Hands‑On AI Coding: My Real‑World Claude Code Project Experience

Preface

Facing a series of urgent backend tasks and a shortage of staff, the author decided to let Claude Code write the code instead of typing manually.

Environment Setup

Claude Code, powered by Zhipu's GML‑5.1 model, was used together with an IDE for quick file browsing and execution. The Superpowers Skills suite was installed (see the referenced Skills article). The programming language was Go, and the base framework was go‑zero.

Coding Results

Claude Code generated the following components:

CLAUDE.md to guide subsequent work

Integration of the Milvus vector‑database SDK

A business table schema for vector data

CRUD business logic and API documentation for the table

Embedding‑model integration for vectorization during Milvus storage

The entire set of features was completed in roughly three to four hours, consuming about 200k‑300k tokens. At a rate of 10 CNY per million tokens, the cost was approximately 2‑3 CNY.

The output quality was satisfactory, but the author still needed to understand the underlying framework and libraries to achieve better results.

Overall efficiency was comparable to a mid‑senior developer, with the token cost being only a few yuan; however, the human "commander" still played a crucial role.

AI Is Still Slow

The dominant feeling during development was slowness, caused by waiting for the AI and repeated confirmations.

Two main reasons were identified:

The Superpowers Skills workflow enforces a Socratic questioning phase, planning, test generation, and verification, making each step serial and thorough, which prolongs the process.

Each step requires manual confirmation, especially for risky operations such as script execution or pulling code.

The model itself may also have inherent latency.

Solution Attempts

The author chose not to alter the Superpowers workflow because its quality guarantees outweigh speed concerns.

Task Parallelism

To improve efficiency, Claude Code's support for Git worktree was leveraged to open multiple windows and run tasks concurrently—for example, one window handled SDK integration, another designed the table schema, and a third integrated the embedding model. This parallelism turned serial tasks into parallel ones.

Reducing Confirmations

Claude Code provides several operation modes: default: read‑only acceptEdits: read, file edit, and common file‑system commands (mkdir, touch, mv, cp) plan: read‑only auto: all operations with backend safety checks (requires Claude Code v2.1.83+ and a Max/Team/Enterprise account) dontAsk: pre‑approved tools only bypassPermissions: all operations with backend safety checks (high risk)

During a session the author cycles modes with Shift+Tab, moving through default → acceptEdits → plan. The personally preferred mode is acceptEdits. The ideal auto mode is unavailable without a newer Claude Code version or a paid account. The bypassPermissions mode is considered risky because the model may request access to local resources such as photos or contacts.

Conclusion

The AI‑assisted coding experiment produced good results and dramatically compressed development time, provided the "commander" has sufficient technical experience. Speed issues can be mitigated by splitting tasks and using permissive modes when safe. Further tips and experiences are still being explored.