AgentSkillsCN

high-agency-ai-building

制定一套框架，打破领导层的决策瘫痪，驾驭那些“所有选项都痛苦”的“深渊式”抉择。适用于面临战略转型、考量高风险的人事变动，或在危机时刻，当没有一条显而易见的“正确”路径可循时。

SKILL.md
--- frontmatter
name: high-agency-ai-building
description: Leverage agentic IDEs to prototype, build, and refactor software. Use this skill when you need to build internal tools without dedicated engineering resources, when a PM wants to ship a functional PR instead of a mockup, or when refactoring complex logic across large codebases.

High-agency building shifts your role from a manual coder to an architect and reviewer. By using agentic IDEs, you bridge the gap between business intent and technical implementation, allowing non-engineers to build functional software and engineers to automate 90% of their output.

The Agentic Building Workflow

1. Initialize with Context

Don't start from a blank slate. Provide the AI with the environment and visual constraints of your project.

  • •Start with Boilerplate: Initialize a basic project (e.g., React, Python) and let the agent install dependencies.
  • •Visual Prompts: Upload screenshots or sketches of your desired UI. The agent can translate image elements into functional components.
  • •Index the Codebase: Ensure the tool has indexed your entire repository (even 100M+ line codebases) to understand local patterns and dependencies.

2. Issue Explicit Instructions

Treat the AI as a high-capacity intern. Being vague leads to irrelevant code changes.

  • •Be Ruthlessly Specific: Instead of "make it look better," say "change the background of the hero section to #FF0000 and add a 'Book Now' button that links to /checkout."
  • •Start Small: Do not ask for a directory-wide refactor in one go. Build one component or one function, verify it, and then expand.
  • •Use Point-and-Click: Use the IDE’s previewer to select specific UI elements and issue commands directly against them (e.g., "Make this specific button retro-style").

3. Review and Refine

In an agentic world, your primary job is Reviewing.

  • •Validate Intent: Check if the agent’s "Plan" matches your goal before it starts writing files.
  • •Check Logic, Not Just Syntax: The AI will get the syntax right, but you must ensure the business logic (e.g., the partner portal's permissions) is correct.
  • •Human-AI Collaboration: If you don't like a variable name, change it manually in one file. A high-agency IDE will recognize your intent and offer to update that variable across the rest of the project automatically.

4. Deploy and Cannibalize

  • •Ship Internal Tools: If a SaaS product doesn't exist or is too expensive for a niche need (e.g., a custom partner portal), build it yourself in a few hours.
  • •Iterate Every 6 Months: Expect your tech stack or form factor to look "silly" within a year. Use the agent to completely refactor or rebuild the UI as your needs evolve.

Examples

Example 1: GTM Internal Tool

  • •Context: A Head of Partnerships needs a custom portal to track lead sharing because the current CRM is too clunky.
  • •Input: A hand-drawn sketch of a table with "Partner Name," "Lead Status," and "Last Contact."
  • •Application: Upload the sketch to the agentic IDE. Instruct: "Build a React app using this layout. Connect it to our existing Airtable API for the data source."
  • •Output: A functional, private web app deployed internally, saving the company $15k/year in SaaS fees.

Example 2: PM Feature Migration

  • •Context: A PM wants to rename a core metric from conversion_rate to signup_efficiency across the entire codebase to align with new branding.
  • •Input: The production repository.
  • •Application: Manually change the variable name in the main config file.
  • •Output: The agent detects the change and prompts: "I noticed you changed this variable. Should I refactor all 42 instances across the backend and frontend?" The PM clicks "Apply" and creates a PR.

Common Pitfalls

  • •The Kitchen Sink Request: Asking the AI to "build an Airbnb clone" in one prompt. This leads to hallucinated dependencies. Fix: Build the landing page first, then the search bar, then the booking logic.
  • •Ignoring the Review Flow: Accepting large code changes without reading the diff. Fix: Use the IDE's "Review Mode" to step through every file modification the agent made.
  • •Lack of "Dehydration": Hiring a developer to build something that a high-agency PM could have built in an afternoon with AI. Fix: Only hire when the person is "underwater" and the task truly requires deep architectural problem-solving that AI cannot yet plan.
  • •Over-reliance on Frontier Models: Assuming a general chat tool (like ChatGPT) is enough. Fix: Use a tool that has "Local Context" (indexing your specific files) to avoid sending billions of tokens to a cloud model.