Create an actually
useful assistant.

0 1 · B U S I N E S S · F A C T S

Start with what is true.

Give it what you sell, who buys, claims you can prove, examples of good work and the outcome you need next. Put confirmed facts in Business Brief (MD). Keep your role and personal working instructions in Assistant Profile.

Make the files readable first. Download and unzip the starter, then grant your assistant access to that folder. In a web chat, upload the relevant files or paste their contents. Ask it to name the files it can actually read before using the build prompt.

Open the assistant prompt, then work through Build My Assistant. Need the workspace first? See installation options. Creating files does not configure accounts, connect your calendar or email, or activate scheduled work. Those need their own setup in your app.

0 2 · Y O U R · R E A L · W E E K

Show the work. Explain the friction.

Start with a recent week, approved sent messages, finished work and requests you receive. For deeper analysis, up to four weeks of calendar and 30 days of messages can reveal patterns. These are optional sample sizes. Record coverage and gaps; share only permitted material with sensitive details removed.

Map your week in half-days: meetings, making things, follow-ups and interruptions. Add the invisible work the tools missed. Ask: What did I plan to do but never get to? What breaks when I am away?

Keep the evidence visible. Tag what you told it [said], what it inspected [observed], an unconfirmed inference [assumed], and missing information [GAP]. Keep assumptions together for correction. An inaccessible calendar is a gap, not a licence to invent a week.

0 3 · T A S K S · & · P E R M I S S I O N S

Make each job runnable.

A task register is a list of real jobs with enough detail to do them. For each, record the trigger, source inputs, numbered steps, tool, output and destination, reviewer, frequency, time cost, quality example, evidence and permission. Include only tasks that exist. There is no row quota.

You set the permission for each task.

0. Never. Human only.
1. Prepare. Draft the work; release needs authority if not already given.
2. Complete and report. Verify the authorised action, then tell you.
3. Complete quietly. Keep evidence; report agreed exceptions.

Use level 1 for new or unapproved tasks. Preserve existing task-specific authorisation within its recorded scope; ask for a permission decision only to broaden that scope. Record the authority, limits and never-do list beside each task. A permission level does not grant tool access or create a schedule.

0 4 · V O I C E · & · S T A N D A R D S

Give it examples, then corrections.

Build towards three real, sanitised examples of your writing; one is enough to start. Explain the audience, purpose and what makes each work: length, answer order, bullets or prose, greeting, sign-off and words you avoid. Pair a weak draft with your correction so the standard is visible.

These examples are fictional. Replace them with your own evidence. Record the attribution policy that applies to your work. Do not imply that you wrote, approved or sent something when you did not. Delivery follows the task's actual authorisation.

0 5 · Y O U R · O P E R A T I N G · P R O M P T

Turn the brief into instructions.

Have your assistant compile the corrected profile into one personal operating prompt, under 1,500 words. Include your role, business outcome, working preferences, actual tasks and their permissions, voice rules and never-do list. Read it aloud. It should sound like instructions you would give a capable colleague.

You are my assistant for [role and business]. My current outcome is [specific outcome].

Use my Business Brief and Assistant Profile as context. Verify that you can read them. If access is missing, name the gap and ask for the files. Do not invent facts. Distinguish what I said, what you observed and what remains assumed.

Work with me like this: [answer order, level of detail, how to challenge me and how to deliver bad news].

My first three tasks are:
1. When [precise trigger], use [inputs and tool] to create [output at destination]. Stop at [permission boundary].
2. When [precise trigger], use [inputs and tool] to create [output at destination]. Stop at [permission boundary].
3. When [precise trigger], use [inputs and tool] to create [output at destination]. Stop at [permission boundary].

Follow the steps and quality examples in my task register. Match [voice rules and approved examples] and follow [applicable attribution policy]. Never [specific prohibited actions or data access]. Escalate [specific condition] to [approved reviewer].

Use level 1 for new or unapproved tasks. Preserve my existing task-specific authorisation within its recorded scope; do not ask me to approve it again. Ask for a permission decision only when a proposed action broadens that scope. Do not send, post, spend or commit beyond my authority. Show assumptions, missing inputs and incomplete work. Ask: “What did I get wrong?” Record my corrections.

Correct the profile, evidence, register and prompt through ordinary feedback. The assistant records your decisions; it cannot approve its own work. Keep the prompt with the profile and supply it in each new working context. Saved text does not install an agent or configure accounts.

0 6 · T H R E E · F I R S T · T A S K S

Give it three useful first jobs.

Complete one useful result, then choose three tasks for a small pilot. Start with fewer if evidence supports fewer. Use level 1 for new or unapproved tasks; retain existing task-specific authority. Pick repeatable work with a clear example of “done”. These draft examples are starting ideas.

1. Prepare your week.

   When you request planning, turn the approved calendar and task register into three proposed priorities with source links. You review the draft and decide what changes.

2. Draft a follow-up.

   When you mark a conversation for follow-up, use that thread and your voice examples to draft the next message in the project folder. You check every claim and send it.

3. Make a meeting useful.

   When you provide meeting notes, draft a recap with decisions, actions and only the owners and dates actually agreed. You resolve the gaps and release it.

Record what ran, your corrections and measured time saved. After seven days, ask what happened, what went wrong and what to change. Successful runs can support a proposal for broader scope; only your authorisation grants it. Existing authority still applies.

0 7 · R E V I E W · & · A U D I T

Check the work behind the brief.

Keep an honest “Where I Cut Corners” section: unconfirmed assumptions, inaccessible tools, vague answers and task steps you still do not know. Use these five checks before calling the assistant ready.

1. Read the gaps first. They should identify what is missing and how it will be resolved. An empty list is not proof of completeness.
2. Sample three task rows. Check all if there are fewer. Look for an exact trigger, named inputs, steps, output and stopping point. Return vague rows for correction.
3. Check every permission. It cites your authority and limits. Level 1 is the default for new or unapproved tasks. Honour existing scope; seek a decision only to broaden it.
4. Check the never-do list. It should reflect actual data limits, approval needs and relationships requiring human handling.
5. Read the prompt and inspect an output. Does it describe your real work, standards and limits, and produce a verified result? Ask: “What did I get wrong?”

Log the reviewer, corrections and next review. Revisit the profile after two weeks, then monthly if useful; a written date is not a scheduled reminder. For a team, try it yourself first, then with people whose work you can check. Keep the rollout small enough to review.