Your first Chief of Staff twin should not be a mystery skill you downloaded from the internet. It should be a piece of plain text you can read, question, copy, and ask your own agent to build locally.
Why start with a Chief of Staff twin?
Most people do not need their first agent to be clever everywhere. They need it to reduce the fog: what came in, what changed, what needs a reply, what needs a decision, and what can safely wait.
That is Chief of Staff work. It is not replacing judgement. It is gathering context, compressing noise, drafting carefully, and asking before anything leaves the room.
What this twin is, and what it is not.
A good first twin is read-heavy, context-aware, and approval-gated. It can prepare a briefing, draft an email, organise a meeting note, or explain what changed in a folder.
It should not send, delete, pay, approve, publish, update live systems, or speak as you without explicit human approval. Preparation first. Automation later.
Give it a safe identity.
If you are just learning, start locally with sample data. When you are ready for real work systems, use a clean permission boundary. A dedicated twin account such as tony.wood.twin@my-domain.com is often safer than overloading your own account.
The account should have MFA, no admin rights, and only the minimum access it needs. In Google Workspace, an email alias is not the same as a separate user account. In Microsoft 365, a shared or delegated mailbox may be a better pattern when the twin mainly supports reading, briefing, and drafting.
The rule is simple: the twin should have enough access to prepare useful work, not enough access to cause damage on day one.
Build the Library before the skills.
The Library is the boring folder that makes the twin useful. It gives the agent names, roles, systems, priorities, policies, and tone before it starts guessing.
Start with people and roles, systems and tools, strategy and priorities, recurring meetings, current projects, policies and boundaries, glossary, tone of voice, and your daily briefing format.
Do not put passwords, secrets, private client material, HR data, medical details, or confidential information in the starter Library. Use safe examples until the access model is approved.
Make a draft first, then a live-read version.
The first useful output should be a draft built from sample or local-safe material. This proves the shape: sources used, facts, assumptions, questions, risks, suggested actions, and draft-only replies.
A good-looking sample report is not a full briefing. The report must say whether it read sample files, copied notes, Gmail, Calendar, Drive, MailerLite, FreeAgent, or something else. It should also name the things it did not read.
Only after the draft is useful should you approve a bounded live-read pass. Keep the scope small, read-only, and visible: the system, the date range, the maximum items, and the stop lines. Preparation first. Automation later.
Copy these into your agent.
These prompts are deliberately visible. Read them first. If you would not be comfortable with the instruction, do not paste it.
The first 10 skills.
These are the useful first skills because they start with reading, briefing, and drafting. They reduce drag without pretending the agent is ready to run the company.
Inbox triage and draft replies. Group what matters and prepare replies in your voice.
File review and document Q&A. Read approved documents and answer with links back to source.
Company, person, and web research. Prepare a grounded brief before a meeting or decision.
Summaries and briefings. Turn scattered material into a short daily or weekly view.
Calendar and meeting prep. Show commitments, clashes, and preparation notes.
Strategy and plan generation. Turn goals, constraints, and evidence into options.
Knowledge capture and memory retrieval. Keep durable context findable and dated.
Tone-aware reply rewriting. Improve drafts without impersonating you unchecked.
Slack or Teams triage. Surface mentions, requests, and blockers from chat.
CRM or contact follow-up. Prepare relationship notes and next steps without polluting live records.
What the first day should feel like.
The first day is not about connecting everything. It is about proving the shape works. Give the twin a safe Library, ask for one draft briefing, review what it used, and improve the instructions.
If it cannot explain what it read, what it did not read, what is uncertain, and what needs your decision, it is not ready for more access.
The second step is a small live-read test, not a full automation. If that live-read report is still useful, then you can decide whether a weekly reminder, a recurring report, or a stable machine such as a home server should run it later.
Check before you trust.
Check the prompt. Check the account. Check the permissions. Check the sources. Check the draft before it leaves. The training is successful when the person stays in charge and the twin makes the work easier to understand.
Official product features change, so always confirm the current account and permission model for your environment before connecting real business data.