thinkhard.SI
Add API key

The bench notebook.

Every change worth knowing about, newest first. Model prices are rechecked whenever a provider changes them.

  1. Three new tools: Prompt Builder, Custom Instructions and Copy Cleaner

    • The Prompt Builder asks nine plain questions about the task and writes a prompt that scores 10 out of 10, laid out for ChatGPT, Claude, Gemini or any other model. Open it in ChatGPT or Claude, save it, or share it as a link.
    • Custom Instructions writes the standing instructions an assistant reads before every chat, in the format ChatGPT, Claude, Gemini and Copilot each expect, with where to paste them and three prompts to check they worked. Start from one of nine roles.
    • The Copy Cleaner turns an AI answer’s stars and hashes into real headings, bold, bullets and tables for Word, Google Docs, email, Slack or Teams, and can drop the long dashes, chatbot openers and emoji on the way.
    • Templates, collections and packs now say what they are: prompts for any AI model, with the layouts for ChatGPT, Claude and Gemini where they differ.
    • Charts and diagrams in the guides, on the home page and on the template index, drawn from the site’s own numbers.
    • Every email from thinkhard.si, from the sign-in link to the Tuesday confirmation, has a new design.
    • Fixed: some sign-ups from browsers that fill in forms automatically were mistaken for spam and dropped.

  2. A REST API, what a one-line prompt scores, and a new footer

    • A free REST API at /api/v1, with an OpenAPI description: search and fetch the prompts, grade a prompt and price an AI feature in plain JSON, with no key. Every template page shows its own API address.
    • New research: what a one-line prompt scores, across 319 everyday work tasks, and the two short additions that lift the average from 1.2 to 5.7.
    • Answers to the questions people ask about the 30 most-searched templates, such as whether employers can tell AI wrote a cover letter.
    • Your workspace now searches everything you saved, scores a prompt as you type, reuses any prompt in ChatGPT or Claude in one click, and backs up and exports your work.
    • The Tuesday prompt email can be joined from every template, guide and post, and is now sent through Buttondown.
    • A calmer footer, with the name drawn in type.

  3. Developer docs for the MCP server, collection pages and three guides

    • A tools reference with every parameter, the fields each tool returns and a real answer for each call, and a playground that calls the server from your browser and copies the call as curl, JavaScript or Python.
    • Use it from code: working examples for plain HTTP, the official MCP SDKs, the Claude API and the OpenAI Responses API.
    • One-step installs for Cursor and VS Code, and a Claude Desktop extension that needs no Node or terminal. Setup steps for Codex, Gemini CLI, Devin Desktop, Zed, Cline and more.
    • The MCP server is now version 1.1.0: every tool declares an output schema, so code that reads the answers knows their shape.
    • A page for each of the 20 collections in the repository, such as prompts for HR and recruiters and prompts for teachers.
    • New guides: how to fact-check an AI answer, AI words to avoid, how to anonymize text before you paste it into AI, SI vs AI and MCP servers for people who don’t code.
    • A redrawn footer, made of text: noise at the top settles into letters, words and whole prompt lines, and at the bottom a globe of prompt lines turns, with the name catching the light. Drag it to turn it. There is a sign-up there for one tested prompt every Tuesday.
    • Your workspace: the prompts you saved and how each one has improved, your Grader reports and tests, and what your API keys spent, on one page. It reads only what this browser already holds.
    • Each of the 20 collections has its own mineral specimen, from barite for email to bismuth for code, and collection and pack pages have their own cards when you share them.
    • Two free prompt packs for the busy seasons, performance reviews and report cards, with every prompt shown in full and no sign-up.

  4. Use thinkhard.si inside your AI, and fewer keys

    • An MCP server at https://thinkhard.si/mcp. Add it to Claude, ChatGPT, Claude Code, Cursor or VS Code, and your assistant can search the prompt repository, fetch and fill in a prompt, grade a prompt out of 10 and price an AI feature. Every prompt also becomes a slash command.
    • A local version for developers, one npx command, that also checks drafts for robot-speak and lists the facts to check, without the text leaving your computer.
    • No key needed to improve a prompt: the Grader now writes the request and opens it in ChatGPT or Claude, using the account you already have. A key still lets you do it here, with a score before and after.
    • Model Compare without a key opens your prompt in ChatGPT, Claude and Perplexity at once.
    • The “Add API key” button now appears only on the tools that can use one. Eight of the ten tools never need a key.

  5. The prompt repository: 319 prompts, 210 of them new

    • 210 new prompts, chosen from research into what people actually use AI for at work: email, presentations, meetings, job search, marketing, spreadsheets, long documents and more. Every one is scored on the ten checks and invents nothing: your facts go in as blanks.
    • Five new collections: Marketing, Job search and careers, Project management, Research and learning, and Legal and compliance.
    • The templates page is now a repository: each prompt has a registry number, a collection, the model it is laid out for, its score and the date it was added.
    • Ten prompts are open to everyone. Your email unlocks every other one, free, and you can choose to get new prompts every Tuesday. What we keep.

