THE RESEARCH ATELIER

A better answer starts with a better check.

Work through one small task. Keep the reasoning, question the output and decide what happens next.

3 curated examples · no account · no uploads · no live model calls

01 / INSPECT

One task. Five visible decisions.

These are authored teaching examples, not measured client cases. Your selections stay in page memory. Nothing is sent to an AI provider.

GOAL

Turn a small set of agreed facts into a decision-ready brief.

Source

Fictional source note: a four-person team needs an internal guide. The deadline is Friday. The owner has not confirmed the audience or approval process.

1 / 5

What a source-check ledger records

Claim → original source and DOI → exact passage → quotation or paraphrase → inference and limits → human decision → disclosure of actual assistance. A quotation needs exact words and location; this example uses paraphrases, not a purported verbatim quote.

02 / PRACTISE

Start where your work is.

A transparent suggestion from two choices, not a test of aptitude. Change it whenever you want.

Your experienceDesired output

SUGGESTED START

Beginner · understand & check

Rule: new or unsure starts with verification. Tool experience starts with a repeatable method. Only an existing repeatable workflow plus an automation goal suggests advanced planning.

Prerequisite: no AI experience needed; use the supplied fictional example.
Identify one unsupported claim before trying any tool.
Explain which sentence needs a source and which decision belongs to a person.

Choose a different depth

03 / TAKE AWAY

Useful after the conversation.

An inspectable format for a written plan, reusable instruction and checking pack. Sample content is fictional; your actual scope is agreed before paid work.

A useful work brief

Goal: Turn a small set of agreed facts into a decision-ready brief.
Input rules: Use fictional or approved non-confidential notes. No company document is uploaded in this demonstration.
Reusable instruction: Use only these facts. Write: goal, known constraints, missing decisions and next action. Mark anything unconfirmed. Do not invent a budget, audience or approval.
Quality check: Compare every sentence with the note. Keep audience and approver as open questions. The project owner decides both before work starts.
Next three actions: Choose a safe input. Run once. Compare and record one correction.
Owner & stuck-case fallback: The person responsible for the task. Stop for missing sources or permissions; return to the source or ask the owner.

Look inside a first session and its takeaways

Illustrative agenda, adapted by agreement to your task. This is a planning example, not a guarantee that a project is solved in one session.

15 min · goal, existing approach and data boundaries
20 min · map one workflow and a useful baseline
30 min · work through and challenge one example
15 min · repeat the check with less guidance
10 min · next three actions and an owner

Bring an anonymised or synthetic example and the rules that apply. The agreed written plan records priorities and checks; ongoing implementation or unlimited support is not included unless separately agreed.

Inspect a sample plan and checklist

Fictional team assessment: a four-person guide project

Project owner: Sets the goal and approves the final guide.
Writer: Uses approved notes; labels assumptions.
Reviewer: Checks each factual claim against its source.
Tool/data owner: Approves access, retention and provider settings.

Map tasks and permissions, not employee ability. External sharing: none in this sample. In a real engagement, every proposed tool and transfer is listed before use. Missing permission means manual work.

04 / SET LIMITS

Should this become an agent?

An agent can take steps with tools. More capability creates more responsibility. Check only what is actually established; unknown stays unchecked.

Six feasibility gatesPermission: each data source and action is explicitly authorised.Consequence: a failure cannot create an unreviewed high-impact decision.Supervision: a person reviews outputs before any external action.Reversibility: changes have a tested undo or safe stop.Dependencies: tools, data access and failure fallbacks are understood.Ownership: a named person will maintain and recheck the workflow.

Do not automate yet.

Resolve the unchecked gates or keep a manual checklist. This is a safety boundary, not a recommendation to buy more tools.

05 / KEEP LEARNING

Explain it before you rely on it.

Optional private self-check · not scored or collectedI can explain source, instruction, check and human decision.I can spot one unsupported claim and show why it fails.I know what data must not enter the tool and when to stop.

If a point is unclear, revisit that step with another safe example. This is reflection, not a qualification, outcome promise or request for a testimonial.

A small, source-checked shelf

UNESCO · Recommendation on the Ethics of AI

Human oversight, proportionality and responsibility. A principles framework, not a compliance certificate.

OpenAI · Data controls in ChatGPT

Check the settings and policy for the exact account you use. A setting is not permission to upload confidential data.

Wang et al. (2023) · Scientific discovery in the age of artificial intelligence

A review article for practising source scope and claim verification. DOI: 10.1038/s41586-023-06221-2.

Links and teaching scope checked: 2026-10-08. Next review due: 2027-01-08. Editorial responsibility: AI Methods Studio. Check original policies before use; account plans and features can change.

Practical education only. No legal advice, compliance certification or live AI service.

Individual starting format: €90 · standard €150

90 minutes plus a written plan. Existing launch terms and private feedback arrangement apply. Team scope is assessed first; no work is purchased here.

Scope, payment and cancellation terms

Rates reviewed: 2026-10-08

Discuss your task