Skip to content

ideas.md

Latest commit

 

History

History
72 lines (57 loc) · 3.76 KB

ideas.md

File metadata and controls

72 lines (57 loc) · 3.76 KB
Raw

Ideas to extend your agent

You have two starter skills (capture-note, recall) and one example persona (researcher). Here's a menu of next steps. Pick what fits the agent you're building — don't try to do all of it. Each is a self-contained afternoon's work, roughly ordered easiest-first within each section.

More skills (offline — no extra accounts needed)

  • daily-journal — append a timestamped entry to today's note. The "what did I do today" habit, owned by your agent.
  • weekly-review — read every note from the last 7 days and write a summary note. Great demo: the agent reflecting on its own memory.
  • summarise-file — point it at a PDF / Markdown / text file and get a tight summary saved to memory.
  • todo — capture, list, and complete tasks stored as Markdown checkboxes (works in Obsidian/Logseq too). A mini task manager.
  • tag-cleanup — scan memory/ and suggest merging near-duplicate tags (#ai vs #a.i.). Teaches the agent to curate its own brain.
  • link-suggester — when saving a note, find existing notes it should [[link]] to.

More skills (with an MCP server or API — a bit more setup)

  • web-research — fan out a few web searches, read the top results, write a cited brief. Pairs perfectly with the researcher persona. (Add a web-search MCP server.)
  • calendar — read your week, find free slots, draft (never send) invites.
  • inbox-triage — classify email into reply / read-later / ignore. Drafts only.
  • github-watch — summarise new issues or PRs on a repo you care about.
  • expense-snap — read a photo of a receipt and log the amount to memory.

More personas

A persona is a focused mindset your agent (or a subagent) can adopt. Copy .opencode/agents/researcher.md and rewrite it. Good ones to try:

  • planner — breaks a fuzzy goal into ordered, concrete steps. Doesn't do the work, designs it.
  • critic — pokes holes. "What's the weakest assumption here? What would make this fail?" Run it over your own plans.
  • explainer — rewrites anything in plain language for a specific audience (a 12-year old, a busy exec, a sceptical engineer).
  • devils-advocate — argues the opposite case so you stress-test a decision.
  • a domain voice — a gardening coach, a recipe planner, a study buddy. Make it about something — specific agents are more fun to demo than generic ones.

MCP connections (giving your agent hands)

MCP (Model Context Protocol) is how your agent reaches the outside world — Notion, Slack, GitHub, a database, your filesystem. Add servers under "mcp" in opencode.json. The Notion one is already stubbed there as an example. Browse servers at https://github.com/modelcontextprotocol/servers and OpenCode's MCP docs.

Ideas: Notion (notes/tasks), Slack (read & draft messages), GitHub (issues/PRs), a SQLite/Postgres database, a maps or weather server, a web-search server.

Stretch goals (if you're flying)

  • Subagent orchestration — have your primary agent delegate to several personas (researcher → critic → writer) for one task.
  • A scheduled run — a "good morning" briefing the agent writes to memory each day.
  • Semantic recall — upgrade recall from keyword search to embeddings so it finds notes by meaning, not just matching words.
  • Voice or a tiny web UI — put a friendlier face on the terminal agent.

Demo tips

  • A working small thing beats an impressive broken thing. Get one loop solid: capture → recall, demoed live, is genuinely compelling.
  • Show the memory/ folder open in Obsidian beside the terminal — "the agent and I share one brain" lands every time.
  • Have a real, personal use case. Agents are more memorable when they obviously help you with your actual life.