Skip to main content
Back to Insights
Method8 May 20266 min read

What Ganjiang Forges Isn't Steel — It's a Soul for an AI Agent

"Forging a soul for every Agent" isn't a slogan. It corresponds to four concrete Markdown files, an Enneagram-based algorithm, and an engineering practice that moves personality out of the system prompt and into the identity file. This essay walks through what actually happens inside Ganjiang's forge.

Ganjiang's pitch is one sentence: "Forging a soul for every Agent." Short like a slogan. But behind the short sentence is an entire engineering practice. This essay opens it up.

The layer prompts can't write

You write a prompt: "You are a careful code reviewer. Prioritise security. Be terse." Run it ten times — about eight follow you. The other two over-pick on irrelevant details, miss real security holes, or suddenly turn chatty. Prompt engineering can shift that ratio. It can never reach 100%.

Because a prompt is an external rule, while every decision an agent makes needs an internal disposition. Rules tell it what to do. Disposition decides what it does when the rules don't cover the case. What Ganjiang forges is that disposition — a self stable enough to survive model upgrades, context truncation, and prompt rewrites.

Four files that make a soul

What comes out of Ganjiang's forge isn't a single file but a four-file standard set:

  • soul.md — "who this is": Enneagram type, wing, instinctual stack, plus stress / growth directions. The agent's personality DNA.
  • identity.md — "what it knows": background, expertise, context memory. A portable CV that travels between tools.
  • user.md — "who it knows": your preferences, working style, hard constraints. The agent's model of its user.
  • agents.md — "how it writes": prose style, phrasing habits, decision tone.

Together: roughly 600–1,500 words, plain Markdown, your data. Drop into any project under .claude/, .cursor/, or agent.md and the tool picks them up immediately.

What happens inside Ganjiang's forge

When you start a forge run on ganjiang.xyz, the system does this: maps your answers / screenshots / dialogue into one of the 27 Enneagram subtypes, infers the wing and instinctual stack, fixes the integration / disintegration directions, generates tool preferences that match your workflow, and finally renders all of that into four readable Markdown files. The whole process happens between your browser and our Shanghai ECS — no cross-border data transfer.

Unlike traditional prompt engineering, the soul that comes out isn't "works this time" — it's "still itself when the model changes, the tool changes, the project changes." That's the real meaning of the name Ganjiang: smelt the soul into the tool, so the tool has its own temperament.

Prompts manage this conversation. A soul manages whether, ten years from now, this agent is still itself.