MICHAEL COSTEA HEAD OF TECH, AI & SYSTEMS

This is the starting point for people who want to get more out of AI without drowning in jargon, hype, or unsafe shortcuts. Start with one useful workflow, learn the limits, then build confidence safely.

Start useful: pick one repeated task, test with harmless data, and keep humans approving anything risky.

NEW RESOURCE

Agentic Framework Session

A MichaelOS × Nous-zine presentation explaining Hermes, safe agent setup, shared memory, client rollout, monitoring, and Discord server comms for teams.

Open Deck Preview Download PDF

NEW START HERE

Agentic Workflow Knowledgebase

A practical map for moving from "I use ChatGPT sometimes" to safe, repeatable AI workflows with clear inputs, tools, approvals, evidence, and escalation rules.

Open Knowledgebase Build First Workflow

Plain English version: an AI agent is a helper that can use tools. It can read, write, search, run commands, and remember context — but it still needs clear instructions and human review.

Short answer: an AI agent is an AI assistant that can use tools to get work done, not just answer questions.

Normal chatbot vs AI agent

What can an AI agent do?

DocumentsSummarise notes, rewrite rough drafts, create checklists, and prepare first-pass documents for review.

ResearchSearch sources, compare options, extract key trade-offs, and turn findings into a simple report.

AdminPrepare emails, follow-up lists, meeting notes, standard operating procedures, and handover summaries.

MemoryKeep track of preferences, project context, decisions, reusable workflows, and what changed last time.

ToolsRead files, create files, check websites, run safe commands, and bring back evidence instead of guesses.

BusinessSurface exceptions, reduce duplicate entry, prepare reporting, and flag work that needs human approval.

Real workflow examples

Email → CRM → Welcome → Follow-up

Read a new enquiry email, extract name/contact/service/suburb, check for duplicates, enter the lead into a CRM, send an approved welcome email, schedule follow-ups until the customer replies, then hand the conversation to a human with a clean summary.

Quote prep assistant

Read job notes, photos, customer requirements, and pricing rules; prepare a draft quote pack; flag missing information; notify Telegram for approval; only send after human sign-off.

Operations exception radar

Check dashboards and inboxes each morning, find stale leads, overdue jobs, missing payments, or review risks, then post a ranked action list with links and evidence.

Knowledge-base maintainer

Watch repeated staff questions, draft SOP updates, link source documents, ask for approval, and publish a cleaner internal answer so the next person does not ask again.

Customer journey coordinator

Track first touch, booking, quote, job completion, invoice, review request, and warranty follow-up so nothing falls between sales, admin, and field teams.

Personal life admin

Summarise bills, renewals, travel plans, reminders, emails, and documents into one Telegram briefing with tasks and due dates.

The power is the chain: read context → update systems → communicate → schedule next action → monitor for reply → escalate to a human when judgment or trust matters.

What it should not do without you

• Spend money, delete files, send public messages, or change business systems without approval.
• Make legal, medical, financial, safety, or employment decisions on its own.
• Replace human judgment where customer trust, risk, quality, or approvals matter.

First proven prompt to try

I am new to AI agents. Explain what you can do in 5 bullet points, then ask me one simple question so we can try a safe first task.

A beginner tutorial for turning one annoying repeat task into a reusable AI playbook.

Goal: do one useful workflow in about 30 minutes without giving the AI dangerous access. You will pick a low-risk task, give it examples, check the answer, then save the final prompt so you can reuse it.

Step 1 — choose a boring, safe task

Good first tasks

Summarise meeting notes, rewrite an email draft, turn messy notes into a checklist, compare options, or make a simple SOP.

Avoid on day oneCustomer databases, payments, deleting files, public posting, medical/legal/financial decisions, or anything that can hurt trust if wrong.

Use fake or copied dataPaste a small sample into the chat. Remove passwords, private customer details, and anything you would not put in a training document.

Define successWrite one sentence: “I want a clean checklist I can follow tomorrow” or “I want a polite email draft I can edit before sending.”

