Model

• The part doing the thinking. Opus, Sonnet, DeepSeek, take your pick.
• Different models have different pros and cons. Pick the right one per job.

Harness

• A bare model only writes text back. The harness lets it act.
• It reads your files, runs commands, uses your tools.
• Claude Code is the harness. My setup on top is open source.

Agent

• The harness pointed at one job, with its own instructions and limits.
• Each only touches what its job needs. The CRM one never touches your finance files.
• A specialist answers sharper than one general chat.

↳ power move the edge is the instructions you hand-write inside each agent. That is where context engineering pays off.

Skill

• A multi-step task you save once and trigger with a slash.
• One word instead of a paragraph you retype. Teach it once, reuse for good.

↳ power move skills compose. One can call another, so a procedure becomes a building block.

Second brain

• A vault of notes and rules it reads and writes, so it carries what you teach it forward.
• Correct it once and it applies the rule from then on. Day 90 is far better than day 1.

↳ power move a weekly job reads your corrections and infers the pattern, so your taste compiles itself.

Terminal & IDE

• Two front doors to the same agent.
• The bare terminal is the rawest, fastest way in.
• Or run it inside VS Code: your files, its changes, and your work all in view. That is where I demo.

It plugs into
your stack

• MCP: a universal adapter. A tool exposes a server, and every agent speaks to it the same way.
• API: anything with an endpoint, the agent calls directly.

If it has an API or an MCP server, your agent can use it. That is how the live demos touch real systems.

↳ power move wrap any internal API in your own MCP server in an afternoon, and every agent gets it for free.

Say it
out loud

• Talking is ~3x faster than typing, and you think out loud.
• Lowest-effort unlock for non-engineers. No syntax, just speech.
• Cleaned automatically: a glossary fixes names & jargon.

Free and local here. Paid alternative: Wispr Flow, ~€15/mo.

One session
per topic

• A fresh chat is a clean desk. New topic, new session.
• Piling unrelated tasks into one thread poisons the context.
• The single biggest quality win that costs you nothing.

Let it learn
your operation

• Tell it a team rule once and it keeps it.
• It learns your accounts, your deal language, your filing. Day 90 beats day 1.

Run sessions
in parallel

• Don't queue behind one task. Fan the work out.
• One researches while another drafts while a third cleans data.
• The shift: you manage several workers and check back.

↳ power move rename and group your parallel sessions, so you always know which one is doing what.

Batch big topics,
hand context forward

• Cut a large job into chunks, one session per chunk.
• A handover note carries decisions and next steps, so each session starts warm instead of cold.

Comment on the page →
agent revises it

• Review like a designer: click, comment, move on.
• Comments export as JSON; the agent reads and rewrites.
• No copy-pasting feedback into a chat. Tightest loop there is.

↳ power move the same loop runs on contracts, specs, any multi-pass work. Feedback as data.

Schedule repetitive,
recurring tasks

• Same report every Monday, same digest every morning? Don't trigger it by hand.
• A scheduled job runs on a timer and does the work while you sleep.
• You wake up to a finished draft, not an empty to-do.

↳ power move every scheduled job still stages for your yes. Nothing auto-sends unless you allow it.

Persistent parallel
sessions + cheap models

• Named terminals (tmux) keep agents working across turns.
• Route each task to the cheapest model that's good enough.
• A high-volume workflow drops from dollars to cents.

↳ power move route by task type: vision on Opus, summaries on Llama, code on DeepSeek, by cost and latency.

Inbound → done →
you just approve

• Stack it all: dictation, the right agent, memory, parallel work.
• A Cliq / Slack message becomes a finished draft awaiting yes.
• Real, from my own commercial-ops setup, running today.

↳ power move a confidence gate decides: high-confidence posts itself, the rest waits in draft. An eval loop guards the send.

From one user
to a team

• Everything so far was my personal setup. The frontier is sharing it.
• Shared skills: the whole team triggers the same saved moves.
• Shared memory: your accounts, ICP, and deal rules compound across everyone.

↳ power move this is where policy & governance matter most: shared memory means shared responsibility.