Working Surfacelink · 8 Nov 2026
03
An architect's drafting table passes one rolled blueprint through a hatch into a separate carpentry workshop, while a small tray beside the hatch fills with folded question slips, one figure seen from behind at the hatch, the whole scene drawn with wide empty margins on every side.
Qiita

How development started to flow after separating the Claude that writes prompts from the Claude that implements

Reito Yamaura · 6 November 2026

How-to · handoff · agents · Agent context files

Read the original
Takeaway

An Offgrid engineer splits AI work into one session that writes the build instructions and another that builds, and forbids the agent to guess unclear business rules.

Summary

Reito Yamaura, an engineer at Offgrid, published an account in Japanese of a month building business software with Claude Code. Offgrid is a Japanese company that mainly places engineers on client projects and also takes contracted AI work. Giving requirements straight to the AI agent in the code repository mixed reading the specification with writing code, and the session filled up quickly. Yamaura now uses one session to write an implementation prompt and a second session to build from it.

Key points
  • The first session reads the requirement documents, screen list and database schema and writes the prompt, and a person reviews it before any code is written.
  • The prompt template carries rules added after failures: follow existing conventions, reuse shared logic, and list new tables or APIs before creating them.
  • The agent may not decide unclear business rules; it marks each one in the code and lists it in its completion report, which becomes the list for the client.
  • After one change touched thousands of lines in four files, every prompt now names the files the agent must leave alone.
Implication

A team building client software with AI agents can review a written implementation prompt before code exists, and make open questions a required output. This is one engineer's practice over one month, not a team study.

Plays
Product

Have the AI agent record unclear business rules as open questions in its completion report instead of deciding them, and take that list to the client.

Engineering

Use one AI agent session to write the implementation prompt from the requirements, and have a person review it before a second session writes the code.

Source issue

8 November 2026: teams using AI agents are moving human review from reading everything an agent produced to settling the decisions it hands back, with evidence attached.