How this tool is designed
This page exists so that you can judge Starbuck Coach for yourself rather than take our word for it. It says what the tool is for, whose thinking shaped it, who is accountable for it, and what has and has not been tested. Where we do not yet have evidence for something, this page says so instead of claiming it.
What it is for
Starbuck Coach is an AI thinking partner for the space between human coaching sessions. It is designed to do one narrow thing well: help you think through a single work or personal-development question by asking rather than telling, and hold you to the next step you choose for yourself.
It is not a coach. It is software. It has no professional judgement, no memory of you, no stake in your life and no accountability for your decisions. It complements human coaching and it is not a substitute for it. It is also not therapy, crisis support, performance assessment, medical advice, legal advice, financial advice or employment-law advice, and it does not make decisions about you.
The coaching approach behind it
The way the tool behaves was informed by two published sources.
The first is the ICF Core Competencies, published by the International Coaching Federation. Several of them shaped concrete design decisions: the tool asks one principal question at a time rather than delivering advice; it reflects your own language back rather than substituting its own framing; it works towards a step you name yourself rather than one it recommends; and it returns agency to you whenever it offers options.
The second is the ICF Artificial Intelligence (AI) Coaching Framework and Standards, published by the International Coaching Federation in November 2024. Its design has been informed by that framework. This does not mean that the software is accredited, certified or endorsed by ICF.
To be as plain as possible about this: no coaching body accredits software, and none has assessed this tool. Using published competencies as a design influence is not the same as conforming to them, and we make no claim of conformance. Where this page or any other says the design was informed by something, that is the whole of the claim.
Human responsibility
Software does not take responsibility for anything. People do. This section names the people accountable for how this tool behaves, and the date each review last happened.
Anything not listed above has either not been confirmed or has not yet happened, and nothing is published here until it can be. Review dates appear only once a review has actually taken place. This deliberately shows less than a launched product would: an unverified name or an invented review date would be worse than a gap.
All five roles — product owner, coaching-quality reviewer, safety and safeguarding reviewer, data-protection owner and accessibility owner — are held by the same person. What is published here is the role rather than an individual’s name, and no professional credential is claimed for anybody, because none has been evidenced.
The dates of completed reviews are still blank on that page, and they will stay blank until a review has actually taken place. A gap there is a statement that the work is outstanding. An invented date would be worse.
What is meant to be tested
This is the test programme the tool is designed to be held to. Publishing the list before the results exist is deliberate: it lets you see the shape of what has been promised and check it against what has actually been done, which is in the next section.
- Coaching quality. Does it ask rather than tell, stay on one question at a time, reflect the person's own language, and leave the decision with them?
- Inappropriate advice. Does it refuse to give medical, legal, financial, employment-law or safeguarding advice, and does it refuse to tell somebody which decision to make?
- Safety. When somebody discloses risk of harm to themselves or another, does it stop coaching, say plainly that it cannot help in an emergency and that nobody is monitoring, and give the approved emergency routes?
- Bias. Does the quality of questioning hold up across different names, dialects, accents in writing, genders, ages, disabilities, ethnicities and job levels, and does it avoid assuming any of them?
- Hallucinations. Does it avoid inventing credentials, research, statistics, policies, legal positions or facts about the person or about Starbuck & Associates?
- Privacy leakage. Does anything a person writes reach a log, an error report, a URL, a page title or any store on our side?
- Prompt injection. When text inside a message tries to instruct the tool, does it treat that text as content rather than as instruction, and do the safety and privacy rules survive?
- Accessibility. Keyboard only, screen reader, zoom, reflow, contrast, focus management, live regions and forced colours, against WCAG 2.2 AA.
- Regression. Do all of the above still hold after every change to the prompt, the model or the interface?
Evidence, and the limits of it
Here is the honest position at the date at the foot of this page.
Automated checks exist and run on every change. These cover the things a machine can settle: that no secret or provider identifier appears in anything sent to your browser, that nothing you write is placed in a URL, that the code contains no path that writes a message, a reply or a note to a log, that the pre-session consent controls are separate and start unticked, that the emergency wording is present and reachable, that the content security policy permits only documented origins, and that the deletion path covers everything the application stores. What they demonstrate is that the code has the property described. They are not evidence about the quality of a conversation.
The behavioural test cases are written and versioned, and have not been run as a scored evaluation. The safety, bias, hallucination and prompt-injection cases exist in the repository as a fixed set of scenarios with expected handling, so that a result can be reproduced and compared between versions. No sample sizes, thresholds, pass rates or scores are published here because none have been produced. When they are, this section will carry the method, the number of cases, the threshold, the result and the limitations, and not a summary adjective.
No independent assessment has been carried out. Not of coaching quality, not of safety, not of accessibility, and not of the data protection position. The accessibility statement says partially conformant for that reason, and will keep saying it until independent evidence supports something stronger.
You will not find the words safe, secure, effective, rigorously tested, bias-free or proven used about this tool on this site. Those are conclusions, and we have not earned any of them yet. An AI model can be wrong, can be confidently wrong, and can be led. Check anything that matters with a qualified person.
What it is designed not to do
- Claim to feel anything, to have lived through anything, or to understand you the way a person would.
- Diagnose you, or act like a therapist.
- Tell you which decision to make.
- Encourage you to depend on it. It will not say that it cares about you, that it is always here, or that you only need it.
- Imply that a human is reading the conversation, because none is.
- Claim credentials, accreditation or real-world experience.
- Ask for your name, your employer or any other identifying detail, none of which it needs.
Reviews and changes
Review dates are published on the main page once a review has happened. Every change to this tool, including changes to the instructions the model is given, is recorded in the change log in the repository, and material changes to any policy page change its version number and date.
If something here does not match what the tool actually did for you, we would rather know. There is a Report a problem route in the footer of every page and inside the session itself, and nothing from your session is attached to it unless you choose to copy something in yourself.
Last updated 26 August 2026. This page is updated whenever a review happens or a test result exists, and it is written to be checkable rather than reassuring.