hhow09's Blog

AI productivity boost is real

I was a bit skeptical about AI coding assistant at first. But since Skills came out, I can really feel the power of agentic workflow. My workflow is still heavy human in the loop but I can certainly get better result with better custom rule, templates.

I use Cursor but these features should also work in general agents e.g. Claude Code.

1. Plan Mode #

Plan Mode is extremely useful to explore the complexities between different solutions before making any changes.

Here is how I used it:

  1. Plan
    • Seperation of concerns
      • SPEC.md: user story, requirements, etc. which will discuss with PM.
      • TECHNICAL_DETAIL.md: architecture, data flow, api design, complexity analysis, etc. which will discuss with technical roles.
      • IMPLEMENTATION.md: implementation plan, which will be updated throughout the implementation process.
    • use SPEC.md and TECHNICAL_DETAIL.md to Understand the problem / spec by asking
      • "Given the spec from jira ticket, understand the current user story with following references: . Provide flow chart."
      • "Grill me with specs I need to clarify with product manager. List out possible edge cases."
      • "Explain the current architecture of , . Draw a diagram to visualize the flow and architecture."
    • use IMPLEMENTATION.md to Break down into smaller tasks by asking
      • "Break features into small, focused tasks"
      • "What's the suggested priority of these tasks in order to minimize the risk of breaking existing features?"
      • "What's the required changes for each task?" to understand the complexity for better estimation.
    • Context: Always provide relevant context: specific code reference, document.
  2. Execute first few steps for quick validation: Execute the initial steps to validate the approach before committing to the full implementation.
  3. Implementation: Update the plan accordingly every round based on the real AI + human implementation.
  4. Futher usage: keep the plan for futher use: decision making reference, QA testing instructions, tech sharing materials, work log, etc.

2. Agent Skills #

open standard for extending AI agents with specialized capabilities.

I used #

3. Rules #

Project based rules to provide contex automatically in chat.

Use cases: general coding guidelines, code structure, testing style

examples: awesome-cursorrules

4. Normal Chat / Others #

5. MCPs #

Other tools to try #

What are still heavily human involved #

Thoughts: Bottleneck shifted #

Bottleneck of delivery shifted from coding to review and other processes (testing, external dependency, release process), as stated in several posts (1, 2). New challenges already emerged: large amount of low quality PR from unexperienced engineer, human QA as relase blocker...

Ref #