Ward

Every agent answers in the same five movements, always in this order, and never as a labelled list — the structure lives in the writing, not in headings the reader has to parse.

It opens by describing what is actually visible in the input: the lesion on the underside of the leaf, the rust weeping from the shaft seal, the clause buried on the third page. This lets the reader check the machine against their own eyes before trusting anything that comes after it.

Then it names what the thing is, and says how sure it is in the same breath — not in a badge parked elsewhere on the screen, but in the sentence itself, so certainty and claim cannot be read apart.

Then it names the most plausible alternative, and with it the single observation that separates the two. Not a list of differential possibilities, but one thing to go and look at: check the stem where it meets the soil, and if it is dark and soft it is the first, if it is firm it is the second. One observation, decisive, performable by the reader in a minute.

Then it says what to do, ordered cheapest and safest first, so the reader can stop after the first step and still have taken the best available action. Expense and risk both increase as the list descends; nothing at the top requires a purchase or a specialist.

It closes by saying what the answer rests on — the regulation, the institution, the standard, the extension guidance. Never "studies show" and never a bare assertion. The reader must be able to leave the app and check the claim against something that exists outside it.

Built With

• canvas-api
• content-security-policy
• ffmpeg
• google-ai-studio
• google-gemini
• ics
• idb-keyval
• indexeddb
• next-font
• next.js
• node.js
• openrouter
• playwright
• pwa
• react
• serverless
• tailwindcss
• typescript
• vercel
• web-share-api
• web-speech-api
• zod