The framework

GAD is planning + evaluation for AI coding agents.

A small CLI, a strict five-step loop, and an experiment harness stapled to the side. The CLI re-hydrates context in one command. Skills tell the agent what to do. Subagents do the expensive work off the main thread. Templates scaffold new projects. Runtime-specific command wrappers are generated only when a coding agent needs them. Evals measure whether any of this actually helps. Read the loop diagram for how it fits together.

Core concepts

Three moving parts

93
Skills
Methodology docs the agent follows. Official consumer skills live under sdk/skills and are the installable public surface; root skills are internal methodology for evolving GAD itself.
-> browse skills
21
Subagents
Specialised workers the framework spawns for planning, research, verification, UI audits, and more. Subagents receive a task plus context and return a concrete artifact such as PLAN.md, RESEARCH.md, or VERIFICATION.md.
-> browse subagents
44
Templates
The planning-doc scaffolding and workflow templates the CLI uses to bootstrap new projects. These live under sdk/templates and ship with the framework's runtime assets.
-> download pack

Skill bootstrap sets

Framework-level vs eval-inherited

GAD ships 93 official skills as the canonical consumer/runtime surface. But eval projects (bare, emergent) do not get the full framework - they start with a minimal bootstrap set copied into their template/skills/ directory. The rest of the framework is withheld by design so we can see what they build without it.

bootstrap

Inherited by bare + emergent (10)

create-skill

>- Capture a reusable pattern, recipe, or failure-mode fix as a skill document so future agents (including you after a context reset) can apply it without rediscovering it. Use this skill whenever you solve a non-obvious problem, discover a working pattern after two or more failed attempts, hit a bug whose fix isn't self-evident from the code, or finish a piece of work that future runs will likely repeat. Write the skill the moment you learn the lesson — not at the end. In bare/emergent eval conditions this is the primary mechanism for agent-authored methodology. The agent IS the workflow author, and skills are how that authorship persists.

inherited by: escape-the-dungeon, escape-the-dungeon-bare, escape-the-dungeon-emergent, escape-the-dungeon-gad-emergent, escape-the-dungeon-planning-only

merge-skill

>- Fuse two or more existing skills into a single tailored skill. Use when you discover overlapping skills in the catalog (same trigger keywords, same target workflow, same failure modes being addressed) or when you want to take a fundamental skill like scientific-method or debug and attune it to a specific project domain (e.g. merging scientific-method into a kaplay rune-spellcrafting workflow produces a scientific-method-for-kaplay-forge variant). Merging is the preferred response to skill collision — do not let duplicate skills accumulate in the catalog. Merge them, deprecate the originals in CHANGELOG, and update any inheriting runs. Per decision gad-73 this is one of the three fundamental skills (find-skills + merge-skill + create-skill) that together enable the emergent-evolution hypothesis to be tested in practice.

inherited by: escape-the-dungeon, escape-the-dungeon-gad-emergent

framework-only

Available to full GAD runs only (83)

Client debug · NEXT_PUBLIC_CLIENT_DEBUG=1
0 lines

No events yet. Window errors, unhandled rejections, and React render errors appear here. Set NEXT_PUBLIC_CLIENT_DEBUG_CONSOLE=1 to mirror console.error / console.warn.