Brand Profile and Advise: Teach the Agent Your Style

Two small mechanisms keep an agent consistent with your publication's voice without magic.

Brand Profile: free-form, never parsed

md2wechat brand init
md2wechat brand show --json

brand init creates ~/.config/md2wechat/brand.md (idempotent). The file is free-form Markdown — voice, theme and layout preferences, conversion rules, whatever your brand needs. The CLI never parses it; it exists for the host agent to read as context, so your rules travel with every session instead of living in one chat's memory.

Advise: a deterministic router

md2wechat advise article.md --json

advise is deterministic and side-effect-free. It returns either no_enhancement_needed or enhance_minimally with at most 3 actions and at most 3 layout modules, drawing from the tools title, cover, layout, and micro_edit, plus concrete micro_edits. Because the routing is deterministic, the same article always gets the same advice — safe to gate an automation step on it.

A sane loop

1. brand init once; write your voice and preferences.
2. Draft the article.
3. advise --json — apply only the proposed edits.
4. title suggest --json and a cover plan if advised (Title Generation, Image Generation).
5. convert --preview, then the draft flow on explicit approval (Publishing API).

The point is restraint: at most three changes per pass keeps the article yours.