Briefing

Strategy as Code: Using MCP Tools to Enforce LLM Constraints

ai-dev
by HIROKAZU YOSHINAGA ·

Patch your LLM agent to load STRATEGY.md via MCP tools so constraints are enforced.

What to do now

Patch your LLM agent to load STRATEGY.md via MCP tools and test constraint enforcement.

Summary

The article introduces the concept of writing marketing strategy as code in a Markdown file that an LLM agent can load before acting, addressing the problem of generic best‑practice advice that ignores business specifics. It explains the six‑section structure of a STRATEGY.md file—Persona, USP, Brand voice, Goals, Operation mode, and Constraints—each of which becomes a constraint the agent must obey. The author shows how the mureo v0.8.0 release (2026‑05‑02) ships five MCP tools that expose the strategy file and state JSON to the agent, allowing read, write, and state updates via commands such as mureo_strategy_get and mureo_state_upsert_campaign. By loading the strategy through MCP, the agent can reference explicit constraints like “no competitor‑name bidding” or “Meta CPA must stay below ¥4,500” and refuse to propose actions that violate them. The article also demonstrates a real‑world scenario where the agent detects a 50 % sibling‑channel divergence and recommends a rescue audit instead of pausing a campaign, illustrating the practical benefit of contract‑based reasoning. Finally, it argues that this pattern generalises to other domains—SRE incident response, investment theses, threat models—by simply swapping the strategy file for a domain‑specific contract.

Key changes

  • v0.8.0 release (2026‑05‑02) introduces five MCP tools: mureo_strategy_get, mureo_strategy_set, mureo_state_get, mureo_state_action_log_append, mureo_state_upsert_campaign.
  • STRATEGY.md file is loaded into the agent via MCP, turning business rules into enforced constraints.
  • The agent can read, write, and update strategy and state JSON through MCP commands.
  • Example constraint: “no competitor‑name bidding” triggers a rescue audit instead of pausing campaigns.
  • The agent references explicit constraints like Meta CPA target and operation mode to guide decisions.
  • The pattern generalises to other domains (SRE, investment, threat modeling) by swapping the strategy file.
  • The agent logs actions to STATE.json, ensuring all changes are tracked.
  • The MCP integration allows any MCP‑compatible host (Claude Desktop, Cowork, web UI) to access the strategy file.

Affects

ads-customers internal

Customer impact

Analyzing matches…

Ask about this story

Impact on an agency? Which customers? Compare historically Risks of waiting