SMF·HYPERWRITE
Can a prompt replace HyperWrite?
AI writing — AI drafting, research and content optimization
Exhibit tracking slip
Verdict
The idea worth copying is that a prompt becomes a named tool: you describe the task once, attach a few examples of the output you want, and afterwards it is a menu item rather than something you retype. Building that library, with example management and a diff view, is a few days rather than an evening — and it needs enough discipline around versioning that improving a tool does not silently change what your old outputs would have looked like. The browser agent side of HyperWrite is out of reach entirely.
Exhibit A — The prompt
Received on31.07.2026Build a local library of user-defined writing tools, each a reusable few-shot prompt.
Tool definition: a name, a short description, an instruction, a list of example pairs (input text and the desired output), an output format (prose, bullet list, structured fields), and the model to use. A tool may declare parameters the user fills at run time — a tone, a length, an audience — which are substituted into the instruction.
Running a tool: paste or select input, choose a tool, fill any parameters, run. The request sends the instruction, the examples in order, and the input. Show the assembled prompt before sending, collapsed by default but always available — a tool whose behaviour you cannot inspect is one you cannot debug.
Examples are the substance, so treat them properly. Adding an example from a run's own output is one click, since that is how a tool actually gets good: run it, edit the result until it is right, and save the edited version as the new example. Show how many tokens the examples consume and warn when they crowd the context.
Versioning, which matters more than it looks. Editing a tool creates a new version rather than overwriting; every run records which version produced it. A results history lets you compare the same input across two versions side by side, so "I improved the tool" is a claim you can check rather than assume.
Output handling: results shown as a word-level diff against the input where the tool is a transformation (rewrite, shorten, adjust tone), and as plain output where it is a generation (draft, summarise, expand). Copy, save to a file, or feed straight into another tool as its input, so tools can be chained by hand.
Library management: tools organised by folder, exportable and importable as JSON, so a set can be shared as a file. Ship no bundled tools — the first-run experience is a guided creation of one tool from a real task the user names.
Use Anthropic or OpenAI, whichever key is present, selectable per tool since different tasks suit different models. With no key, the library, examples, versions and history must still be browsable, and say why running is unavailable.
Storage: SQLite, holding tools, versions, examples and every run with its input, output and version id.
Out of scope: a browser extension, any agent that operates a browser or takes actions on the web, sync, and personalisation learned automatically from your writing.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 the browser extension that runs these tools wherever you type
- B.2 the AI agent that operates a browser to complete multi-step tasks
- B.3 personalisation learned from your writing over time
- B.4 sync of your tool library between machines
Prior art
Exhibit C — Why people still pay: workflow, data, and model tuning
Because the value compounds with where it lives. A tool library in a separate application gets used when you remember it; the same library inside your email and your documents gets used all day.
Questions
Can I bring my HyperWrite tools across?
There is no export, so each custom tool is re-created by hand. That is usually quick — the instruction is a sentence or two — but the examples are the valuable part and those you will have to write again from your own past outputs.
Do the examples really matter that much?
More than the instruction does. Three good input-output pairs pin down tone, length and structure far more reliably than a paragraph of description, and the save-output-as-example loop in the prompt exists because that is how a tool goes from roughly right to consistently right.
What does it cost to run?
Each run sends the instruction, the examples and the input, so a tool with three long examples costs more per call than a bare prompt — typically a cent or two. Heavy daily use lands in the low single-digit dollars per month, well under the subscription.
What is the one thing that does not survive the rebuild?
The agent. HyperWrite's Personal Assistant navigates websites and completes multi-step tasks, which is a different product from a prompt library and is not something a local tool with no browser control can approximate.
Related tools
Receipt