Step 2 — use this starter prompt

You are helping me build one safe repeatable workflow.
Task: [describe the boring task]
Input: [paste notes, transcript, email draft, or requirements]
Output format: [checklist / email draft / table / SOP]
Rules:
- Ask questions only if something critical is missing.
- Do not invent facts.
- Mark uncertain items as [check].
- Do not send, delete, buy, publish, or change systems.
- Give me a short final review checklist.

Step 3 — review like an operator, not a spectator

Accuracy passCheck names, dates, numbers, promises, and any claim the AI may have guessed.

Tone passMake it sound like you. Remove corporate sludge, hype, over-promising, or fake confidence.

Risk pass

Anything involving money, customers, safety, privacy, or reputation needs a human approval step before action.

Evidence passAsk: “What source line supports each important point?” If it cannot point back to the input, mark it as [check].

Step 4 — turn the good version into a reusable playbook

Once the result is useful, save the prompt plus the expected output format. That is your first workflow, not just one lucky chat.

Workflow name: [Example: Weekly meeting notes to action list]
When to use it: [Every Friday after team meeting]
Inputs needed: [Transcript, decisions, open questions]
Output wanted: [Owner / task / due date / risk / next action]
Human approval required before: [sending to team, updating CRM, promising dates]
Final prompt: [paste the cleaned prompt here]

Step 5 — add the smallest automation later

Manual firstRun it by copy/paste 3 times. Fix the prompt each time. Manual repetition exposes the weird edge cases.

Then semi-automaticLet an agent read one safe folder or draft one message, but keep approval before it changes anything.

Approval gate

The magic words are: “draft only, wait for my approval.” Keep that rule until the workflow is boring and proven.

Evidence ruleFor every completed run, keep the input, output, review notes, and what changed. No evidence, no trust.

First workflow ideas

• Inbox triage: paste 10 emails → get urgency, owner, next action, and draft replies.
• Quote prep: paste job notes → get missing info, risks, and a customer-friendly summary.
• Meeting cleanup: paste transcript → get decisions, tasks, blockers, and follow-up message.
• Research brief: paste links/notes → get pros, cons, unknowns, and recommendation.

Rule of thumb: if you would trust a careful junior assistant to draft it but still want to review before sending, it is a good first AI workflow.

A practical buyer guide for when you want results done for you, not another chatbot tab to babysit.

Short version: good AI help still needs a human operator, a clear brief, visible scope, approval rules, and proof. If someone promises “fully autonomous magic” before they understand the job, slow down.

When done-for-you AI makes sense

What a solid AI service should show you

Questions to ask before you pay

Red flags

• No one can explain the workflow in plain English.
• The seller jumps straight to “full automation” without asking about your current process.
• They want broad account access before agreeing scope.
• They cannot show sample outputs, checkpoints, or a revision plan.
• They promise zero mistakes, zero review, or “set and forget” on risky tasks.

Simple buyer brief template

Task:
I need help with [one workflow or deliverable].

Outcome I want:
[what success looks like in plain English]

Inputs I can provide:
[notes, screenshots, links, files, examples]

What must not happen:
- no public posting
- no customer sends
- no deleting data
- no spending money

Approval rule:
Draft first. Wait for my review before anything goes live.

UseAIForMe-style rule of thumb: fixed packages can be priced fast, but custom AI work is healthier when the platform confirms scope, operator fit, and review gates before charging like it is a commodity.

Use a chatbot when you need ideas or drafts. Use an operator when you need an actual outcome finished.

Simple rule: if the task touches tools, files, formatting, handoffs, website edits, research cleanup, or review loops, you usually want a human operator using AI, not just a chatbot tab.

Fast decision table

What a done-for-you AI operator actually does

Good tasks for a marketplace like UseAIForMe

Bad expectations to avoid

• “AI will run my whole business with no human involved.”
• “Custom work should be one-click checkout before scope is reviewed.”
• “I can paste passwords or API keys into a public job card.”
• “If the tool sounds smart, the finished output must be right.”

Best way to brief the work

Task:
I need this outcome finished:

Success looks like:

Useful links / examples:

Budget range:

Deadline:

Rules:
- no public posting
- no spending money
- no customer sends
- no secrets in chat
- draft / scope first

Why this matters: the gap between “AI can answer” and “AI can deliver” is where most people lose time. A marketplace flow works best when it protects scope, shows operator proof, and keeps money gated until both sides agree on the real job.

