About Me
A college student who likes getting hands-on and prefers practice-oriented tasks — rapid prototyping, front-end design, UI design and other hands-on kinds of design. Finding a design flaw in the middle of an interaction and then polishing the details is a joy like no other. I also think of myself as a perfectionist, though I am not sure whether that is a good thing or a bad thing (awkward laugh); it happens to match what the new agent-assisted coding model needs — taking a rapid prototype through acceptance, testing and detail polishing until it reaches product quality — and it matches the identity shift programmers are going through in coding: from coder to reviewer and optimizer. I use agent tools fluently, such as Claude Code, Codex and DeepSeek Harness, and I have also raised a "lobster" (OpenClaw) and a "horse" (Hermes); I understand some of the basic working principles of agent frameworks, and I am still learning the finer professional knowledge. I am good at designing prompts — turning a real situation into a prompt that an LLM grasps at a glance so the agent understands the requirement constructively, which cuts down on trial and error — and at planning the agent's "construction blueprint". Combined with software-engineering thinking and agent automation workflows, I only need to take part at the two ends of a task, design and review, to get a basically usable prototype (raising the code quality further takes finer optimization). I understand what agent tools such as MCP, file tools and RAG knowledge bases do and how much they matter, and I understand just as well how important it is to tell an agent about each tool's properties (when to use it, how to call it, its schema); I have a complete agent-collaboration coding system of my own (global coding rules, automated review workflows, and more). Frontier AI and agent technology still feels fresh to me, and I am especially excited about personally customized, AGI-like agents such as OpenClaw — how that category is developing, and the realization of a future all-purpose personal assistant! AI is getting closer and closer to growing "limbs" out of a "brain" and actually doing work.
My AI Coding Trajectory
AI Completion and Local Assistance
The trajectory begins with code completion and local generation. Here AI assists: completing code, suggesting between the lines. Authorship stays with the coder, who writes and reviews line by line, and whether each implementation holds is still a human judgment.
AI becomes a coding accelerator
Output Stuck in the Chat, Manual Integration and Debugging
AI can produce a complete file in one pass, but the file still has to be placed by hand, fitted into the existing structure, run, and corrected error by error. Coding gets faster; context does not hold together, so debugging and verification cost more.
The center of gravity moves from writing code to describing needs, integrating, and debugging
Agentic Coding: From Task to Automated Tests
A task is handed over whole: the agent reads context, rewrites files, runs the project, writes tests, and iterates until they pass. I hold the two ends — setting direction and boundaries at the start, reviewing and accepting at the end.
The role shifts from author to direction-setter and reviewer
Settled into Workflows and a Development System
Coding rules that keep growing, tools (MCP, CLI, RAG knowledge bases), agent memory and ways of collaborating settle into a personal development system and workflow that are reusable and can run in parallel.
From single tasks to a reusable, parallel personal development system
My Way of Using Agents
Rather than setting an agent loose, we write the development document together: requirements analysis, module design, task breakdown, tool configuration, acceptance criteria, following a software-engineering process shaped for agents. What comes out is one document read two ways, structured enough for an agent to build from and complete enough for a person to audit.
Define the Problem
Confirm the goal, boundaries, and acceptance criteria before moving in the wrong direction.
Draw the Blueprint
Break requirements into tasks, dependencies, and expected outcomes with clear context for the agent.
Execute with Agents
Use the right models, tools, and sub-agents for each task while continuously realigning the direction.
Verify and Deliver
Automated tests run alongside manual ones: check the code and the features, keep the evidence, and iterate from there.
My Capabilities & Direction
From current focus to practical output, and the way I organize agents around real tasks.
Application & Mastery
- Prompt design
- Rapid prototype output
- Building Agent workflows
Management & Orchestration
- Goal definition / requirement clarification / acceptance criteria
- Task decomposition / Agent delegation / documentation constraints
- Process review / quality verification / continuous iteration
Explore and Develop
- Agent Automation Workflow
- Rapid Prototyping Output
- Explore AGI
My Interests & Exploration
Agent Development Teams
What I want to reproduce is not a single agent but the way a development team operates: requirements, design, implementation, testing, and review each have an owner, and tasks pass between stages with explicit handoffs.
The Agent as Creator
Push the edge of making inside the frame: a concept-driven image, a walkable three-dimensional scene, a complete creative video. The hope is for an agent to own the whole path from idea to finished piece, not one link in it.
Toward AGI
The goal is not a stronger model but the all-purpose assistant of science fiction: one that understands intent, executes tasks, and creates when creation is needed. Understanding, execution, and creation in one place — that is what the word assistant should mean.
Embodied Agents
Let intelligence step out of the screen and gain a body: observing its surroundings, acting when it is needed, learning and growing across long use. Not only a tool, but a companion.