Research: Existing Tools and Building Blocks

Crosspost (@humanwhocodes/crosspost) is almost exactly what we need for the publishing layer. It's a single Node.js package that works as a CLI, a programmatic API, and an MCP server — meaning Claude Code can post directly to X, Bluesky, LinkedIn, Mastodon, Discord, Telegram, and more within a session. No browser, no credential juggling. For polishing, there's no off-the-shelf skill, but that's fine — we'd build custom Claude Code skills with platform-specific prompts and tone guidance. Claude itself is the polishing engine. The LangChain Social Media Agent is interesting as a reference but is more heavyweight than we need.

Top pick: Crosspost (@humanwhocodes/crosspost)

A Node.js package that is a CLI tool, programmatic API, and MCP server in one. Supports X/Twitter, Bluesky, Mastodon, LinkedIn, Discord, Telegram, Dev.to, and Nostr. The MCP integration means Claude Code can publish directly within a session. This is the single best fit for our "agent workflow" approach.

• Repo: https://github.com/humanwhocodes/crosspost
• npm: @humanwhocodes/crosspost

Multi-platform publishing tools

Platform-specific libraries (if we need fine-grained control)

• X/Twitter: tweepy (Python) — requires Basic plan ($100/mo) for write access
• Bluesky: atproto (Python) — full AT Protocol SDK, free
• LinkedIn: linkedin-api-python-client (Python) — official SDK

AI agent / content workflows

• LangChain Social Media Agent — LangGraph-based agent that generates Twitter/LinkedIn posts from URLs, with human-in-the-loop approval. TypeScript. Repo: https://github.com/langchain-ai/social-media-agent
• Claude Code custom skills — no existing social media polishing skill found, but the SKILL.md infrastructure is purpose-built for this. We'd create our own polishing skills with platform-specific prompts, character limits, and tone guidance.

MCP servers for social platforms

Recommended stack

1. Polishing: Custom Claude Code skills (SKILL.md) — Claude itself is the polishing engine, guided by platform-specific prompts and tone rules
2. Publishing: Crosspost as MCP server + CLI — one tool for all platforms, native agent integration
3. Scheduling (if needed later): Late CLI or Postiz API

First Prototype Validation

The first prototype is intentionally lightweight:

• one top-level fast social skill
• local context docs for stream, tone, and channels
• loose markdown examples as test inputs

The question for this validation pass is not "is it complete?" but "does this shape feel fast enough to use?"

Example: feature-note.md

What worked:

• the source is thin but still enough to infer a Flowershow or Datopian publishing angle
• the "less friction" angle is clear and easy to turn into short social copy
• this is a good example of the kind of real input the workflow should handle

What felt weak:

• without reading the PR, the skill would need to stay generic
• product identity is a little ambiguous, so this is a likely case for one short clarifying question

What to tighten:

• make it explicit in the skill that when product identity is ambiguous, it should choose the safer framing instead of pausing too long

Example: dataset-update.md

What worked:

• this is ideal MVP input because the product, audience, and value are easy to infer
• the "publishing data well matters" angle is stronger than a generic release note
• should produce good X and LinkedIn drafts quickly

What felt weak:

• the input does not include any concrete dataset detail, so posts may risk sounding generic

What to tighten:

• encourage one specific concrete detail where available, but do not require it

Example: thought-leadership-note.md

What worked:

• the note is strong enough to produce a clean, opinion-led post
• this tests whether the skill can choose one angle without turning into a workshop

What felt weak:

• this kind of input can easily trigger too much ideation and too many alternatives

What to tighten:

• bias the skill toward choosing one strong framing and drafting immediately

Example: video-link.md

What worked:

• the desired angle is already well constrained: one practical takeaway rather than full summary
• this is a good pattern for YouTube or podcast-derived posts

What felt weak:

• if the linked content cannot be inspected, the workflow depends entirely on the short context note

What to tighten:

• tell the skill to ask for one takeaway sentence only if the link context is too thin

Validation Summary

The current prototype direction looks viable for the MVP.

Strengths:

• the workflow is lightweight enough to stay inside one thread
• the skill can work from rough markdown rather than a structured brief
• the context pack removes the need for repeated tone and product questions

Risks:

• generic outputs when the source lacks one concrete detail
• over-questioning or over-ideating on thought-leadership inputs
• publishing still depends on external tooling that is not yet integrated here

Most important next refinement:

• tighten the orchestrator skill so it defaults to drafting faster and explaining less

Publishing Decision For The Next Iteration

The smallest publish path worth prototyping next is:

• keep the skill producing copy-ready output by default
• separately test Crosspost as the first real publishing integration

Recommendation:

• do not block the MVP on full publish-in-thread support
• test the polish workflow first
• if the polish flow feels genuinely fast, add one publish helper using Crosspost rather than building custom platform code

Main blocker to true publish-in-thread behavior:

• credential setup and tool configuration, not content generation