跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

Ten Unorthodox WorkBuddy Moves: Persistent-Memory Cron, Session Hijacking, and One-Click Replication

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The concrete value is that WorkBuddy can operate with your real login state and local files, so it turns manual browser backends into scriptable data sources and automates repetitive tasks with persistent memory. For developers, the patterns described here are immediately copyable, and the quantified 3–5x credit cost of expert teams is a real budgeting factor when planning agent-heavy workflows.

Summary

WorkBuddy pairs cron with persistent memory and the ability to bind experts, skills, and working directories. The article shows how to use it for a nightly work log that summarizes sessions, detects file changes, and writes a redacted markdown report, plus a project health report that checks ports, recent activity, dependency health, and staleness. Both produce structured output with strict rules: only report measured facts, never fabricate, and reply NOOP when nothing needs attention.

Browser control is the standout hack. Because WorkBuddy connects through a browser extension to your live Chrome or Edge session, it can borrow your login state and open tabs. That means a backend with no export button becomes a data source: it can paginate through order lists, click detail popups, and compile a structured CSV. The same trick works in reverse for filling multi-step web forms from a local file, and for periodic patrols of logged-in dashboards with diffing against stored snapshots.

The article also covers Skills (packaging a personal workflow into a reusable capability), human-AI co-writing (select text and modify in place), table analysis, custom MCPs for private integrations, a shared data library for humans and agents, expert teams that consume 3–5x credits, a WeChat assistant for remote control, and an inspiration community where one-click replication beats starting from scratch.

Takeaways
— Automation in WorkBuddy is a cron with persistent memory, and it can bind experts, skills, and working directories.
— A daily work log automation can retrieve conversation records, check file modifications, and write a redacted markdown report with fixed sections.
— Redaction rules in the work log prioritize privacy, replacing names, clients, systems, and paths with generic placeholders.
— A project health report checks port liveness, recent activity, dependency health, and stale output files, writing findings to health-YYYY-MM-DD.md.
— Skills require a YAML frontmatter with `name` and `description`; the description is the critical field for whether a skill gets triggered.
— Browser control borrows the user's live login state, cookies, and open tabs via a browser extension.
— A browser session can extract CSV data from a backend that has no export button, by clicking through pages and detail popups.
— The same browser control can fill multi-step web forms from a local file and save drafts without hitting publish.
— Custom MCPs let users connect private resources like a Typora note library, and newly added MCPs need to be explicitly trusted before they take effect.
— Expert teams run multiple experts in parallel and typically consume 3–5 times the credits of a single expert.
— The inspiration community allows one-click replication of works like dashboards and analysis templates, which the author says is almost always cheaper than starting from scratch.
— Some content sources lack callback APIs for new articles, so information collection must rely on periodic polling.
Conclusions

The article treats 'unorthodox' as a compliment, but the real power comes from reusing the user's own context: login state, local files, and markdown documents. That is a design choice many agent platforms do not expose, and it is what makes the browser-control and co-writing features genuinely useful.

The author's insistence on redaction rules, 'only write evidence', and 'do not guess unconfirmed information' shows a practical approach to AI hallucination: the automation is acceptable only because it caps what the model is allowed to infer.

Concepts & terms
Skill (in WorkBuddy)
A packaged set of best practices, consisting of a SKILL.md manual with YAML frontmatter, optional scripts, references, and assets. It turns a personal workflow into a reusable capability that triggers based on its `description` field.
MCP (Model Context Protocol)
A standard interface that lets an agent connect to external tools and data sources. Custom MCPs allow users to attach private resources like local note libraries or internal APIs that official integrations do not cover.
人机双写 (Human-AI Co-Writing)
A WorkBuddy feature where you select a portion of a document in-place and have the AI modify only that selection, preserving everything else. It works in Word, Excel, PPT, and Markdown.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