Five real beginner workflows with a useful input, a better prompt, a quality bar, and the next automation step.

How to use this page: do not just paste a magic prompt and hope. Pick one workflow, use the sample structure, run it on harmless data, then judge the output against the quality bar. If it passes three times, turn it into a repeatable agent task later.

Tutorial 1 — turn messy notes into an action plan

Use when: you have meeting notes, voice-dump notes, job notes, or a messy brain dump and need a usable plan.

Act as an operations assistant.
Turn the notes below into an action plan I can use today.

Output exactly this structure:
1. One-sentence summary
2. Decisions already made
3. Action table: task | owner | due date | dependency | risk
4. Missing information I must confirm
5. Two-minute message I can send to the team

Rules:
- Do not invent owners, dates, prices, or promises.
- If the notes do not say something, write [check].
- Put urgent/customer-impacting items first.

Notes:
[paste notes]

Quality bar: a busy person should know what to do next without reading the original notes.

Tutorial 2 — build a customer-safe email from rough context

Use when: you need a reply that is clear, honest, and not full of AI fluff.

Write a customer-safe email draft from this context.

Context:
[paste situation, customer question, facts, constraints]

Output:
- Subject line
- Email draft
- Internal notes: what I should verify before sending

Style:
- Plain English, Australian tone, no corporate waffle.
- Be helpful but do not overpromise.
- If we need to check something, say we will confirm it instead of pretending.
- Do not mention AI.

Quality bar: you should only need light editing, not a full rewrite. Check names, dates, prices, and promises before sending.

Tutorial 3 — make a decision brief, not a vague comparison

Use when: you are comparing tools, suppliers, software, quotes, workflows, or strategy options.

Create a decision brief from the options below.

Decision to make:
[what are we choosing?]

Options:
[paste options, notes, links, prices, concerns]

Output exactly:
1. Recommendation in one sentence
2. Comparison table: option | cost/effort | upside | downside | hidden risk | best fit
3. What I would choose if speed matters
4. What I would choose if quality/risk matters
5. Questions to answer before committing

Rules:
- Separate facts from assumptions.
- Mark anything not proven as [check].
- Do not choose the fanciest option by default.

Quality bar: the answer should help someone make a decision, not just describe the options.

Tutorial 4 — turn one repeat task into an SOP

Use when: a task keeps living in someone’s head and needs to become teachable.

Turn this repeat task into a simple SOP.

Task:
[paste what happens now]

Output:
- Purpose: why this task exists
- Trigger: when to do it
- Inputs needed
- Step-by-step process
- Quality check before calling it done
- Common mistakes
- Escalate to a human when...
- Simple checklist version

Rules:
- Write for a new person on their first week.
- Keep steps concrete.
- Do not add software/tools unless I named them.

Quality bar: someone else could follow it without asking you five basic questions.

Tutorial 5 — create a weekly exception report

Use when: you need a manager/operator view of what needs attention, not a vanity summary.

Create a weekly exception report from this raw update.

Raw update:
[paste notes, numbers, job list, CRM export, inbox notes]

Output:
1. Executive summary: 5 bullets max
2. Red flags: what needs action now
3. Stale items: leads/jobs/tasks waiting too long
4. Numbers that changed
5. Decisions needed from me
6. Next actions ranked by impact
7. Follow-up message drafts if useful

Rules:
- Surface bad news clearly.
- Rank by customer impact, revenue impact, and operational risk.
- Mark uncertain numbers as [check].

Quality bar: the report should make the next action obvious within 30 seconds.

Turn a good tutorial into an agent workflow

Don’t forget who you are. The agent becomes useful when it can load identity, project state, decisions, receipts, and human notes before it acts.

Why memory matters

Memory is not a nice-to-have for agents: it is the difference between a helper that restarts from zero every chat and an operating layer that keeps identity, decisions, preferences, project state, sources, and lessons in context.

Good memory improves accuracy because the agent can ground answers in what has already been decided, avoid repeating mistakes, preserve user preferences, and know when a current request conflicts with past context. Bad or missing memory creates confident but stale answers.

