EP127 从 Skills 到自动化工作流，论 Agent 如何接管真实生产力 ⚙️

Summary

This 硬地骇客 episode extends AI Skills from plugin recommendations into a practical operating layer for Agentic Workflow. The hosts argue that useful skills should be chosen by repeated work, weak domains, and verification value, not by popularity. Coding examples center on requirements clarification, architecture maps, Playwright or computer-use validation, TDD, review, release, and production checks; non-coding examples with Podwise, notes, 微信读书, email, traffic analytics, server-cost monitoring, and investment tracking introduce Routine Agent Automation.

Key Claims

• A skill is worth installing when it solves work that repeats daily, weekly, or monthly, or when it covers a domain where the user lacks taste or experience.
• Project-specific skills can live inside a repo when they encode stack, design-system, testing, CI/CD, deployment, or codebase conventions; global skills should be kept more selective.
• The best moment to create or search for a skill is often after seeing where an agent repeatedly gets stuck or where the user keeps writing the same prompt.
• In coding, verification-oriented skills may be more valuable than framework knowledge because they help the agent run the product, test behavior, inspect diffs, and close the feedback loop.
• “Grill me” style requirement questioning is treated as a pre-coding skill: clearer user stories, acceptance criteria, and context files reduce long-running agent drift.
• The described Codex workflow moves from discussion to planning, implementation, self-review, fixes, commit/release, and live verification, with agents.md and skills enforcing smaller checks along the way.
• When building from zero, one speaker asks AI to propose the tech stack and architecture map first, then checks whether later diffs stay inside the right modules rather than reviewing every line for style.
• Visual landing-page work can begin with generated reference images and asset extraction before Claude Code or Codex turns the design into responsive frontend code.
• Non-coding skills are framed as personal infrastructure: podcast transcripts, reading notes, email triage, cost monitoring, analytics, industry tracking, and portfolio news can become scheduled agent work.
• Popular skill packs can still be deleted if they do not match the user’s actual work; too many skills may make agent behavior noisier and consume more context.
• Cross-agent review, such as having Codex write and Claude Code review, can raise confidence, but responsibility for bad agent output remains with the human operator.
• The episode applies a DRY-like rule to agents: when a workflow is repeated enough times, it can become a skill instead of another manual prompt or shell script.

Key Quotes

“现在不写提示词，而是写 loops” — how the coding discussion reframes agent work around repeated feedback cycles.

“重复三次以上” — the practical threshold used to decide when a repeated workflow should become a skill.

Connections

• 硬地骇客 — show context for the episode.
• AI Skills — core topic; skills become reusable workflow, verification, and automation packages.
• Routine Agent Automation — new concept for scheduled, recurring, low-glamour agent work.
• Agentic Workflow — the broader pattern of agents using tools, context, and checks to complete real work.
• Agent Harness — skills, computer use, CLI tools, permissions, and review loops sit inside the harness.
• AI Coding Verification and AI Engineering Thinking — coding skill value is tied to TDD, real testing, review, and release checks.
• Podwise — sponsor and recurring example of podcast-to-knowledge workflows.
• Codex, Claude Code, Open Cloud, and Superpowers — agent environments and orchestration references.
• Playwright — browser automation example for agent-driven verification.
• 微信读书 and Data Portability And Sustainable Tools — reading notes and personal knowledge assets as skill targets.
• AI Investment Research — investment monitoring skills are useful only when paired with user judgment and domain understanding.
• Human Judgment Under AI and Agent Permission Boundaries — trust, privacy, permissions, and responsibility remain human concerns.

Contradictions

• No direct contradiction with prior wiki content. The source reinforces the existing Agent-Optimized CLI, AI Skills, and Agent Harness threads while adding a stronger selection rule: a skill should survive only if it serves repeated work, weak-domain support, or real verification.