Trellis is an agent skill I wrote to scaffold the way I want to work with an AI coding agent on non-trivial features. I had been noticing two specific failure modes recurring in my Claude sessions: the model would happily fabricate design decisions instead of surfacing the choice back to me, and it would balloon a single change into a sprawling unreviewable PR if I didn’t keep yanking it back to scope. Trellis is a structured workflow that pushes back on both.
The skill organizes any feature into three explicit layers, each with its own create/iterate/review operations:
design.md. Decisions
get rationale; open questions get severity tags ([blocks-v1],
[deferred]); explicit “complete” / “not-yet-complete” gates keep the
agent from moving on prematurely.reviews/.The skill is manually invoked — Claude will not “helpfully” jump into planning mode unprompted. A typical session looks like:
Use trellis to start a new design plan at
docs/features/refunds/
…and then a few rounds of iteration on open questions, optionally an external-agent review pass, then bootstrapping implementation, then:
Use trellis to execute Sprint 02 step 3 through 5 in
docs/features/refunds/02-manager-core.md
Each execution step spawns its own subagent, implements one step, gets reviewed by another subagent, and produces an isolated commit. The artifacts left behind are real markdown files that survive the session and serve as the durable record of what was decided, why, and what was done.
It’s architecture-agnostic — backend, frontend, CLI, library, infra — and
inherits project conventions from whatever CLAUDE.md or AGENTS.md is
already in the repo.