GitHub Spec Kit — the assistant works from a spec, not from inspiration
You ask for a feature, the assistant writes code, and half an hour later it has solved the wrong problem. GitHub Spec Kit puts the written spec first, and we tested its flow. It is GitHub’s own setup for making an assistant read before it writes. The tradeoff is time: you slow down the first pass so you waste less later.

What it does
You ask your assistant to build a feature. It starts writing code, and half an hour later you realize it solved a different problem. That is what happens when you go by instinct and skip the paper.
Spec Kit is GitHub’s official tool for doing the opposite. First you write what you want. Then how you will build it. Then which tasks it breaks into. Only at the end does code get written. Each step leaves behind a text file that feeds the next one. The assistant does not guess; it reads.
The main flow has five stages: Specify, Plan, Tasks, Implement, Converge. In plain terms: what I want, how I will do it, which tasks it splits into, let me build it, and then check what is left. Beside those there are two optional extensions, one for fixing a bug and one for judging whether an idea is worth the effort (documentation).
This is not a program that writes code. It is a set of instructions and templates that the assistant reads inside the tool you already use. In practice it installs a series of instruction files. Those files are called skills: they load on their own when the assistant needs them, without you asking.
What it does not do
Spec Kit is not a machine that spits out finished features when you give it only a fragment of an idea. We tested the installation and saw that the tool produces templates and skills — it does not execute, it does not compile, and it does not check whether the code it will produce actually runs. Do not ask it for automatic testing, debugging, or deployment. You or the CI tools you already have do those steps.
It does not replace the developer or the architect. If your spec is weak, the assistant will follow the wrong instructions. Do not expect it to clear up ambiguity on its own. It also does not handle repository management, dependencies, or task tracking — it leaves the files behind and waits for human verification.
The trap is the false sense of safety: the flow slows down fast production, and if you hate paperwork, you will skip it and write the wrong things again. Use it on large jobs. Do not load it with tiny changes or with the expectation that it will automatically solve every assistant-understanding problem.
Who it is for
For anyone tired of explaining the same thing from scratch in every new request. Spec Kit puts a sheet of paper in front of you before you burn hours in code.
It is useful for large features, for new projects, and for work you will pick up again after weeks and want to remember what you had in mind.
It is not worth it for changing a color. There the paper weighs more than the work.
Installation
You need Python 3.11 or newer, uv, and one assistant from the list. The tool is called specify-cli.
uv tool install specify-cli
Check that it is installed:
specify --version
In our test it returned “specify 1.1.3”. There is also a readiness check:
specify check
It answered “Specify CLI is ready to use!”.
Then you create a project. This is the point that confuses people: the assistant key is not always what you expect. For Cursor, for example, the key is cursor-agent, not cursor (integration list).
If the assistant tool is not installed on the machine, init complains. Then you add a flag:
specify init my-project --integration claude --ignore-agent-tools
The –ignore-agent-tools flag means “I only want the files, do not ask me whether I have the agent.” We needed it for real, because there is no Claude Code CLI here.
Adding a second assistant
If you want the same project to be worked on by several assistants, you add them in the same folder. They share .specify/ and specs/.
specify integration install codex
specify integration install cursor-agent
specify integration status
Use in each LLM
The installation is done. Now you open the assistant inside the project folder and call the stages in order. How you call them changes from assistant to assistant, and that is the part articles usually leave out.
Claude Code
First you create the project with the claude key.
specify init my-project --integration claude
Then you open Claude Code inside the folder. You ask for the stages in chat, one by one:
/speckit-constitution
/speckit-specify
/speckit-plan
/speckit-tasks
/speckit-implement
The skills sit in the project’s .claude/skills/ folder, ten in all. You run constitution once at the start. Then specify writes the spec, plan the technical plan, tasks the task list, implement the code. Watch this: these are skills, not terminal commands. You type them in the assistant chat, not in the black window.
Codex CLI
Same idea, different key.
specify init my-project --integration codex
Here two things change. The skills are written to the .agents/skills/ folder instead of .claude/skills/. And you call them with a dollar sign instead of a slash:
$speckit-specify
In Codex you write $speckit-specify, $speckit-plan, $speckit-tasks, $speckit-implement. The substance is the same, the syntax changes.
Cursor
And here is the trap. Cursor does not come in with the word cursor.
specify init my-project --integration cursor-agent
The skills are written to the project’s .cursor/skills/ folder. You call them with a slash, like in Claude Code:
/speckit-specify
Cursor is a windowed app, not a terminal. So there is no need for a command-line tool to be installed for the integration to work; the project folder is enough.
Where it wins, where it loses
It wins on order. Before a single line of code is written, there is paper with what you asked for. The mistakes show up early, when fixing them is cheap.
It wins because it does not lock you in. The documentation counts 38 integrations, and there is a generic key for anything not on the list. You can start with one assistant and add another tomorrow.
It also wins on maturity. The repository has 140,656 stars, 12,594 forks, and an MIT license, with the latest release v1.1.3 on 9 October 2026 (releases).
It loses on first contact. Ten skills, five stages, and two extensions is vocabulary that takes time to get used to.
It loses on small work. When you are changing one color, the process weighs more than the job.
And it loses when you are in a hurry. The flow stops you for a check at every step. Some people feel that as a brake.
Test
It really ran, on this machine, with uv.
The installation passed, and specify –version returned “specify 1.1.3”. specify check answered “Specify CLI is ready to use!”.
Then we ran a real init with Claude as the integration:
specify init demo-app --integration claude --non-interactive --ignore-agent-tools
The –ignore-agent-tools flag was needed because there is no Claude Code CLI here. It passed. The demo-app folder filled with .claude/skills/ containing ten skill folders, including speckit-specify, speckit-plan, speckit-tasks, speckit-implement, and speckit-constitution. Beside them, .specify/templates/ with five templates: spec, plan, tasks, checklist, constitution. I counted them: ten and five.
What was not counted. We did not run a full cycle from specify to implement inside an assistant. We did not test integrations in Codex or Cursor. Anything I write about the steps inside each assistant and their syntax comes from the documentation, not from hands-on testing.
Verdict
Spec Kit solves a problem everyone feels and few name: the assistant answers quickly, but not always the question you asked. Here the flow comes before speed, and it does that part right.
It is mature, active, free, and under an MIT license. The fact that GitHub itself maintains it counts for trust.
The price is the learning time. Two or three jobs will pass before it feels natural.
If you work with Claude Code, Codex, or Cursor, and you build things you will not remember in two days, try it.
Score: 8.8/10.
See also
If this was useful, see also:
- ty: Astral’s type checker that catches the mistake before it runs
- cf: Cloudflare’s new CLI that does not guess flags
- ast-grep: grep that understands code, not text
Sources:
https://github.com/github/spec-kit
https://github.github.io/spec-kit/
https://github.com/github/spec-kit/blob/main/docs/reference/integrations.md