Yet Another Spec-Kit. Go figures! I based this spec-kit mainly using Kiro's rules and guidelines and complimented with documentation from a few others, like user-created and GitHub's own spec-kit. I wasn't satisfied with any of them. This format should be compatible with Kiro.
Differences to Kiro: This is Kiro's methodology made explicit and platform-independent. Same principles, same workflow, same results, but documented for use with any AI system or IDE. Potentially scales better across teams, tools, and complex projects using explicit documentation patterns. See report.md for an AI-generated evaluation.
Reports on performance and suggestions for improvements are welcomed! Best of luck with your projects.
(Windows) Requires: Git for Windows, for Bash
Transform feature ideas into structured specifications through AI-guided Requirements → Design → Tasks → Implementation workflow.
- Copy into your project workspace, change the agent prompt to
gemini.mdand delete the installers.
- Extract the files somewhere, and run
./install-gemini.sh - Use
/specin Gemini to initialize spec-dev in your current workspace. Restart Gemini.`
- Extract the files to your project directory and run
./install-cursor.sh - Cursor will automatically load the spec-dev agent rules and system files
If your agent is having troubles with the scripts, windows is likely using WSL's bash instead of Git's. Run 'fix_bash.bat' as admin to add Bash to your path.
The core components of this system are:
spec-dev-agent.md: The instructions for the AI agent..yask/: Contains all supporting system configurations, tools, and detailed documentation.blueprint/: This is where you'll store your project's development specifications.
All in-depth documentation, covering core principles, methodology, interaction patterns, prompting strategies, and implementation support (templates, scripts), is organized within the .yask/ directory.
Special thanks to @jasonkneen for their Kiro documentation set as a point of reference for the creation of this instruction set.
> Need to test both 2.21 and 2.21_cf side by side.
* Must make installer(s) also remove the install-cursor file.
- Must fix cursor implementation automation, currently as a 'mode' or modified custom instruction for Plan mode, only tried as a custom mode.
- Add VS-based profile for Roo and Kilo.
-! remove dependency on spec.sh script. Almost fully achieved.
? ...should we remove installer dependency altogether and bundle different zip releases for specific agent wrappers?
-! refine cross-documentation workflow, check that when updating design, it checks if it needs to update requirements, make more open to also changing ALL docs
? Centralized prompt, must check performance, as preemptive evaluation indicates degradation of context comprehension. More testing required.
* Initial testing suggests 'collage' centralized prompt produces worse results in models with sub-par instruction adherence, results degrade versus central+3 instruction method.
* Initial versions of 'jigsaw' centralized prompt are created, but must be further refined, validating results and focusing on format and wording inducing results and adherence.
> This seems to suggest that actually having 4 different files with redundancies, even if loaded at a secondary point, improve functionality.
> This seems... odd, as it all ends up in the same context, so 'collage' centralized prompt should perform identically.
* Future:
* Adding Validation and Testing guidelines and procedures
* Testing efficacy and then adding procedures for Pseudocode-based code parsing and creation
- look into maybe incorporating some more best practices and principles from the kiro guide docs.
? Work on integrations. Methods to recognize and use specific MCP tools such as Tree-Sitter and code health and parsing/validation tools would be meaningful, but drift the project away from its current flexiblity.
* Evaluate what adding a script could meaningfully contribute with, as to keep the prompt flexible and agnostic. *Could help with auto-updating it and with code validation.*