  6. Three new tools for everyday AI, and plainer names

    • Robot-Speak Checker: paste an AI draft and see every phrase that makes it sound machine-made, from long dashes to “not just X, but Y”. Fix the safe ones in one click, check it against your house rules, and copy instructions that stop it next time.
    • Fact Check List: paste an AI answer and get a checklist of every number, quote, source, link, law and court case to check, riskiest first, with a search link for each.
    • Safe Paste: swap names, emails, phone, card and account numbers for placeholders before you paste into an AI, then put them back in the answer.
    • All three are free, need no API key and run entirely in your browser.
    • Tools are now named for what they do: Prompt Lab is now the Prompt Tester, Model Arena is Model Compare, the Injection tester is the Chatbot Safety Test, Prompt diff is Version Compare, and the Library is My Prompts. Links still work.
    • The Tools menu is split into tools for everyday AI, which need no key, and tools for building with AI.

  7. Think hard, SI: a new mark

    • Read the address out loud: think hard, SI, where SI is super intelligence. The name explains it.
    • A new mark, a speech bubble cut like an emerald: what you say to the model, shaped with care, with SI in gold at its heart. It is on every page, the browser tab and the home screen icon.
    • The home page shows where your words go: through a typical AI app’s server, or straight from your browser to the AI company with thinkhard.si out of the way.
    • Every guide, post and use case now has its own preview card when you share a link.

  8. Straight into ChatGPT or Claude, and plainer words

    • Wherever a finished prompt appears, in the Grader and on every template, Open in ChatGPT and Open in Claude start a new chat with the prompt already in it. If a detail is still blank, you are asked first.
    • The ten checks and the score out of 10 now use the same plain names on every page, including the templates, the Library and Prompt diff.
    • A new address for questions and reports: contact@thinkhard.si.
    • Tidier tables on phones, easier-to-read card titles and summaries, and shorter lines in notes and captions.

  9. Start here and the reading pages, redesigned

    • Start here is now a set of goal cards: pick what you want to do, see which tool does it and whether it is free.
    • Guides, posts and glossary groups each carry a topic icon, so you can find what you need before you read.
    • The blog opens with its latest post and a card for the next one, and the guides are grouped by who they are for.
    • The glossary is searchable and grouped by subject.
    • Every guide and post starts with an In short summary, numbers its sections and shows how far you have read.

  10. A friendlier Grader, built with people new to AI

    • A plain score out of 10 with a word for it, from Needs work to Excellent, as you type.
    • Questions instead of jargon. Each thing your prompt is missing becomes a short question with quick answers, and a better prompt builds as you answer. Free, with no key.
    • Improve with AI is faster. The review and the rewrite now start together, the new version appears as it is written, and the default setting is Fast.
    • Results lead with your score before and after and your new prompt, ready to copy. The ten checks and the part-by-part notes are one click away.
    • Every tool’s How it works now opens with a diagram: what happens, in what order, what is free and what needs a key. The ten checks are shown on one real prompt, and the cost formula is drawn term by term.
    • Step charts are drawn as a route, and the home page shows where thinkhard.si fits next to observability platforms.

  11. Use cases for every team, a price index and a faster site

    • Use cases by team and by industry, from a DCF review in finance to patient letters in healthcare administration. Each one names the jobs AI does well there, the templates to start from and what it costs to run.
    • The AI model price index: what every major model charges for input, output and cached input, and what common jobs cost on each one a month.
    • One place to learn: every guide, post, use case and reference page, with a short reading path for people new to AI at work.
    • Charts in the guides and posts, so the numbers are easier to compare at a glance.
    • New guides, including how to write a system prompt and how to reduce made-up answers.
    • The blog now publishes a new post every Tuesday. The next one is always shown on the blog, and you can sign up to get it by email.
    • Pages open faster. Scripts load in one round instead of four, the main interface font is a quarter of its old size, and the next page you are likely to open starts loading before you click.

  12. 109 templates, seven new guides and a blog

    • The template library grows from 30 to 109, with new sections for hiring, managers, finance, everyday email and meetings, small businesses and training. Every template now has its own page with the full prompt, what to fill in and related templates.
    • Seven new guides on prompting at work, from what an API key is to a risk checklist for using AI at work.
    • A blog, with worked numbers and opinions, and an RSS feed.
    • A log of what your API keys were used for: every call, the tool that made it, the model and an estimated cost. Your prompts are never stored in it.
    • Plain-words help on every tool, a glossary whose words you can tap anywhere on the site, and a one-minute key to which tool you need.
    • The cost calculator now starts from what you are building and how often it runs, and can plan to a monthly budget.
    • A new footer: a section through the site, logged like a borehole, with the diamond at the bottom.
    • Corrected the cached-input price of Claude Sonnet 5.5 to $0.10 per million tokens.

  13. A new look, and a way to reach us

    • Every page has been redrawn: calmer type, an engraved specimen plate for each tool, and a Tools menu that shows all seven at once.
    • The home page now starts with the Grader. Paste a prompt and watch its hardness change as you type.
    • A contact page, so a bug report or a request reaches a person.
    • Anonymous usage counts: daily totals of page views and tool runs, with no cookies and nothing stored about you. Browsers that send Global Privacy Control or Do Not Track send nothing.
    • This notebook.

  14. thinkhard.si opens

  15. The first Grader

    • A one-page prototype: ten plain tests, a hardness from talc to diamond, and a rewrite that fixes what failed. It became the Prompt Grader.