跪拜 Guibai
← All articles
Frontend

OpenSpec, Spec Kit, Harness: A Clear Framework for Choosing AI Development Tools

By J船长 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Ignoring the layer distinction means duplicated rules, missed requirements, and session resets regardless of which tools you install. A focused fix lets teams get value from AI agents without a full stack of overlapping software.

Summary

Using Codex, Cursor, and Claude Code together spawns three distinct failure modes: per-tool rule inconsistencies, AI missing stated requirements, and lost progress after session switches. They need different fixes. OpenSpec and Spec Kit organize specs and development flow; Harness is the execution concept covering context, tool calls, progress, and verification; Rulesync and Ruler sync config across tools; Trellis is a team-level harness.

Choosing among them comes down to which problem dominates. Spec Kit suits teams that want explicit phase-gated workflows; OpenSpec suits change-centric iteration on an existing product; Trellis becomes necessary when session and team context keeps fragmenting; Rulesync handles rule distribution. A pilot with a single real requirement can reveal whether the team repeats less, rules load correctly, and verification evidence exists before delivery.

Takeaways
— OpenSpec and Spec Kit both support spec-driven development but differ in structure: Spec Kit guides teams through a staged flow from constitution to tasks; OpenSpec bundles proposal, design, requirements, and tasks into a single change that gets archived.
— Harness is an engineering concept covering how the model gets context, calls tools, tracks progress, and verifies outcomes; Trellis is a concrete implementation.
— Rulesync generates per-tool configs from a unified source, covering rules, skills, and MCP; Ruler distributes central Markdown rules to supported tools.
— Config consistency alone does not ensure output quality; final code still needs tests, static checks, and human review.
— When combining Trellis and OpenSpec, teams must assign ownership for each generated file and designate the canonical source of task state to avoid overwrites.
— A first-round pilot should observe three things: reduced repetition, correct rule loading, and explicit verification evidence before delivery.
Conclusions

The tool landscape for AI development suffers from overlapping marketing; separating layers by failure mode is a pragmatic way to avoid over-installation.

The distinction between config consistency and result compliance is often lost; syncing rules can give false confidence if the verification stage is missing.

The suggestion to test tools on the same real requirement is a strong heuristic because feature lists don't reveal how a tool fits your workflow.

Concepts & terms
Harness
The infrastructure that supports long-running AI agents: context provisioning, tool calling, progress saving, and verification. It is not a single product but a set of mechanisms.
Spec
A document or set of documents that define the behavior, edge conditions, and acceptance criteria for a specific feature, distinct from long-term engineering principles.
Project constitution
A set of long-term engineering rules for a codebase, such as architecture boundaries, dependency choices, and testing expectations, written to guide every development task.
From the discussion
Featured comments
组合搭配函数

I totally agree with this approach of picking based on pain points; installing everything at once is just digging a pit for yourself. In a previous project I stacked several different workflows, and the states ended up conflicting every day, and in the end we had to manually patch things up — it was really troublesome.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