GitHub Spec Kit vs SpecWeaver: Judged Specs, Checked Specs
Spec-driven development (SDD) means writing a rigorous spec before an AI coding agent builds anything, so the agent works from stated intent instead of guessing. GitHub Spec Kit and SpecWeaver are both SDD tools. The difference is not the philosophy. It is who checks the result: in Spec Kit, the agent judges its own convergence; in SpecWeaver, a command exits 0 or 1.
What is GitHub Spec Kit?
Spec Kit is GitHub's free, MIT-licensed toolkit for spec-driven development, and in 2026 it is widely adopted: roughly 130k+ GitHub stars, a 1.0 release in August 2026, and steady releases since. It installs a set of slash commands into your coding agent: /speckit.constitution for project principles, /speckit.specify, /speckit.plan, and /speckit.tasks to produce markdown artifacts (spec.md, plan.md, tasks.md), /speckit.implement to build, and /speckit.converge to compare the codebase against the spec and append whatever work remains. It supports 30+ agents, including GitHub Copilot, Claude Code, Cursor, Gemini CLI, Windsurf, and Amazon Q.
It is good at what it claims. If your agent work suffers from vague intent, Spec Kit's structure fixes real problems, and the price is zero.
What is SpecWeaver?
SpecWeaver is our $37, one-time-purchase CLI (a command-line tool: install it, run it, no accounts or API keys). It is built around one idea: acceptance criteria should be machine-checkable. A done_when entry in a SpecWeaver spec can carry a check, a shell command whose exit code is the verdict, so nothing has to interpret the criterion and nothing can interpret it generously. Around that sit validate (schema and structural rules), diff (compares the file paths and symbol names the spec declares against the code on disk), gate (every check in one command with one exit code, with a mode that annotates GitHub pull requests), and verify (builds a review prompt from the spec and the diff, never from the builder's transcript). It works with any agent system that accepts structured specs. Python 3.10 or newer, three dependencies.
The real difference: who does the checking
Here is the part most comparisons get wrong. Spec Kit does have a spec-versus-code check: /speckit.converge reads your spec, plan, and tasks, inspects the code, and classifies the gaps. But that check is a prompt template the coding agent executes. The judge is the same kind of model that did the building. That scales to nuance a script cannot see, and it also inherits the model's generosity: an LLM reviewing work against a checklist can decide the checklist is close enough.
SpecWeaver's checks are deterministic. A done_when check either exits 0 or it does not. A declared file either exists or it does not. The gate either passes or blocks the merge. No judgment is involved, which is precisely the point of a gate: it holds when the model is having a bad day. The cost of that rigor is scope: a shell command cannot evaluate "the design feels coherent." The two approaches are not competing claims about the same thing. They check different things.
| Category | GitHub Spec Kit | SpecWeaver |
|---|---|---|
| Cost | Free, MIT license | $37 one-time, keyless |
| What it structures | The upstream workflow: constitution, spec, plan, tasks | The downstream gate: validation, drift, acceptance checks |
| Drift check (spec and code disagreeing) | /speckit.converge, executed by the agent as judgment | diff command, deterministic, against declared files and symbols |
| Acceptance criteria | Scenarios in markdown the agent traces | done_when entries carrying shell-command checks; exit code is the verdict |
| CI gate | None built into core | gate command, one exit code, GitHub pull request annotations |
| Agent support | 30+ coding agents | Any agent system that accepts structured specs |
Can you use both?
Yes, and the split is natural: they meet the spec at different ends. Spec Kit structures how intent becomes a spec. SpecWeaver checks whether the code honored it. If you only want structure, Spec Kit is free and you should use it. If your failure mode is agents reporting success on work a script would fail, the deterministic gate is the missing piece. That is the failure mode we built for. Inside our own stack, that is also what Level 2's spec workflow teaches: writing specs agents can act on, and gating what comes back.
SpecWeaver is $37, one-time, with the worked example shipping 22 done_when entries, 20 of them carrying an executable check. The honest comparison of everything we sell, including what Dark Factory is not, is on the compare page.