AI Costa · Useful AI, at your own pace

Episode 01 · 8

Start with a real conversation

What you will achieve

Open ChatGPT and turn a real work situation into a useful conversation without learning a formula.

Why it matters

The best first experiment is not asking what AI can do. Bring something that already occupies your mind: a decision, a difficult email, or a half-formed idea.

A little help if you get stuck
Speak as you would to a new person on your team. If the answer is not useful, do not start over: point out what is missing and continue.

Try it now

Episode 01

I am working on [describe the real situation in two sentences].

Before suggesting anything, ask me three questions, one at a time, to understand the outcome I need and the boundaries we must not cross.

When you have my answers, summarise the problem in plain language and suggest one small first step I can review today.

Episode 02 · 8

Think and work by voice

What you will achieve

Use voice to develop an idea, add context and correct course without having to write everything.

Why it matters

Speaking releases detail that often disappears when we type in a hurry. Voice is especially useful for sorting out a situation, rehearsing a message, or discussing work while looking at it.

A little help if you get stuck
Use voice chat for a live exchange; use dictation when you want to review the text before sending it. Availability can vary by account and workspace.

Two ways to work by voice

Talk with ChatGPT. Dictate with Háblame.

ChatGPT voice

You have a conversation with AI: think out loud, receive questions and correct course. The conversation uses the ChatGPT service.

Háblame, on your computer

Hold a key, speak and the text appears wherever your cursor is. Dictation audio is processed on your PC or Mac. If you later send that text to ChatGPT, the text does leave your computer.

I built Háblame with AI assistance without writing the code from scratch. I started from Handy, an open-source project under the MIT licence. I decided the name, privacy, behaviour and tests; AI helped plan, transform and review. The lesson is not “AI makes apps”: speaking, deciding and checking are also ways of building.

Read how I built Háblame → See Háblame ↗

Try it now

Episode 02

I want to think out loud with you about [topic].

Listen first. Do not give me a solution until I say “that is all”. Then return:
1. what seems decided;
2. what remains open;
3. the most useful question to answer next.

If I contradict myself, point it out calmly and quote the two conflicting ideas.

Episode 03 · 8

Assign work instead of asking for magic

What you will achieve

Give AI an outcome, sources, boundaries and a clear definition of “done”.

Why it matters

“Make something beautiful” forces the system to guess. A strong assignment says which truth it may use, what it must deliver, what is forbidden and how the work will be checked.

A little help if you get stuck
You do not need a long document. For important work, include context, objective, deliverable, boundaries and verification.

Try it now

Episode 03

Context: [what is happening and who it is for].
Source of truth: use only [the document, notes or data I attach].
Outcome: I need [specific result].
Deliverable: give me [a format I can review].
Boundaries: do not invent facts, and do not send, publish, buy or change anything without my approval.
Verification: finish with a list showing what you checked and what remains open.

Episode 04 · 8

Work with documents and office files

What you will achieve

Turn real files into a document, spreadsheet, presentation or PDF that you can review.

Why it matters

AI contributes more when it starts with your material and ends with a useful object. The final file matters as much as the answer in the chat.

A little help if you get stuck
Only attach material you are allowed to share. State the file type, structure and review criteria. Check names, figures, formulas and pages before using the result.

Try it now

Episode 04

Use only the file I attach.

Create a one-page executive summary for someone new to the topic. Include: situation, three findings, decisions needed and next steps.

Keep names and figures exactly as written. If two facts conflict, do not choose one: create a “To verify” section with both and their locations.

Return it in an editable format and check for cropped text or blank pages.

Episode 05 · 8

Create something people can see

What you will achieve

Move from an idea to an image or visual piece and improve it through concrete instructions.

Why it matters

The first version helps you discover what you actually wanted. Describe purpose, audience, composition and restrictions; then correct one thing at a time.

A little help if you get stuck
Do not invent logos, awards, testimonials or customer faces. If an image represents a real product, compare the details before publishing it.

Try it now

Episode 05

Create a horizontal image that explains [idea] to [audience].

Show [main element] in [setting], with a clear composition and clean space for a short title. The mood is [three concrete adjectives].

Do not include logos, small text, brands, awards or recognisable people.

Before generating, summarise the composition you will create in one sentence.

Episode 06 · 8

Use Codex for real work

What you will achieve

Give Codex a folder or project, request a concrete outcome and require proof of what it did.

Why it matters

Codex is not only for programmers. It can investigate a project, organise information and build deliverables within the permissions you give it.

A little help if you get stuck
Start with reversible work. Say what it may read, what it may change and which real action needs your approval. Never paste passwords or secrets into a chat or file.

Try it now

Episode 06

Work in this project to achieve [outcome].

First find the existing rules and files. Tell me what you will use as the source of truth.

You may read and prepare changes inside the project. You may not send, delete, buy, deploy or publish without asking me.

When finished, run the appropriate checks and return: changed files, test results and any real limitation.

Episode 07 · 8

Make good work repeatable

What you will achieve

Understand skills, hooks, MCP and automations as four simple ways to repeat, watch or connect work.

Why it matters

A skill stores a way of working; a hook checks something at a specific moment; MCP connects tools and context; an automation runs an assignment again when the time comes.

A little help if you get stuck
Do not start with the technology. Choose a task you have already done well twice. Save it, test a normal case and an exception, and only then consider automating it.

Try it now

Episode 07

Turn the work we just approved into a reusable procedure.

Keep its objective, sources, steps, boundaries, final test and situations where you must stop and ask me.

Add two test cases: one normal and one with an essential fact missing.

Then tell me, without jargon, whether this should be saved as an instruction, run on a schedule, connected to another tool or given an automatic check. Do not configure anything yet.

Episode 08 · 8

The SELLO Method: finish with evidence

What you will achieve

Separate what looks right from what has been checked, and reserve real-world actions for explicit approval.

Why it matters

Giora’s SELLO Method turns the end of a job into a series of gates: define the outcome, do the work, check it, ask permission before a real action, and then verify the public result.

A little help if you get stuck
A draft exists. Reviewed work has received a critical look. It is only finished when every criterion is met and the evidence is in front of you.

DraftA first version exists.

ReviewedSomeone has examined it.

FinishedEvidence and approvals complete.

Try it now

Episode 08

Apply the SELLO Method to this job: [describe the job].

1. Define the outcome and checks with me before acting.
2. Do only the reversible work that is authorised.
3. Check the result with visible evidence.
4. If the next step sends, publishes, buys, books or changes something real, stop and ask for my explicit approval.
5. After that action, verify the real result again.

Finally label the status DRAFT, REVIEWED or FINISHED. Do not use FINISHED if evidence or approval is missing.

Do not study AI. Put it to work with you.

One minute: AI actually working

Episode guide

Start with a real conversation

What you will achieve

Why it matters

Try it now

Think and work by voice

What you will achieve

Why it matters

Two ways to work by voice

Try it now

Assign work instead of asking for magic

What you will achieve

Why it matters

Try it now

Work with documents and office files

What you will achieve

Why it matters

Try it now

Create something people can see

What you will achieve

Why it matters

Try it now

Use Codex for real work

What you will achieve

Why it matters

Try it now

Make good work repeatable

What you will achieve

Why it matters

Try it now

The SELLO Method: finish with evidence

What you will achieve

Why it matters

Try it now

AI does not finish the work. You decide when it is finished.