What has changed on WriteMyAIPrompt and when. Entries are added when something ships. The compiler, the guides, the glossary and the prompt library all move independently, so each entry says which part changed.
A faster path to a strong prompt
Added "Sharpen it" to the builder: a few one-tap questions chosen for the detected task type — audience, tone, length, format and must-have rules — plus a suggested quality bar. Almost every prompt copied so far came from a one-line brief scoring under 40; each answer is real information, so it is counted by the same published rubric.
Added "Run it in" links that open the finished prompt directly in ChatGPT or Claude, pre-filled, and copy it for Gemini. Nothing passes through this site on the way.
The builder now suggests the image format when a task looks like an image but a chat model is selected, since image models follow a descriptor list better than a written brief.
Redesigned the site: a self-hosted typeface, syntax-coloured prompt output, a score ring, a compact header and a mobile copy bar that appears while the prompt is off-screen.
In-text links are now underlined as well as coloured, fixing a link-in-text-block failure that only shows up when axe runs against the rendered page.
Fixed AI polish, which had returned nothing since 17 October: the model was spending its whole token budget reasoning before answering. Thinking is now off for this wording-only step, and a polish takes about 3 seconds instead of 30.
The polish guard no longer discards a rewrite for a vague word inside a bracketed note to you, such as "[Add an example of a good answer]" — notes the rewrite is asked to write.
The site now carries ads from Google AdSense, which is how it stays free. The privacy page explains the cookies Google uses and how to opt out; the builder still never sends your prompt anywhere unless you press Polish.
Accessibility and polish
Audited every page with axe-core. Zero violations across all 180 pages.
Fixed three colour-contrast failures found by a token-level audit: muted-faint text was 3.36:1 against 4.5:1 required, and form control borders were 1.56:1 against the 3:1 non-text requirement.
Added a light/dark theme toggle that applies before first paint, and copy buttons on every prompt page — both in about a kilobyte of vanilla JavaScript, so content pages stay effectively script-free.
Ready-made prompt pages
Published 52 use-case prompt pages, each compiled by the real engine and explaining what makes that specific task hard to prompt for.
Added build-time quality gates: a page under 600 words of unique content, or sharing more than 70% of its eight-word shingles with another page, fails the build.
Guides, glossary and library
Published 15 guides covering prompt structure, few-shot and chain-of-thought prompting, system prompts, structured output, prompt injection, and evaluation.
Published a 69-term glossary, each definition written to stand alone as a single sentence.
Published 31 hand-written library prompts with the techniques behind each one broken down.
Search and machine readability
Added llms.txt and llms-full.txt, plus a markdown twin of every page, so language models can read the site without executing anything.
robots.txt now names 30 AI crawlers explicitly, because several honour only a block naming themselves.
Structured data across the site, with stable entity identifiers so the whole site resolves to one organisation rather than 180 unrelated pages.
The builder and the improver
Launched the prompt builder: describe a task and it compiles a structured prompt for Claude, ChatGPT, Gemini, Grok, Llama or an image model.
Launched the prompt improver, which reads an existing prompt back into its parts and rebuilds it, explaining every change.
Published a library taxonomy covering 508 use cases.
Added optional AI polish through DeepSeek, with a guard that discards any rewrite which drops part of the prompt structure.
The prompt compiler
Built the deterministic engine behind the site: twelve task archetypes, five model syntaxes, and a published scoring rubric.
The compiler runs entirely in the browser, which is why the tool is free at any level of traffic.
The scoring rubric and the full compilation pipeline are published on how it works.