Turn dictated specs into tickets and code with LLM agents
I write a spec by speaking, not much thought given to grammar and sentence structure. Task: The author dictates rough specs via Whisper, manually edits for speech errors, and uses an LLM to densify them into structured Markdown. An LLM generates implementation plans and scaffolding, while agents create top-level tickets and parallel commit-level issues. Parallel agents implement code, adversarial reviewers check it, and the human merges only after verifying green CI and reviewing diffs.; tools: custom-built or unnamed agent. Human role: The original post describes the human role this way: Any failures spawn rework agents until the feature branch is up to snuff.I review the feature branch diff and merge it, having read the review and rework comments on the PR.; limitations: Success depends heavily on modular code design to enable parallelization. The workflow requires strict adherence to global style guides and repo policies. The author notes that while agents are cheap, they still require significant human oversight for merging and correcting wrong implementations..
Tools
custom-built or unnamed agent
Published on Agent Work Atlas:
Overview
I write a spec by speaking, not much thought given to grammar and sentence structure.
Workflow
The author dictates rough specs via Whisper, manually edits for speech errors, and uses an LLM to densify them into structured Markdown. An LLM generates implementation plans and scaffolding, while agents create top-level tickets and parallel commit-level issues. Parallel agents implement code, adversarial reviewers check it, and the human merges only after verifying green CI and reviewing diffs.
Reported result
The author reports results are 'quite good,' comparing the workflow to managing junior developers. They note it is vastly cheaper than human labor but acknowledge inefficiencies remain. No specific metrics like lines of code or time saved are provided, only qualitative success.
Human role
The original post describes the human role this way: Any failures spawn rework agents until the feature branch is up to snuff.I review the feature branch diff and merge it, having read the review and rework comments on the PR.
Limitations
Success depends heavily on modular code design to enable parallelization. The workflow requires strict adherence to global style guides and repo policies. The author notes that while agents are cheap, they still require significant human oversight for merging and correcting wrong implementations.
Try it
I want to replicate a spec-to-code workflow. First, I will dictate a rough feature spec for [replaceable feature name] into a text file. Use Whisper or similar to transcribe if needed, then manually fix obvious speech errors. Next, use an LLM to convert this into a dense, structured Markdown technical spec. Save this as 'spec.md'. Then, ask the LLM to generate an implementation plan and basic project scaffolding (Makefiles, tests, logging) based on 'spec.md'. Save the plan as 'plan.md'. Finally, create a prompt for an agent to generate top-level implementation tickets from 'spec.md' and 'plan.md'. Do not execute code generation yet. Show me the generated 'spec.md' and 'plan.md' for my review and approval before proceeding to ticket creation.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.