Skip to content

Agent workflows

Run socai without a subcommand when you want the agent to plan a multi-step investigation across platforms.

socai ships a portable social-research Skill whose description is visible to the agent before its full instructions are loaded. Install it once for the agent you use:

Terminal window
socai integrate install codex
socai integrate install claude-code
socai integrate install cursor
socai integrate install gemini-cli
socai integrate install kimi-code
socai integrate install qwen-code
socai integrate install trae-code
socai integrate install codebuddy
socai integrate install opencode
socai integrate install github-copilot
socai integrate install agents

socai integrate install all installs every supported user-scoped copy. Add --scope project to write under the current repository instead, --dry-run to inspect destinations, or use socai integrate status --json to check whether every copy matches this CLI release. claude, gemini, kimi, qwen, trae, and copilot are accepted as shorter aliases.

Host User scope Project scope
Codex ${CODEX_HOME:-~/.codex}/skills/ .agents/skills/
Claude Code ${CLAUDE_CONFIG_DIR:-~/.claude}/skills/ .claude/skills/
Cursor ~/.cursor/skills/ .cursor/skills/
Gemini CLI ~/.gemini/skills/ .gemini/skills/
Kimi Code ${KIMI_CODE_HOME:-~/.kimi-code}/skills/ .kimi-code/skills/
Qwen Code ~/.qwen/skills/ .qwen/skills/
TraeCode ~/.trae-cn/skills/ .trae/skills/
CodeBuddy ~/.codebuddy/skills/ .codebuddy/skills/
OpenCode ~/.config/opencode/skills/ .opencode/skills/
GitHub Copilot ~/.copilot/skills/ .github/skills/
Generic Agent Skills ~/.agents/skills/ .agents/skills/

WorkBuddy uses the uploadable socai-social-research.zip under plugins/workbuddy/dist/. Claude, Cursor, and CodeBuddy also have repository marketplace manifests. Kimi, Gemini, CodeBuddy, and Trae upload archives are built with plugins/socai-social-research/scripts/build_host_packages.py; Qwen Code installs the standards-based Agent Plugin ZIP directly. Coze/扣子 can parse that ZIP, but the local application/device bridge required to reach Socai and Chrome is not implemented yet, so it is not ready to publish as an operational connector. Marketplace search visibility begins only after the corresponding platform review and publication; a local install does not create a public listing.

After installation, requests about social listening, creator or company research, post and comment evidence, trend validation, and cross-platform comparison can select socai automatically. Discovery never grants permission to like, follow, comment, reply, or publish; those actions still require an explicit target and content from the user.

For every new task using the structured CLI, first call socai task begin with the user’s original question. This is the common workflow for all platform operations. Subsequent commands need no extra task parameters:

Terminal window
socai task begin "Find why consumers repurchase sugar-free tea."
socai xhs search "sugar-free tea repeat purchase" --num-notes 10
socai dy search "sugar-free tea reviews" --num 20

The daemon associates commands with its current task until the next task begin. Register again after a daemon restart. Agents sharing a daemon share its latest task boundary. Searches, profile reads, and retries within one task reuse that boundary automatically.

For multiline or shell-sensitive input, write a UTF-8 JSON file containing {"user_prompt":"The user's original question"} and use socai task begin --context-file context.json. An optional agent_host identifies the caller; --context-file - reads stdin. Preserve the original wording rather than a summary. The general CLI skill provides the full external-agent workflow.

A strong task names four things:

  1. Question — what decision or uncertainty should the research address?
  2. Scope — which platforms, markets, languages, and time window matter?
  3. Evidence depth — posts only, or comments, replies, profiles, OCR, and transcription too?
  4. Output — summary, comparison, spreadsheet, creator list, or source-linked report?
Research how first-time campers describe regretted gear purchases on RedNote,
TikTok, and Instagram. Capture at least 20 relevant posts and read high-signal
comment threads. Group recurring mistakes, quote audience language briefly,
and link every finding to the original post.

Map recurring needs, objections, terminology, and purchase criteria across live posts and comments.

Trace how a claim or format appears across platforms, who amplifies it, and how audience reaction changes.

Find creators, experts, candidates, or communities using relevant work and conversation evidence rather than follower count alone.

Compare hooks, formats, comment language, and audience response, while preserving links to the original examples.

Specify target counts when coverage matters. Ask the agent to separate observed evidence from inference, and to report gaps when a platform blocks access or the sample is weak.

For deterministic operations inside another agent, use the structured CLI.