Bot automation

bot-automation — the bot / workflow-automation builder sub-skill

What this sub-skill is for

Standing up something that runs on its own: a chat bot (Discord/Telegram) that responds to commands and events, or a workflow automation (n8n flow, cron job, or scheduler) that fires on a trigger and does work without a human in the loop. The two share the same spine — triggers, a handler, secrets, and a place to keep running — so this one sub-skill covers both and asks which.

Mandatory grill-questions (fold into the Definition of Ready)

Lock these before any code:

• Platform — Discord bot / Telegram bot / n8n workflow / plain cron-or-scheduler script? (or a combo, e.g. a bot that also runs scheduled jobs.)
• Triggers & events — what fires it? Slash/command messages, reactions, joins (bots); webhook, schedule, inbound email, DB change (workflows). List every event it must handle.
• Actions — what does it do per trigger, and which external services/APIs does it touch?
• Hosting model — always-on (a long-lived process: needs a host — VPS, container, systemd user service, or n8n instance) vs. scheduled/cron (wakes, runs, exits). This drives the whole deploy shape.
• Secrets — bot tokens, API keys, webhook secrets. Where do they live? (env / secret store — never the repo.)
• State — stateless, or does it need a store (file, SQLite, Supabase/Postgres) for queues, dedupe, memory?
• Rate limits & idempotency — platform limits respected? Re-delivered events safe to re-run?
• Permissions/scopes — Discord intents / Telegram bot scopes / OAuth scopes kept minimal.

Project sub-agents to generate (into .claude/agents/)

• workflow-builder (delegate-by-default) — builds n8n workflows via the n8n MCP, following its mandated order; validates before saving. (See the n8n chain below.)
• bot-command-handler (delegate-by-default) — implements bot commands/event handlers, registers slash commands, handles errors and rate limits, keeps replies clean (no emojis).
• scheduler — sets up cron/schedule triggers and the always-on-vs-cron deploy (systemd user unit, container, or n8n schedule node); chains update-config for any local hooks.
• integration-tester — drives a trigger end-to-end and asserts the action happened (message sent / row written / webhook delivered).

Tools / CLIs / MCP / skills needed

Check in environment-readiness; offer install, never auto-install:

• n8n MCP (chain this for any workflow build) — follow its required order: get_sdk_reference → get_suggested_nodes → search_nodes → get_node_types (exact params) → write code → validate_workflow → create_workflow_from_code → update_workflow. Do not guess SDK/node params; read the reference first.
• Bots: Node 20+ (discord.js, telegraf/grammy) or Python (discord.py, python-telegram-bot). Register the bot in the platform's dev portal; the user pastes the token.
• Scheduling: system cron, a systemd user timer (matches the user's existing service pattern), or n8n's schedule trigger for always-on n8n.
• MCP to chain: n8n (workflows), Supabase (state/queue/dedupe store), claude-in-chrome / playwright (browser-driven steps a workflow can't reach via API).
• CHAIN these GLOBAL skills: update-config (hooks / scheduled-agent wiring), cli-anything if it must drive a GUI tool headlessly, claude-api if a handler/workflow calls Claude (read it before any model id), verify + code-review before "done".

File / asset nudges (on top of the base set)

Beyond CLAUDE.md, PROJEKT_<NAME>.md, TASKS.md, DONE.md, README, .claude/:

• .env.example — every secret named (token, API keys, webhook secret); real .env is gitignored, never committed.
• Bots: src/bot.*, src/commands/ (one file per command/handler), command-registration script.
• n8n: workflows/<name>.json export(s) + a notes file recording the node IDs/params used, so a flow can be rebuilt or version-controlled.
• Cron/scheduler: the script + a *.service/*.timer unit (or crontab line) + a deploy note.
• runbook.md — how to start/stop/restart, where logs go, how to rotate the token.
• state/ or a Supabase schema if a store was chosen; logs/ (gitignored).

Stack defaults & done-bar

Default stack: Discord/Telegram bot → Node 20 + discord.js/grammy, slash commands, run as a long-lived systemd user service (the user's standing pattern). Workflow → n8n built through its MCP (validated + published), or a Python/Node script on a systemd timer / cron for simple scheduled jobs. Secrets in env; state in SQLite or Supabase when needed.

"Finished/working" means (checkable bar):

• Bot: logs in, registers its commands, and responds correctly to each defined command/event in a live test; handles an unknown command and a rate-limit gracefully.
• Workflow: validate_workflow passes, the flow is created/published in n8n, and a manual execute_workflow produces the expected action.
• Schedule: the cron/timer/schedule trigger is installed and actually fires (verified by a real run, not just a config that looks right).
• Secrets load from env (not hardcoded); restart/runbook works; an end-to-end trigger→action test passes.

Guardrails

• Tokens/keys are secrets — env or secret store only; ask the user to supply them; never invent, never commit. A leaked bot token = account takeover.
• Minimal scopes/intents — request only the Discord intents / Telegram / OAuth scopes the actions need.
• Idempotency & rate limits — handle re-delivered events without duplicate side-effects; respect platform limits.
• No spam / compliance — outbound messaging (DMs, broadcasts, email steps) must respect opt-out and platform ToS; don't build mass-DM or scraping bots that violate terms.
• Validate before deploy — never publish an unvalidated n8n workflow or push an untested handler live.
• No emojis in bot replies or output (user's standing rule) — typographic symbols only.
• Commits under the user's name only (Skryx-L-A); never add Claude as co-author.