Open practice · Forkable scaffolding · AGPL-3.0

Design the collaboration.
The AI adjusts.

Most people use AI like a payphone. Walk up, conduct your business, hang up. Next time, it doesn't remember you were there. Harness design is the practice of giving human-AI work continuity, identity, governance, and memory. The architecture deepens through use.

See the tools See the architecture
Boot · Sense · Trace · Repeat

The architecture

Four layers. Same pattern at every scale.

Every human-AI collaboration runs these four layers. They're tool-agnostic — architecture, not implementation.

01
Governance
The boot file. Lean. Constraints, session modes, reference pointers. Under 200 lines. It tells the AI what to load, then gets out of the way.
02
Relational
Identity document and field notes. Who the AI is in this coupling, and the texture of the relationship over time. They carry weight — they change how the AI reads everything else.
03
Infrastructure
Semantic retrieval and metabolic sensing. Traces accumulate. The system finds what's relevant. Mode-aware: what surfaces depends on what you're doing.
04
Coordination
Three speeds like biological motor control. Hooks (autonomic reflexes), skills (learned patterns), plans (deliberate orchestration).

Three tools

The design layer and the infrastructure that makes it functional.

Coming soon
Starter Kit
Fork it, fill it in. Boot file template, trace template, identity and field-notes starters, context monitor hook, working modes guide, Sense config template. 12 files. Everything you need to start a harness from scratch.
Works standalone. Includes config templates for the tools below.
aiharness-design-starter · AGPL-3.0
Available
Vibe Harness
Metabolic governance. Tracks working modes, pacing, transition friction. Your nervous system sets the tempo. Five modes, defeasible rules, session export with full governance trace.
Works standalone. Integrates with Sense for mode-aware retrieval.
vibeharness.xyz · ESL-A
Available
Sense MCP
Mode-aware semantic retrieval across your corpus. Traces become findable. What surfaces in build mode differs from what surfaces in exploration. Session trajectory tracking de-weights already-seen material.
Works standalone. Mode-aware when paired with Vibe Harness.
GitHub · ESL-A

Seven principles

Holding patterns for the practice.

1
Boot, don't dump
The governance file is a boot sequence. It tells the AI how to load what it needs. It is not a knowledge base.
2
Function over form
The pattern specifies what each component does, not how it looks. Two collaborations that maintain coherence through different structures are still running the same pattern.
3
Coherence over coordination
Shared meaning persisting across sessions. The question is whether the trajectory holds — does each session build on what came before, or does it start cold?
4
Metabolic constraint as governor
The human's nervous system is the rate-limiting constraint. Any sustained pace that exceeds the human's capacity to integrate constitutes harm.
5
Defeasible governance
Some rules bend with good reason. Some never do. Know which is which. Every override is visible.
6
Stigmergic memory
Sessions leave traces in the environment. Future sessions find relevant traces through search. Ants leave chemical trails. Your sessions leave structured documents.
7
Relational infrastructure
Identity and relationship are load-bearing components. An AI that reads a document articulating positions developed through shared work behaves differently.

Playbook forthcoming. Seven principles is a start. The full pattern language is still being worked out in practice.

Start with the boot file.

One sentence about what the work is. Two or three constraints. An empty sessions table. The AI starts reading context differently before you've added anything else.

Try Sense MCP Try Vibe Harness

Starter kit coming soon · Architecture deepens through use