Seven-phase agent pipeline writes a help center from a product login
Velu rebuilt help-center generation as a seven-phase agent pipeline that only needs a product login. Task: Velu implemented a seven-phase agent pipeline to generate help centers from product logins without repo access. The workflow uses Haiku for research mapping and screenshot capture, Sonnet for planning, writing, and editing, and a non-AI CLI for publishing. Human teams customize editor rules via SKILL.md and flag changed screens for selective updates.; tools: Browser agents, Claude Haiku, Claude Sonnet, SKILL.md, velu CLI. Human role: Not stated in the source; limitations: This is a vendor engineering post from Nov 3, 2026, with narrative rather than audited metrics for cost and coverage improvements. The workflow requires a valid product login and assumes the UI is stable enough for automated mapping. Human approval is necessary for final publishing and handling of changed screens..
Tools
Browser agents, Claude Haiku, Claude Sonnet, SKILL.md, velu CLI
Published on Agent Work Atlas:
Overview
Velu rebuilt help-center generation as a seven-phase agent pipeline that only needs a product login.
Workflow
Velu implemented a seven-phase agent pipeline to generate help centers from product logins without repo access. The workflow uses Haiku for research mapping and screenshot capture, Sonnet for planning, writing, and editing, and a non-AI CLI for publishing. Human teams customize editor rules via SKILL.md and flag changed screens for selective updates.
Reported result
The split architecture resolved token cost explosions and context loss found in single-agent runs, ensuring cost scales linearly with page count. Editors successfully caught hallucinated UI labels, and mock data generation maintained visual consistency without exposing real customer records. Exact financial savings or page counts are not published in the source.
Human role
Not stated in the source
Limitations
This is a vendor engineering post from Nov 3, 2026, with narrative rather than audited metrics for cost and coverage improvements. The workflow requires a valid product login and assumes the UI is stable enough for automated mapping. Human approval is necessary for final publishing and handling of changed screens.
Try it
Act as a technical writer agent. I need to generate a help center for [Product Name] using a seven-phase pipeline. Inputs: - Product URL: [URL] - Test Login: [Username/Password] - Feature Checklist: [List of key tasks] - Style Guide: [SKILL.md content or rules] Steps: 1. Research: Use a browser agent to map screens based on the checklist. Record exact button names and step orders. 2. Plan: Define task-based topics (e.g., 'Cancel Order') rather than screen-based ones. 3. Write: Draft one page per topic using only relevant map notes. 4. Edit: Review drafts against the style guide and map. Remove any claims not supported by the map. 5. Mock Data: Generate consistent fake data for screenshots. 6. Capture: Take screenshots/videos using the mock data. 7. Publish: Run a link/image check script before finalizing. Constraints: - Do not write back to the product database. - Flag any missing screenshots. - Output a draft site structure and sample pages for my review before full generation. First, perform a dry run on [One Specific Feature] and show me the map notes and draft page for approval.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.