quality-bar.md

| # quality-bar.md | | | A template for encoding your quality criteria before building with AI. | | | Fill this in as a team before the first prompt is written. | | | The conversation it forces is the work. | | | Update it as your understanding evolves. Version it. | | | This is a living document, not a one-time exercise. | | | --- | | | ## 1. Purpose | | | What problem does this solve? | | | One sentence. The real problem, not the feature description. | | | Who experiences this problem, and when? | | | Be specific. Not "users" — a real person in a real moment. | | | What would they say if this worked perfectly? | | | Write the sentence they would say to a colleague the next day. | | | What would they say if it failed? | | | Write the complaint. Be honest about what failure looks like from their side. | | | --- | | | ## 2. The user in the moment | | | What state is the user in when they encounter this? | | | Rushed, confused, expert, anxious, curious? The emotional and cognitive state matters. | | | What do they know that the system doesn't? | | | What context, history, or preference exists in their head that AI cannot see? | | | What do they assume the system can do that it can't? | | | Where is the gap between user expectation and actual system capability? | | | --- | | | ## 3. Behavioral specification | | | In normal conditions, the AI should: | | | When it doesn't know, the AI should: | | | When it is uncertain, the AI should: | | | When the user is wrong, the AI should: | | | When it makes a mistake, the AI should: | | | --- | | | ## 4. What good looks like | | | Describe a perfect interaction, step by step. | | | What does the output look, sound, and feel like when it's right? | | | What would a senior designer say about it in critique? | | | What specific quality markers would they call out? | | | --- | | | ## 5. What wrong looks like | | | Describe an output that looks right but is subtly wrong. | | | The failure mode that passes a visual check but fails the intent check. | | | Describe an output that is confidently wrong. | | | The hallucination or overreach case specific to your product. | | | Describe an output that is technically correct but unhelpful. | | | The case where the AI answered the question but missed the point entirely. | | | --- | | | ## 6. Voice and tone | | | How does this AI speak? | | | Three adjectives that describe the voice. Three it explicitly avoids. | | | Sounds like: | | | Never sounds like: | | | What does on-brand output look like versus off-brand? | | | A specific example of each for this product. | | | What does the AI never say, even if prompted? | | | --- | | | ## 7. The not-doing list | | | This product refuses to: | | | This product will never do, even if asked: | | | We have considered and deliberately declined: | | | List the things that would technically work but that you are choosing not to do, and why. | | | --- | | | ## 8. Ethical boundaries | | | What user autonomy must this always preserve? | | | What data does this handle, and what does it never do with it? | | | Where might this cause harm, and how have we addressed it? | | | Who could be disadvantaged by this, and what have we done about it? | | | --- | | | ## 9. Criteria for AI-generated output | | | Every output must: | | | An output fails the bar if it: | | | Who reviews output, and what are they specifically looking for? | | | At what stage does review happen? Who has authority to reject and send back? | | | --- | | | ## 10. How we know | | | What user signal tells us the bar is being met? | | | What signal tells us quality has slipped? | | | How often do we review these criteria, and who owns that review? | | | --- | | | ## Metadata | | | Last updated: | | | Updated by: | | | Version: | | | Approved by: | | | --- | | | Embed this file in your tooling so it travels with the work. | | | This document should live where the work lives, not in a folder nobody opens. | | | MIT Licensed — Owl-Listener (MC Dean) |