Breakdown is a development instruction language tool for AI-automated development using TypeScript and JSON/YAML/Markdown.
BreakDown is a prompt generation tool for AI development support. It generates prompts in a format that AI can easily understand based on input Markdown files and specified conditions. The generated prompts function as instructions for AI systems to perform actual transformations and analyses.
Key features:
Prompt Generation Process
When executed from the command line, it simply receives prompts.
sequenceDiagram
participant Developer as App Developer
participant Breakdown as Breakdown Tool
participant Files as File System
Developer->>Developer: Write project overview and requirements in Markdown
Developer->>Breakdown: Execute command (e.g.: breakdown to project)
Breakdown->>Files: Identify required files
Note over Breakdown,Files: - Prompt files<br>- JSON schemas<br>- I/O paths
Files->>Breakdown: Read files
Breakdown->>Breakdown: Prompt generation processing
Note over Breakdown: - Prompt selection<br>- Variable substitution<br>- Schema application
Breakdown->>Developer: Output generated prompt
The terminal command → prompt part can be run on AI development agents. The goal is to generate recursive instructions based on these instructions.
AI Processing Process
The basic structure is simple. By establishing a prompt generation process for AI automated development, instructions are automatically given. Instructions create new subdivided instructions, which are used for subsequent instructions.
sequenceDiagram
participant Developer as App Developer
participant AI as AI Development Agent
participant Breakdown as breakdown command
participant Result as Processing Result
loop Instruction・Prompt Generation・AI Processing Loop
Developer->>AI: Input new instruction
AI->>Breakdown: Execute breakdown command based on instruction
Breakdown->>AI: Output generated prompt
AI->>AI: Processing based on prompt
Note over AI: - Markdown parsing<br>- Structuring<br>- Task decomposition
AI->>Result: Output processing result
Result->>Developer: Receive result
Developer->>Developer: Result = New instruction
end
Internal Processing
Pre-prepared prompts are used. The appropriate prompt is selected from the command, and parameter values are treated as variables and embedded.
Number of combinations:
This results in the following.
When multiplied by the versatility of LLMs, this tends to diverge.
When development deliverables are the final product, methods are needed to control divergence and converge toward the goal. Additionally, the specification definitions and decomposition methods necessary for development become finite patterns due to accumulated theories. (For example, the 23 design patterns of GoF.)
Therefore, Schema definitions are used to guide diverse expressions into specific patterns.
The Breakdown tool pre-defines three layers: Project > Issues > Tasks, and the to instruction performs decomposition.
At this stage, pre-prepared prompts are classified into either project/issue/task.
When multiplying the combinations of input → decomposition destination (output), 5 combination patterns (P->I,T, I->I,T,T->T) are pre-prepared.
Also, for cases different from the decomposition direction, the summary instruction performs aggregation (project -> Project, Issues -> Project, etc.). In particular, the defect instruction is provided to handle error processing, bug response, and requests for unmaintained states. Please see Usage for details.
Making the prompt selection part AI-driven would make it more dynamic. This part can be separated from this tool as it would be more efficient to use external services.
JSON Schema references are used to converge to finite patterns. The goal is not to fit into Schema definitions but to reference them as checklists to control generation results. The final output does not need to be JSON. More versatile YAML/Markdown/Text formats are expected.
In a future version, it is planned to make Schemas referenceable by URL. This will enable data processing using Schema.org and utilization of other Schema definitions.
If appropriate Schema definitions for development requirements exist, they will be referenced, but as none have been found yet, we plan to create what this application requires.
| English | 日本語 |