跳到正文
cccskills
免费在 GitHub 公开

create-x-content

Draft voice-tuned X (Twitter) post variants from a free-form brief. Reads a personal voice guide (generated via generate-voice-guide), produces 2–5 variants with distinct framings (simple-howto, problem-first, hype, mechanism-breakdown, etc.), and self-checks against the voice guide's banned phrases before returning.

中文简介正在准备中。当前显示原文说明。

集成工具X (Twitter)

实测评分 未实测

查看安装方法

包含的文件(2)

  • SKILL.md6.9 KB
  • skill.meta.json268 B

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Create X Content

Draft X/Twitter post variants that sound like you, not like an AI. Reads a voice guide the user has already generated (or prompts to create one), produces multiple framings of the same idea, and saves each variant as its own markdown file with frontmatter.

This is an agent-executed skill — the agent does the drafting and self-check inline. No Python script.

Quick Start

/create-x-content --brief "New open-source CLI that turns Figma files into React components. Called figma2react. Free, MIT licensed."

Or interactively:

/create-x-content

Inputs

FlagRequiredDefault
--briefYes (asked interactively if missing)—
--variantsNoSkill decides based on brief richness (2–5)
--voice-guideNoResolved via chain below
--outputNo./content/YYYY-MM-DD-<topic-slug>/
--topicNoDerived from brief

Voice Guide Resolution

Resolve in this order, stop at first hit:

  1. --voice-guide <path> flag
  2. ~/.goose-skills/config.json → voice_guides.x
  3. ~/.goose-skills/voice-guides/voice-x.md (default path)
  4. Fallback prompt — no guide found. Give the user three options:
    • (a) Paste a path to an existing guide
    • (b) Run /generate-voice-guide --platforms x now to create one (recommended)
    • (c) Proceed with a neutral default (warn that variants will sound generic)

Never silently skip the voice guide. Generic posts are the failure mode to avoid.

Workflow

Phase 1 — Load the voice guide

Read the resolved voice guide into context. Extract, for use throughout drafting:

  • Banned phrases — treat as hard blocks
  • Hook patterns — sample these for variant framings
  • Format rules — length ranges, line-break style, emoji policy
  • Dos/don'ts — apply to every variant
  • Example posts — use as calibration targets

Phase 2 — Parse the brief and decide variant count

Decide variant count by richness:

  • 2 variants — one-line opinions, single-angle hot takes, simple tool mentions with no mechanism
  • 3 variants — typical tool spotlights, single how-tos, observations with one clear angle
  • 4 variants — multi-angle topics (a how-to + a problem-first frame + a hype frame + a mechanism breakdown)
  • 5 variants — rich, mechanism-heavy content where several distinct hooks all have real substance

If --variants is explicitly set, respect it. Otherwise pick the smallest count where each variant genuinely adds a different angle. Quality > quantity. 2 strong variants beat 5 watered-down ones.

Phase 3 — Generate variants with explicit framing

Each variant gets a distinct framing label that determines its structure and hook:

FramingStructureWhen to use
simple-howtoBare steps, no mechanismTool install + usage, 2–3 line how-to
howto-plus-mechanismHow-to + "here's how it works" breakdownTools where the mechanism is interesting
problem-firstOpen with the pain, then solutionWhen the problem is visceral/relatable
hypePunchy "someone just dropped X" energyLaunches, new OSS releases, cool builds
mechanism-breakdownFocus on the howTechnical builds, systems, architectures
ecosystem-mapCurated list of N related tools/companiesLandscape posts
contrarian"Most people do X wrong" openerOpinion pieces with a clear counter-take
personal-experience"We've been doing X. Here's what I learned"Field notes, lessons learned

Use framings the voice guide's hook patterns actually support. Don't force framings the user never employs.

Phase 4 — Self-check against the voice guide

Before saving, run each variant through these checks:

  1. Banned phrase check — does the variant use any phrase from the voice guide's banned list? If yes, rewrite that line.
  2. Length check — does it fit the voice guide's format rules (short-form <280, long-form 100–1000, etc.)?
  3. "Meat" check — does it have at least one concrete reference: a specific tool, number, workflow, or build? If not, it's filler — rewrite.
  4. Voice-match spot check — would this plausibly appear in the example posts section of the voice guide? If it reads like a generic AI tweet, rewrite.

If a variant fails any check and two rewrite attempts don't fix it, drop the variant rather than ship something weak.

Phase 5 — Save outputs

Write each variant as its own .md file. File naming:

variant-<letter>-<framing-slug>.md

Examples: variant-a-simple-howto.md, variant-b-problem-first.md.

