emergent-thinking is a project-local skill + CLI for coding agents.

In plain language: it tries to stop the model from improvising everything in its head, and forces it to write the work down, validate against reality, and iterate in loops.

emergent-thinking is a local board workflow for AI agents.

It combines:

• a skills-compatible skill that tells the agent how to work
• a project-local CLI that stores the board and exposes the protocol

The skill is the working method. The CLI is the external memory and move engine.

This is intentionally not a global MCP memory service. The natural scope is the current repository and its cwd.

Most coding agents fail in a predictable order:

1. they form a rough intuition
2. they start implementing too early
3. they discover conflicts too late
4. they patch over the mess
5. entropy keeps increasing

emergent-thinking tries to change that order.

The core loop is:

1. write the problem down
2. write the goal down
3. choose one bounded path
4. validate against reality early
5. record findings and proof debt
6. repair the cheapest meaningful stage
7. loop until the board converges

So the promise is not "the model becomes smarter". The promise is "the model is pushed into a more reliable working order".

Install it with:

npx emergent-thinking install

or:

bunx emergent-thinking install
pnpx emergent-thinking install

The normal usage is not "manually type lots of CLI commands".

The normal usage is:

1. install the skill
2. tell your coding agent to use emergent-thinking
3. let the skill drive the workflow
4. let the CLI persist the board inside the current project

If you want the actual operating protocol, read:

• skills/emergent-thinking/SKILL.md
• skills/emergent-thinking/references/board-protocol-rules.md
• skills/emergent-thinking/references/how-to-use-subagents.md

If you want contribution, benchmark reproduction, or protocol iteration guidance, read:

• CONTRIBUTING.md

Latest checked-in protocol_micro summary from evals/inspect_ai:

• On anthropic/glm-5.1, baseline scored 0.875, sequential scored 0.875, emergent scored 1.0, and emergent_compact also scored 1.0 while using 1,876 fewer total tokens than emergent.

This is not a universal proof. It is one concrete signal that the protocol can improve move ordering, and that prompt compaction can sometimes preserve quality while reducing cost.

• Skill: skills/emergent-thinking/SKILL.md
• Core protocol rules: skills/emergent-thinking/references/board-protocol-rules.md
• Subagent extension: skills/emergent-thinking/references/how-to-use-subagents.md
• Contribution and benchmark reproduction: CONTRIBUTING.md