AI Agents — Solo Empire

Overview

Use this skill to design and operate AI agents that follow a reliable tool-use loop. Treat skills/ as the agent instruction layer and infra/ as the implementation and tool layer: invoke existing CLI commands, scripts, MCP tools, and tests before adding new mechanisms.

Response language

Use Thai for all human-facing agent responses by default. Keep code, commands, paths, identifiers, logs, protocol fields, and quoted source text exact. Follow another language only when the user explicitly requests it for the response or artifact.

Workflow

1. Establish context

• Read architecture guidance from architecture-docs/docs/ (via SOLO_ARCHITECTURE_ROOT or ../architecture-docs) and operational command ownership from platform-ops-cli/README.md (via SOLO_OPS_ROOT or legacy sibling ../platform-ops-cli). Neither checkout depends on the hub. Use focused rg/sed ranges rather than loading long documents.
• Read the relevant operations/ or integrations/ document when the task concerns a documented platform or external integration.
• Inspect git status --short and preserve unrelated user changes. Do not use destructive Git commands.

For an agent feature, identify the complete loop before editing:

request → agent instructions → tool selection → infra operation → structured result → next decision

2. Route to the existing capability

Follow the tool-selection and context-efficiency policy in skills/ai-agent-workflows/SKILL.md. Keep inputs explicit and bounded, and use structured results rather than scraping human-formatted output when a structured interface exists.

3. Make the change safely

• Keep implementation in its owning repository; do not put runtime code in this skill. Portable skills live in agent-skills.
• Use apply_patch for source and documentation edits.
• Follow the owning service repository's documented backup/restore procedure before schema or data migrations; the retired local SQLite backup command is unavailable.
• Keep credentials in the documented environment or secret store. Never print, copy, or commit values from the owning repository's .env, tokens, cookies, or private data.
• Treat data/, generated reports, backups, and nested repositories as separate boundaries.
• For external services, verify that the required credential is available and report it as unavailable instead of asking the user to paste a secret.

4. Verify proportionally

Run the narrowest relevant checks, then the repository checks required for the change:

• General ops-toolkit changes: task test, task typecheck, and task docs:check from platform-ops-cli.
• Database changes: relevant checks from the owning service repository.
• Platform changes: task test:platform and the relevant validation scripts in bookchaowalit/platform-infra-kit.
• MCP changes: run the checks documented by mcp-aggregator.
• Python automation changes in ops-toolkit: task test:python and the focused test task listed in its Taskfile.

If a check cannot run because an external service, credential, or local daemon is unavailable, record the exact limitation and do not hide the failure.

Decision rules

• Need a new AI-facing capability → implement or extend an MCP tool in mcp-aggregator, then document its selection and safety rules here if it is reusable.
• Need a repeatable deterministic operation → implement or reuse a helper in the owning repository; do not embed it in SKILL.md.
• Need agent judgment, routing, or safety procedure → update this skill.
• Need durable operational policy or a runbook → update the relevant architecture-docs/operations/ boundary, not this skill alone.
• Need a blank reusable document or artifact → copy from architecture-docs/templates/ into its owning workspace and leave the source template unchanged.
• Need to update learning progress → use the existing skills CLI or MCP tool; do not confuse database-tracked learning skills with this AI skill package.

Handoff

Report the files changed, commands and tests run, important outputs or failures, commit/push/deploy evidence, credential/service limitations, and any follow-up task. Follow the repository's automatic release closeout preference for scoped Git and documented non-production deployment actions; retain approval gates for production, destructive, paid, credential, and external-message actions.