Which AI Coding Tool Solves Control Loss and Forgetfulness: mattpocock/skills vs Trellis

Background and Pain Points

When using Claude Code or Cursor for AI‑assisted programming, developers often encounter two problems: the generated code becomes increasingly difficult to control and verify, and each new session forgets the project’s architecture, naming conventions, and technology choices.

Projects Compared

The two GitHub projects that aim to solve these problems are mattpocock/skills (a collection of TypeScript‑based AI coding skills released under the MIT license, installable with npx) and Trellis (an out‑of‑the‑box engineering framework from Mindfold, distributed as the npm package @mindfoldhq/trellis v0.6.6 under AGPL‑3.0).

Target Problems

mattpocock/skills positions itself against frameworks such as GSD, BMAD, and Spec‑Kit that “take away your control and make bugs hard to resolve.” Its four failure modes are misalignment, verbosity, no feedback loops, and “ball of mud.” The solution is a set of small, composable skills that force the engineer to think before acting, share a domain model, write tests, and conduct code reviews.

Trellis defines its single failure mode as “every session starts from scratch.” It persists specifications, tasks, and memory in a .trellis/ directory and drives development through a four‑stage loop.

Architectural Philosophy: Toolbox vs Framework

mattpocock/skills provides a toolbox of independent markdown‑based skill files. Each skill lives in its own SKILL.md and is invoked manually (e.g., /grill-with-docs or /implement).

Trellis is a full project‑management system with a CLI, a defined directory layout, state machines, hooks, and sub‑agent scheduling. The framework orchestrates the workflow while the developer makes decisions.

Core Capability Comparison (Six Dimensions)

Dimension            mattpocock/skills                     Trellis
-----------------------------------------------------------------
Architecture         Loose skill files + router            CLI + directory conventions + hooks + sub‑agent scheduler
Philosophy          You arrange the tools                 Framework arranges, you decide
Requirement Alignment grilling (independent primitive)   trellis‑brainstorm (task‑system entry)
Specification Management CONTEXT.md (glossary + ADR)   .trellis/spec/ (layered, package×layer)
Task Management      External tracker (GitHub, Linear…)   Built‑in task.py with lifecycle hooks
Cross‑session Memory handoff (manual markdown)        journal‑N.md (automatic, rotating)
Implementation Check  15‑line implement skill (doc‑driven) sub‑agent enforce, no git access, guarded execution
Platform Compatibility Claude‑only (symlink)          17 platforms via configurators
Enforcement          Optional (skill‑driven)               Mandatory per‑turn breadcrumb injection

When to Choose Which

Choose mattpocock/skills if you work mainly with Claude Code, are an individual or small team, already have an issue‑tracker, prefer TDD, domain modeling, and want lightweight discipline without a heavy framework.

Choose Trellis if you need multi‑platform support (Claude, Cursor, Codex, etc.), require shared specifications and session logs for a team, have complex, layered projects, and are willing to let the framework orchestrate the development flow.

Combined Use

In theory the .trellis/ directory can coexist with the skill files from mattpocock/skills, allowing you to use Trellis for project structure and task flow while importing the grilling, TDD, and code‑review skills.

Observations on Industry Trends

Explicit requirement alignment (grilling/brainstorm) is becoming a standard first step for AI coding.

Sub‑agent patterns are solidifying; Matt’s skills use prompt‑based parallel agents, while Trellis implements process‑level scheduling.

Managing spec/context lifecycle is a key challenge; both projects propose different solutions (single‑file CONTEXT.md vs layered spec with precise JSONL injection).

Cross‑platform compatibility is a hard barrier; Trellis’s 17‑platform configurators illustrate the effort required.

Final Guidance

If your main pain is low code quality, adopt the discipline‑oriented toolbox (mattpocock/skills). If the pain is repeated forgetting of project context, adopt the management‑oriented framework (Trellis). Most real‑world scenarios involve both layers, but weighing flexibility against consistency will guide the right choice.

Note: This analysis is based on the source code of mattpocock/skills (v1.1.0) and Trellis (v0.6.6). Both projects are under active development; features may evolve.