Describe what you want in plain language. Five passes later you get a structured specification with scope, output contract, error handling, and a quality block that scores it on eight axes.
Prompts, Agent, Workflow, Pipeline, Multi Agent or MCP Tool. The choice selects a whole engine, not a template. Each has its own spec sections and its own eight quality axes.
One sentence is enough. Start from a curated example if it is easier. If something load bearing is missing, the forge pauses and asks rather than inventing it.
Each builds on the last, producing a structured document with scope, output contract and error handling, closing with a quality block scoring it on eight axes.
Open the editor to rewrite any block or ask for alternatives, and watch the score move. Then copy it, open it in Claude, or export it.
Generation is the first half. The editor and the scorecard are what turn a draft into something you would ship.
The quality block is read out of the spec itself and rendered as an overall percentage plus a per axis breakdown, so the document and its grade can never drift apart. On multi agent builds the numbers come from a separate evaluator that regrades the finished spec rather than from the model that wrote it.
The spec opens as its real sections, each with its type and line count. Rewrite a block or ask for alternatives; a structure guard rejects an edit that would break the document.
The simplest of the six: one precisely worded instruction, with no loop of its own to go wrong. There is no runtime to guard, so the discipline goes into the wording. Every prompt is scored on the same five axes, Completeness, Accuracy, Depth, Actionability and Quality, so one can be measured against another.
One AI actor that loops, calls tools and picks its own next step. The engine resolves an iteration cap from the complexity of the build and writes it into the spec twice, as the loop bound and again as a stop condition, and every tool has to carry a named fallback for failure rather than the words handle the error.
A fixed sequence of steps, more predictable than an agent because the path is decided in advance. Every workflow sets a retry budget of three attempts with exponential backoff and jitter for temporary failures, and marks each irreversible stage with an idempotency key, so a run that fails at step 7 of 10 does not redo steps 1 to 6.
Built for volume, moving many records through the same steps. Failures are sorted before they pile up: a temporary one is retried with exponential backoff up to three attempts, while a bad record is routed to a dead letter queue on its first failure instead of being retried or quietly dropped.
More than one agent working toward the same goal. The engine reads the coordination round ceiling straight from its own constants, 10 rounds for simple, 25 for moderate and 50 for complex, and every spec has to carry a failure isolation section saying what happens when one agent fails so the rest keep going.
A tool you hand to any MCP compatible client, not just your own product. Because it leaves your control, its spec always carries a security section listing what must never appear in output or an error message, names the environment variables the server checks before it starts, and scores security as one of its eight quality axes.
Pushing every record through a model is the expensive way to work. On a Pipeline build a deterministic filter sits ahead of every model stage and clears the obvious cases first, so you only pay for the ambiguous middle, and a build that leaves a stage uncovered gets rejected rather than shipped.
Sign in and everything you build saves to your own library. Make folders, drag prompts to reorder or move them, star the ones you keep coming back to. Every read and every write is checked against your account, so nobody else reaches your saved work.
Every change is rebuilt and reparsed before it lands, and one that would break the section structure is refused. Rewrites, alternatives and rescoring all cost no credits, so revision is not something a team has to ration.
Professional grade specs, no shortcuts.
Solo gets you a spec. Pro gets you the system.
Everything Pro can do, with people behind it.
Credits are what runs cost: one for a Prompts run, two for a build in any of the five builders.
A prompt is a paragraph you end up rewriting. A spec is a structured document with identity, explicit scope, an output contract and error handling, so the model has fewer places to improvise, and you can review it before you run it.
A Prompts run costs 1 credit. A build in any of the five builders costs 2, taken once upfront when the run starts. Choosing a type and describing the idea are free, and editing the result afterwards costs nothing.
Nothing you made is locked or deleted. Saved prompts stay in your library. What stops is new builder generation and the editor. The trial runs 3 days from signup and expires on its own, and no card is involved.
No. It is plain text with no proprietary format, so paste it into any model. There is a one click handoff to Claude because many people use it, but it is a convenience, not a dependency.
No. Prompts run through Anthropic's API under its commercial terms, where inputs are not used to train the models, and we operate in zero data retention mode. Our gateway strips metadata and sends only the prompt text.
Not in the middle of a run. Picking a build type selects a whole engine, not a template, so a different type means starting a new build. Nothing is lost either way: the spec you already made stays saved in your library.
You are never limited to one. All six are unlocked from the first day, on the trial and on every paid plan, so you can use whichever one fits the job rather than the one you happened to start with.