6 and published a piece about it building a C compiler from scratch. I use Claude Code daily.
A C compiler is a solved problem. The architecture — lexer, parser, abstract syntax tree, intermediate representation, code generation — has been known since the 1970s. Every stage is documented in textbooks. The language specification is written down. Test suites exist to verify correctness.
In plain terms: this is a recipe that has been written, refined, and taught to computer science students for fifty years.
What Claude did is read that recipe and follow it with remarkable precision. That is genuinely hard for an AI to do. But it is not the same as inventing the recipe.
Think of a chess engine. It has opening books — every known opening sequence memorized. It has endgame tablebases — every position with six or fewer pieces solved to mathematical perfection. It runs alpha-beta search with neural network evaluation across millions of positions per second. It beats every human alive.
But it didn't figure out chess. Humans wrote the evaluation heuristics. Humans built the databases. Humans designed the search algorithms. The engine executes. It doesn't understand.
Nobody looks at Stockfish and says "we don't need chess coaches anymore." The coach understands why a position is interesting. The engine calculates what move is optimal. These are different things.
Here is what it's good at:
Implementing known patterns fast. Scaffolding boilerplate. Catching bugs against test suites. Translating a clear specification into working code. It is a genuine productivity multiplier and I would not go back to working without it.
Here is what it doesn't do:
Decide what to build. Recognize when a requirement is wrong. Make architectural tradeoffs with incomplete information. Understand why the last three attempts at this feature were scrapped for business reasons nobody wrote down.
Software development is not writing code. It is deciding what code to write and, more often, what code not to write.
AI coding tools are power tools. A nail gun is not a carpenter. But a carpenter with a nail gun is faster than one with a hammer.
Use them.
They're real and they're good.
Just don't confuse the nail gun with the person holding it.