Write the product brief once. Loopcraft turns it into a focused scope, ordered implementation prompts, and review gates that keep Claude or Codex moving without prompt archaeology.
Runs locally in this page. Your brief is used to derive the packet and is never sent anywhere.
How does Loopcraft translate raw brief text into structured scopes, prompts, and rework reduction metrics?
Loopcraft is a deterministic client-side structuring utility rather than an AI agent or language model. When a user submits a product brief, the script strips non-alphanumeric punctuation and splits text into whitespace-delimited tokens. Tokens longer than four characters are treated as keyword nouns to fill predefined scope templates, while the prompt sequence and review gates follow static multi-pass templates. The reported rework reduction metric is a simple inverse linear heuristic based strictly on brief word count, capped at 62%.
The rework reduction percentage is an arbitrary heuristic tied to word count, not an empirically measured software engineering outcome or benchmarked token savings. Word extraction relies on string length rather than natural language parsing or part-of-speech tagging. The tool generates local text templates; it does not connect to Claude, Codex, or execute automated coding workflows.
Clicking 'Use sample brief' populates the textarea with 'A mobile app for scheduling neighborhood tool lending.' Clicking 'Generate build packet' evaluates the 8-word input. Words longer than 4 characters ('mobile', 'scheduling', 'neighborhood', 'lending') populate the scope list, 3 sequential multi-pass agent prompts are rendered, and the rework reduction calculates to 28% (18 + 10).
The script sanitizes input by removing non-alphanumeric characters and splitting on whitespace. It extracts up to four unique words exceeding four characters in length. If none exist, it falls back to a default tuple ('product', 'workflow', 'user', 'outcome'). These words are interpolated into fixed template strings to define MVP scope slices.
The 3 generated prompt passes and 4 review gates are predetermined template strings. Prompt 1 incorporates the brief title (truncated to 72 characters), while passes 2 and 3 provide generic instructions to implement happy paths and review against criteria. The review gates evaluate observable user outcomes, testable transitions, failure paths, and handoff instructions.
The estimated rework reduction percentage is calculated via Math.min(62, 18 + Math.max(0, 18 - wordCount)). For brief lengths of 18 words or more, the metric defaults to 18%. For shorter briefs, each word below 18 adds 1%, capped at a ceiling of 62%. It serves purely as an illustrative heuristic rather than an empirical software delivery benchmark.
The sample brief has eight words. The local parser lowercases it, strips everything except ASCII letters, digits and spaces, and keeps the first four unique tokens longer than four characters. Mobile, scheduling, neighborhood and lending become the focus. It is length filtering, not linguistic noun recognition. The generated packet always contains four scope slices, three prompt templates and four review gates; it runs no model or deployment. The source labels an estimated rework reduction. For the eight word sample the formula gives eighteen plus ten, or twenty eight percent. At eighteen words or more it shows eighteen percent. Nothing in the formula measures actual rework, product quality or agent performance. The nominal sixty two percent cap is unreachable because the inner expression is at most thirty six even with zero parsed words. A nonempty punctuation-only brief can reach that upper value. Generate stores the trimmed brief in browser local storage and renders the planning packet. JSON includes the original brief, generated timestamp, MVP scope, prompts and review gates. Text exports title, scope, prompts and gates but omits the full brief and estimated saving. Restoring a saved brief rebuilds a new packet locally. The gates are suggestions rather than enforced tests; exported prompts do not execute Claude, Codex or any backend.