Comparison and limits
The family's overview compares it with eleven projects nearby. Each statement there about another project was read from that project's own repository, documentation or registry entry on 2026-11-09, and links to where it was read. Those projects change every week, so this page does not repeat what the overview says of them: it names them, and says what the spec family does in each area. Read the overview, and follow its links, before relying on a comparison.
Neighbours
| Area | Projects | What the spec family does there |
|---|---|---|
| Spec-driven development | Spec Kit, OpenSpec | Keeps a specification true after it is written, deterministically, on every change. spec-graph reads the layouts Spec Kit, OpenSpec and Kiro write with no configuration (layouts). |
| ADR enforcement | Archgate, Mneme, adrkit | A spec-guard rule is an HTML comment under the sentence it enforces, in the ADR itself, and is not run once the ADR is superseded (document status); spec-graph holds the ADRs' lifecycle. An exception to a protected file counts only as a commit signed by a key the base branch's allowed signers list (rulings). |
| Documentation checks | DocGuard, ctxgrd | An Implemented claim needs a directive that counts a test by its name, which spec-guard prove shows can fail (claims); a ticked box needs a commit git holds (checklists). The code is held to the documents, as well as the documents to each other. |
| Scoped agent work | Trailbun, spec-gates | A round of agent work runs under a brief, a Markdown file a person approves; each brief protects its own files for its own round, and a protected path opens only by a ruling the person signs (rulings). |
| Import rules | dependency-cruiser, import-linter | spec-guard's import rules are written in the ADR under the sentence they enforce, in JavaScript and TypeScript, Python, Go, Rust and C#, and stop being enforced when the ADR is superseded (imports). The two kinds of tool run side by side. |
What the tools do not do
The overview's full list gives the reason for each and where it is decided.
- The guard before a write works under one agent. spec-harness refuses a
write to a protected file through Claude Code's
PreToolUsehook, in Claude Code 2.1.139 or later (requirements). No other agent's hooks have been tested with it. Every agent, or none, still meets git's pre-commit hook where it is installed, the audit, the archive and CI. - A write through a shell passes the guard. The hook sees the agent's file tools, not a shell command. The audit and spec-brief's archive measure everything the round changed, whatever wrote it, so they find it, as the first story shows.
- A ruling proves the agent did not sign it, not that two people reviewed. One person may write the brief, run the round and sign its rulings; nothing checks that the signer is someone other than the brief's author. A forge that squashes or rebases on merge drops the signature.
- They do not run your tests.
spec-guard claimsasks anImplementedclaim for a directive that counts a test by its name, andspec-guard proveshows the directive can fail; whether the test passes is your test runner's to say, in the same CI run. - They do not decide whether the code is correct, or whether a document asks for the right thing. That is the tests' job, and the person's.
- They read only what is marked. A sentence with no directive under it, a number nothing marks, a file no configuration names: none is checked. No model reads the prose.
- They do not protect their own rule files. An agent that can edit the ADR, the configuration or the CI workflow can loosen the rule it is breaking; protected branches and code owners close that gap.
- Status words are read in English. A document may be written in any language and gives its status in English.
- The output is not frozen yet. The tools are before 1.0. This site shows what the releases it pins print, and is played again for each new pin.