# Using tweet-factory with Claude Code

**Last updated 2026-10-11.** “Claude Code Twitter” here means using the Claude Code CLI as the model that drafts posts. tweet-factory’s default provider is `claude-code`. It does not send your likes to a tweet-factory server, and it does not create the post on X.

## What the process runs

`src/llm/claudecode.ts` spawns:

```
claude -p <prompt> --output-format json --tools "" --disallowedTools "*,mcp__*" --permission-mode dontAsk
```

No API key is read on that path. The command uses the `claude` login already on the machine. `--tools ""` removes built-in tools, so a tweet, article, or Notion page inside the prompt cannot make it run a shell. If `claude` is missing, the error tells you to install Claude Code or set `TF_LLM_PROVIDER` and a key. The timeout is 240 seconds. `TF_LLM_MODEL` is passed as `claude --model` when you set it.

The prompt is `voiceSystem()` from `persona/PERSONA.md` and measured stats, plus `OPTIMIZER_SYSTEM` built from `src/rank/weights.ts`. Scraped text is fenced as reference data. The model is told to return JSON and to leave scores at 0. After it returns, `src/rank/score.ts` fills the 0–100 number. See [the tweet generator](ai-tweet-writer-your-voice.md) and [how drafts are scored](how-tweet-factory-scores-drafts.md).

## A run that stays on your machine

```bash
tf auth
tf fetch
tf onboard
tf generate
```

`tf generate` writes `drafts/YYYY-MM-DD.md` and `onboarding.json`. Default shape: four ideas, two or three variants each. `--n` changes the idea count. `--json` prints the bundle. You publish the variant you want. OAuth scopes in `src/x/auth.ts` are `tweet.read`, `users.read`, `like.read`, `bookmark.read`, and `offline.access`. There is no write scope.

Claude Code has to be logged in on that machine, including a server you cron from. `tf` does not ship a Claude credential.

## The optional Claude Code skill

The repo includes `.claude/skills/tweet-factory/SKILL.md`, triggered as `/tweet-factory` inside Claude Code. That skill calls `tf` for fetch, flush, and data, then drafts. It is optional. `tf generate` does not load it. An extra skill at `~/.claude/skills/twitter-algorithm-optimizer/` can replace the playbook when a person runs the skill. The binary still scores with `src/rank/weights.ts`.

Direct APIs, if you would rather not use Claude Code, are `TF_LLM_PROVIDER=anthropic`, `openrouter`, or `openai`. Temperature is 0.8 and the token cap is 8,000. Details are on [LLM providers](llm.md).

## Related

- [Tweet generator in your voice](ai-tweet-writer-your-voice.md)
- [Install](install.md)
- [LLM providers](llm.md)

## FAQ

### Can Claude Code write tweets?

Yes, as tweet-factory’s default backend. `tf generate` runs `claude -p` with your existing Claude Code login and with tools disabled. The output is a local markdown file. You post it.

### Does Claude Code post to Twitter or X for me?

No. The CLI’s X scopes are read-only. Claude Code is the language model, not an X client. Publishing stays outside `tf`.

### Do I need an Anthropic API key?

Not for the default path. The default uses the `claude` CLI login. Set `ANTHROPIC_API_KEY` only if you switch `TF_LLM_PROVIDER` to `anthropic`.