Skip to content
Antoine Le LouëtIndependent senior developer (Home)
Back to the Journal

Ideas6 min read

AI: how can you be more ambitious in your requests?

Felix Rieseberg's four forms of ambition, followed by a two-conversation method for preparing a precise request with verification criteria.

Updated on

Four axes of ambition with a model: speed, altitude, depth and breadth.
The four axes proposed by Felix Rieseberg. Original Journal diagram based on his article of 13 October 2026.

What should you ask after a useful first response?

You ask an AI to fix a button. It proposes a coherent change. Yet the form remains difficult to use: errors arrive too late and some fields request information the customer does not have. The original request left much of the problem outside its scope.

In 'How to be more ambitious: A tutorial for humans with models', published on 13 October 2026, Felix Rieseberg proposes four ways to broaden requests. An Engineering Manager at Anthropic, he says he has led engineering for Claude.ai, Cowork and Claude Code Desktop. Here are the principles from his article, followed by an application I suggest for a web project.

Felix Rieseberg's four forms of ambition

  • Speed: do the same work in less time. This is the familiar use on which the author spends little time.
  • Altitude: move from the task up to the intended outcome, then give the model the means to pursue it.
  • Depth: work with raw data and the discipline's methods, challenge assumptions and verify the result independently.
  • Breadth: borrow concrete techniques from other disciplines, avoiding mere role-play.

His advice on validation deserves to be retained verbatim: 'Do not ask the model to judge itself.'

What is missing from 'You are an expert in…'?

'You are a UX expert, improve this form' gives direction but leaves essential decisions open. Which users are involved? What can be changed? Which observations should inform the work? The assigned role does not answer these questions. I suggest addressing them in the request itself.

For our fictional form, we can ask it to examine field validation, distinguish errors detected before submission from those returned by the server, and check that input remains available after failure. These technical terms describe behaviours the team can examine. Adding jargon without explaining its connection to the problem would make the request harder to review.

A well-constructed request also specifies what should be delivered. For example: for each observed problem, give reproduction steps and propose an acceptance criterion. We can then discuss a precise case: after an oversized attachment is rejected, does the customer understand what to correct and retain their other answers?

An example: improve a quote request form

Take a fictional case: a small business receives incomplete quote requests. The submit button seems a good starting point, but I would first state the intended result: enable customers to send the information needed for an initial sales response. The team must specify which information is truly essential.

To prepare the work, provide a test copy of the form and anonymised examples of incomplete requests. The question becomes: where is information missing, and does the form make the expectation understandable? A suggestion about button colour must justify its connection to that difficulty.

We can also examine the journey using usability and writing methods: name fields in the customer's vocabulary, explain why a document is requested and observe the feedback after an error. Each avenue must lead to a precise change that can be examined in a pull request.

For this trial, I would keep a simple limit: work on a copy, with no real submissions or production changes. The check could verify that a customer can identify a missing field, correct it and retain information already entered. An automated test can check that behaviour; observation with a user would help evaluate whether the explanations are understandable.

First conversation: develop the request

You do not always know the discipline's terms or which checks to request. I suggest devoting a first conversation to preparation. Describe the subject in your own words, state what you have and ask AI to help construct a more precise request. The two-conversation method below is a Journal proposal.

text
Aide-moi à préparer une demande pour une seconde conversation IA.
Ne réalise pas encore la tâche.

Sujet : notre formulaire de devis reçoit des demandes incomplètes.
Je dispose d’une copie de test et de quelques exemples anonymisés.
Je veux comprendre ce qui manque avant de demander une modification.

Repère les ambiguïtés et pose les questions qui changeraient la démarche.
Propose les méthodes pertinentes en expliquant leurs termes techniques.
Pour chaque méthode, précise les informations nécessaires et le contrôle
qui permettrait de décider si la modification répond au besoin.
Garde un périmètre limité à ce formulaire.

Après mes réponses, rédige une demande autonome à copier dans une
nouvelle conversation. Distingue les faits confirmés des hypothèses.
Liste les documents à joindre et les questions encore ouvertes.
N’invente pas de données, de résultats ni d’accès aux outils.

Review that preparation before passing it on. If AI proposes measuring abandonment rate, check that you actually have the necessary data. If it adds a redesign of the sales journey, ask why that would help with the chosen form. A better-documented question can stay short; every addition should help make a decision.

Second conversation: provide self-contained context

Once the useful points are clear, copy the request into a new conversation with the same AI or another one. Attach the authorised documents it needs to examine. The text must restate decisions made during preparation: a reference such as 'use the data seen above' does not provide that context to the second conversation.

For our fictional case, the outcome of preparation could look like the text below. It still needs the list of essential fields approved by the team and the files actually accessible.

text
Sur cette copie du formulaire de devis, aide le client à fournir
les informations nécessaires à une première réponse commerciale.
La liste des champs indispensables validée par l’équipe est jointe.

Examine les exemples anonymisés de demandes incomplètes.
Pour chaque changement proposé, indique l’observation qui le motive
et un critère d’acceptation observable.
Vérifie notamment les messages de validation et la conservation de la
saisie après une erreur, avec les outils déjà disponibles dans le projet.
Si les données manquent, précise ce qu’il faut observer avant de décider.

Prépare une modification limitée et une pull request avec les contrôles
exécutés et leurs limites. Aucun envoi réel ni déploiement.

Changing conversations organises the transition from preparation to execution. It does not guarantee correctness: two AIs may repeat the same mistaken assumption. To decide, return to the form and agreed checks. Request the results of checks actually performed and keep visible anything still requiring observation with a user.

A first conversation clarifies facts and assumptions. Human review prepares a self-contained request and useful documents, passed to the second conversation to carry out and verify the work.
Method proposed by the Journal for the fictional form. The useful transfer is the reviewed request with its documents; changing conversations does not guarantee correctness.

Sources and useful links

Share on LinkedIn (new tab)

Keep reading
All articles in this category