Bias checker for documentation and process writing
Build a bias-checker workflow that lets someone submit a URL, pasted text, or an uploaded file (PDF or text) and get back a structured fairness/bias report — an overall score, a per-category severity breakdown (with emphasis on race, sex, and gender), flagged passages with explanations and neutral rewrites, and per-category learning resources (readings, training, articles, links to relevant studies). Design it as an async job so slow work never risks an HTTP timeout: A web page (single-page app) where a user pastes a URL or text, or uploads a file, and submits it for analysis, then polls for and displays the result. An intake API endpoint that validates the submission, creates a job record in persistent storage with a "pending" status, returns a job id immediately, and hands off to a background step to do the real work. A background analysis step (no route of its own, triggered by the intake step) that: fetches and cleans the source content (strip HTML from a URL, decode text, or pass a PDF through natively) capped at a reasonable size limit; calls an LLM — Claude, or your preferred model provider — with a fairness-review system prompt and a structured tool/schema for the response, retrying once if parsing fails; writes the result (or an error) back into the job's status record. A status/polling API endpoint that reads the job's status record and reports pending, error, or done with the full report, so the UI can poll it every couple of seconds until it resolves. Use persistent storage (a shared volume) to hold each job's input and status, keyed by job id, so the intake, background, and status steps can all read/write the same job directory. Enforce a reasonable upload size cap (e.g. a few MB) for files. Note that any resource links or study citations the model suggests are generated and should be flagged as worth verifying before sharing. Keep the intake and status endpoints private to your workspace (space-authenticated) since this reads/creates job data; the web page can be similarly scoped. Tools used: Anthropic Claude API, 3B, React
What this prompt builds
A bias-detection workflow that analyzes pasted text, uploaded documents, or URLs for fairness issues across categories like race, sex, and gender. It returns a scored report with flagged passages, explanations, neutral rewrites, and curated learning resources for each bias type found. The workflow runs asynchronously to handle large documents without timeout, storing results in persistent storage for polling retrieval.
The problem
Leaders and teams building processes, documentation, or public-facing content need to catch unconscious bias before it ships. Spotting problematic language manually is slow and subjective, and most people lack the training to recognize subtle bias across categories like race, gender, and sex. This workflow analyzes any text, uploaded file, or URL and returns a structured fairness report: an overall bias score, per-category severity ratings, every flagged passage with an explanation and a neutral rewrite, and curated learning resources so teams can understand the issue and do better next time.
Solution and impact
Teams submit content through a simple web interface and receive a detailed bias audit within seconds to minutes, depending on length. The report breaks down findings by category, highlights exactly where bias appears, and provides both immediate fixes (rewrites) and long-term education (readings, training links, studies). By surfacing bias early and pairing correction with learning, the workflow helps organizations ship fairer content and build a culture of continuous improvement around inclusive language.