Frontmatter schema:

---
id: <topic-slug>-<letter>
platform: x
format: short | long
topic: <slug>
framing: <framing-slug>
status: draft
---

Body: just the post text. No commentary, no surrounding markdown.

Phase 6 — Deliver

Print to the user:

  • Output directory path
  • List of files created
  • Variant count + framings used
  • Suggested next step (review, tweak, or run /social-kit for a matching graphic)

Outputs

  • <output>/variant-<letter>-<framing>.md per variant
  • No index file — each variant is self-contained

Examples

Simple brief:

/create-x-content --brief "npx goose-skills install claude-code-hooks — a new skill that adds pre-commit, pre-tool-use, and post-response hooks to Claude Code so you can enforce coding standards automatically."

→ 3 variants (howto, hype, mechanism)

Rich brief:

/create-x-content --brief "Built a Claude-driven lead gen system. Scrapes Reddit for people asking about email deliverability, finds their domains, verifies their business email via Hunter, drafts a personalized DM. 47 leads in 6 hours at $0.03/lead."

→ 5 variants (personal-experience, mechanism-breakdown, hype, problem-first, simple-howto)

Dependencies

  • A voice guide at the resolved path (see Voice Guide Resolution above)
  • generate-voice-guide skill (for creating one when missing)

Tips

  • Default to fewer variants. Five weak ones is worse than two strong ones. The skill should feel picky, not generous.
  • Framings must actually differ. Don't write five variants with the same opening hook and minor tweaks. Each variant = a different lens on the same idea.
  • Quote the brief's specifics verbatim in the variant. Numbers, tool names, prices — these are the "meat." Losing them to paraphrasing kills the post.
  • If the voice guide has a "field-notes deep-dive" hook pattern and the brief is rich enough, include one long-form variant. That format tends to outperform on X.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Research supported advertising promises from a specific buyer's context, current product facts, customer language and observed ads. Turn them into static, video or copy ideas; match video ideas to actual recipes and save a product-scoped angle bank for write-video-ad-script. Product-led discovery can use verified facts without reviews. Reference persistence and reach are observations, not proof of performance.

中文简介正在准备中。当前显示原文说明。

gooseworks-ai/goose-skills12642026年10月23日 更新

Analyze ad campaign performance data (Google, Meta, LinkedIn) to identify what's working, what's wasting budget, and specific cut/scale/test recommendations. Runs statistical analysis, funnel diagnostics, and multi-channel budget reallocation with specific dollar-amount shift recommendations and scenario modeling.

中文简介正在准备中。当前显示原文说明。

gooseworks-ai/goose-skills12642026年10月23日 更新

For paid lead-gen and participant-recruitment ads, replaces vanity CPA with true CAC per qualified lead by joining ad-platform data with downstream funnel events, surfaces tracking gaps, and classifies every creative into Scale / Keep / Investigate / Cut.

中文简介正在准备中。当前显示原文说明。

gooseworks-ai/goose-skills12642026年10月23日 更新

The intake stage of a Meta ad harness. Researches the brand first, states what it assumes, and asks only what it cannot infer and what would change money, claims or legal exposure if guessed wrong — at most four questions a turn, in plain business language. Writes the brand's stable facts (ads/brand.md), the campaign brief, and a log of what was asked vs assumed. Use it when someone wants to run Meta ads and the brand's ad facts are not yet recorded.

中文简介正在准备中。当前显示原文说明。

gooseworks-ai/goose-skills12642026年10月23日 更新

Checks paid ads against the ad platform's rules (Meta, TikTok, Google, LinkedIn and others) and the brand's category rules: the copy, on-image or on-screen text, pictures and people. Rewrites what can be fixed honestly, adds the disclosures that are needed, and flags anything that needs a human or legal review, with one pass, fixed or flag line per ad and platform. Use alone when you already have the ad text or images and know where they will run; otherwise run the playbook for the job (e.g. make-static-ads, make-content, email-lifecycle).

中文简介正在准备中。当前显示原文说明。

gooseworks-ai/goose-skills12642026年10月23日 更新

Analyze the message match between your ads and landing pages. Checks if the promise in the ad copy carries through to the landing page headline, body, and CTA. Flags disconnects that kill conversion rates. Works with Google, Meta, and LinkedIn ads.

中文简介正在准备中。当前显示原文说明。

gooseworks-ai/goose-skills12642026年10月23日 更新

查看 gooseworks-ai 的全部技能

报告此技能的问题