AI Agent Guide

Use plain-language prompts to let Codex, Claude Code, and other local agents operate UniPost through the CLI for safe read and write workflows.

Plain-language agent operation

Tell the agent what you want. Let the CLI do the precise work.

After agent setup, you do not need to memorize every UniPost command. Ask Claude Code or Codex in normal language, tell it to use UniPost CLI, and set the boundary: read-only by default, safe drafts before live publishing.

Default rule

Start with read and summarize. For writes, ask for the plan first. Live publish, schedule, cancel, and retry need explicit approval.

--yes + --idempotency-key

Set up the agent context

Finish CLI auth first from CLI Overview, then install the client-specific agent instructions. The bootstrap commands ground the agent in the active workspace, profiles, accounts, and safe next actions.

A reusable prompt pattern

Profiles, accounts, and recent posts

Read workspace context

Give the agent enough real UniPost context before it explains, plans, or writes code.

Tell the agent

Expected CLI

The agent should summarize profile IDs, account IDs, platforms, connection status, and recent post status in normal language.

Health, capabilities, and metrics

Inspect account readiness

Ask the agent which connected account is ready for a specific social workflow.

Tell the agent

Expected CLI

The agent should call out disconnected accounts, missing platform permissions, unsupported media types, and useful account-level metrics.

Validate, dry-run, and draft

Prepare posts safely

Let the agent turn a plain-language idea into a safe draft without publishing externally.

Tell the agent

Expected CLI

The agent should return the draft ID, validation result, suggested caption changes, and any missing inputs.

Explicit write confirmation

Publish only after approval

Keep live writes deliberate when the user has already reviewed the plan.

Tell the agent

Expected CLI

The agent should restate the target before publishing, then report final status and any platform result IDs.

Safety boundary

Operation	Agent rule
Read profiles, accounts, posts, analytics	Safe. Ask in plain language and tell the agent to summarize.
Validate, dry-run, draft	Safe preparation. These actions should not publish externally.
Create live post, schedule, publish draft	Requires explicit user approval, `--yes`, and `--idempotency-key`.
Cancel or retry	Requires explicit user approval, target IDs, and `--yes`.
Default instruction	Start with read-only; do not publish, schedule, cancel, retry, or mutate data unless the user explicitly confirms.