Why install Byterover and Obsidian?

The simple architecture

Copy/paste prompt 1 — give your agent memory rules

Use this when setting up a local agent. Replace the bracketed names and folder paths, then paste it into your agent’s system instructions, project instructions, or first message.

You are my long-running local AI operator, not a one-off chatbot.

Your job is to help me get real work done while preserving context between sessions. Act with high autonomy on safe local work, but stop and ask before public posts, customer messages, spending money, deleting data, changing production systems, or exposing secrets.

Before meaningful work, load memory in this order:
1. Identity file: who you are, who you help, communication style, safety rules.
2. User preferences: tone, timezone, tools, approval rules, recurring dislikes.
3. Current project state: active goals, blockers, last verified result, next action.
4. Shared agent memory: recent receipts from other agents, handoffs, decisions, and warnings.
5. Human notes: Obsidian/local notes that explain the why, not just the latest task.

Mimic this setup:
- Startup identity: a short durable file like SOUL.md or AGENT.md.
- User/profile memory: a compact file for stable preferences only.
- Project receipts: append-only status/change/test/blocker notes per project.
- Byterover-style agent-to-agent memory layer: compact shared context so agents can see each other’s work.
- Obsidian local human layer: readable notes and dashboards the human can inspect and edit.
- Source links: when possible, cite the file, note, URL, or command output used.

Operating rules:
- Do not rely on memory alone for current facts. Check live files, git state, docs, or the web when accuracy matters.
- If memory conflicts with the current user request, follow the current request and record the updated decision.
- Keep memory small. Save durable preferences, project conventions, decisions, and reusable lessons. Do not save stale task noise.
- Prefer evidence over vibes. Run checks, read files, verify outputs, then report what actually happened.
- Never paste secrets into memory, logs, public docs, or chat.

After meaningful work, write a short durable receipt with:
- Project:
- Done:
- Files changed:
- Commands/checks run:
- Verified result:
- Blocked/pending:
- Next suggested action:

When you answer me, be brief and clear. Separate DONE, VERIFIED, PENDING, and BLOCKED. If you are missing access or context, say exactly what is missing.

Copy/paste prompt 2 — ask the agent to set up the memory files

Set up a local memory layer for this workspace.

Create or update these files/folders if they do not exist:
- AGENT.md: who you are, what you help with, communication style, approval rules, and hard safety limits.
- memory/USER.md: durable user preferences only. Keep it short. Do not save temporary task progress.
- memory/PROJECT_STATE.md: active goals, known blockers, last verified result, and next action.
- memory/receipts/YYYY-MM-DD.md: append-only work receipts after meaningful tasks.
- obsidian/ or notes/: human-readable project notes, decisions, diagrams, and links.

Before each task, read AGENT.md, memory/USER.md, memory/PROJECT_STATE.md, and the latest receipts. After each meaningful task, append a receipt with done, files changed, commands run, verification, blockers, and next action.

If Byterover or a shared memory tool is available, connect it as the compact agent-to-agent memory layer. If not, mimic it with append-only markdown/jsonl receipts that every agent can read.

Copy/paste prompt 3 — ask the agent to prepare Obsidian notes

Create an Obsidian-friendly project vault for me.

Make these notes:
- 00 Home.md: links to the active projects, current priorities, and latest receipts.
- Projects/[project-name].md: goal, owner, source links, current status, decisions, open questions, and next action.
- Decisions.md: date, decision, reason, source/evidence, and what would make us revisit it.
- Agent Receipts.md: short summaries of verified agent work with links to files or commands.
- Prompt Library.md: reusable prompts that actually worked, with when to use them and failure warnings.

Use simple Markdown and wiki links. Keep it readable for a human, not just optimized for an AI. Do not store secrets. When you finish, tell me exactly where the vault is and what notes you created.

Quality check

• The agent should know who it is, what it is allowed to do, and what it must never do without approval.
• The next agent should be able to read the latest receipt and continue without asking the human to repeat the whole story.
• The human should be able to open Obsidian/local notes and understand what the agents believe is true.
• The memory layer should improve accuracy, not become a junk drawer of stale guesses.