moe-execution

moe-execution

MoE-aware coding skill + code-assistant instructions for Mixture-of-Experts models (Qwen, GLM, DeepSeek, Mixtral, local MoE).

Most "weak model" coding pain with MoE models isn't low intelligence — it's unstable expert routing. In an MoE model a lightweight router picks a small subset of "experts" for each token based on your phrasing. Paraphrase a concept, mix domains, or drift your wording, and a different set of experts takes over mid-task — so it forgets decisions, produces locally-correct-but-globally-broken edits, and follows instructions inconsistently.

This repo gives you three layers that fix that:

1. Code Assistant Instructions — an always-on block that re-anchors routing every turn (role + canonical vocabulary + one-domain-per-turn + map-before-edit + verify).
2. moe-execution skill — on-demand depth: routing stabilization, role-anchor table for software-dev concerns, state externalization (impact maps, plan files), and a done gate.
3. workflow-gate skill — workflow bootstrap (workflow-gate:init) and mandatory interface blast-radius discipline (workflow-gate:impact-analysis), then handoff to moe-execution for execution.

Recommended execution path (non-init):

1. planning (brainstorming / writing-plans)
2. workflow-gate:impact-analysis for interface-level changes
3. moe-execution for implementation/refactor/debug

Why MoE changes the game

Symptom	MoE root cause	Fix in this repo
"Lost the plot" mid-task	Wording drift re-routes to different experts	Canonical vocabulary, re-anchoring
Locally-right, globally-broken edits	Routing is per-token; no expert sees the whole task	MAP dependents, persist plan to file
Inconsistent instruction following	Paraphrasing activates different experts	Stable, structured prompt shape
Weaker than expected on a domain	Wrong experts activated	Front-loaded role + tech keywords

Works with any MoE model. DeepSeek is a common example, but the same routing dynamics apply to Qwen, GLM, Mixtral, Llama-4, and local MoE builds.

Install

Claude Code (plugin)

/plugin marketplace add uuie/deepseek-instructions
/plugin install moe-execution@moe-instructions

The GitHub repo is currently named deepseek-instructions; the marketplace it exposes is moe-instructions. If you rename the repo, update the marketplace add path accordingly.

Then add the always-on layer to your ~/.claude/CLAUDE.md:

@<path-to-clone>/instructions/code-assistant-instructions.md

…or copy the block from instructions/code-assistant-instructions.md directly into your CLAUDE.md.

Any other agent (plain files)

This repo works without the plugin system. Copy the instructions block into your agent's persistent instruction file and (optionally) the skill into its skills directory:

• Instructions → see the mapping table in instructions/code-assistant-instructions.md (Copilot CLI, Gemini CLI, Codex/AGENTS.md, Cursor).
• Skill → copy skills/moe-execution/ into your agent's skills folder (e.g. ~/.claude/skills/).

What's inside

.claude-plugin/
  marketplace.json          # marketplace manifest (for `/plugin marketplace add`)
  plugin.json               # plugin manifest (registers the skill)
skills/
  moe-execution/
    SKILL.md                # on-demand MoE execution skill
  workflow-gate/
    SKILL.md                # init + impact-analysis workflow skill
    scripts/
      init_engineering_workflow.py
instructions/
  code-assistant-instructions.md   # portable always-on block + per-agent mappings

The core idea in one line

You can't make the experts smarter — but you can keep the right experts consistently active, and write down the global state they can't hold.

License

MIT — see LICENSE.