Babysitting LLMs in Copilot Agent Mode: Four Full-Stack Builds, Four Models
A developer ran four full-stack builds in GitHub Copilot agent mode (o4-mini, Gemini 2.5 Pro, Claude 3.7 Sonnet twice), preparing a spec, task list, and instructions file before each run. All models initially failed the React/Tailwind v4/ShadCN setup; the poster babysat shell commands throughout. Claude 3.7 Sonnet's second run finally produced styled, working React components. Task: 1. Prepared three files before each run: PROJECT.md (spec), TASKS.md (step-by-step list), copilot-instruction.md (agent rules). [Setup · Preparation] 2. Ran four builds with o4-mini, Gemini 2.5 Pro, Claude 3.7 Sonnet (twice), refining specs between runs. [—] 3. Babysat shell commands: corrected path errors (cd frontend && bun install inside /frontend/frontend) and Windows/Linux CLI mismatches; added "use full path names in CLI" rules. [Oversight · Exception handling] 4. Used ChatGPT deep-research to pin the correct Tailwind v4 + ShadCN install order, then re-ran. [Exception handling]; tools: GitHub Copilot, Claude. Human role: - Oversight: the poster constantly monitored shell commands, correcting paths and CLI issues. Preparation: spec, task list, and instructions files written before runs. - Other division of labor: not stated in the source.; limitations: - Single poster, single private project; anecdotal. - Windows (cmd/PowerShell) runs added CLI friction. - Model behavior changes quickly; results may not reproduce. - ~70-task project; React/Tailwind v4/ShadCN setup is version-sensitive..
Tools
Published on Agent Work Atlas:
Overview
A developer ran four full-stack builds in GitHub Copilot agent mode (o4-mini, Gemini 2.5 Pro, Claude 3.7 Sonnet twice), preparing a spec, task list, and instructions file before each run. All models initially failed the React/Tailwind v4/ShadCN setup; the poster babysat shell commands throughout. Claude 3.7 Sonnet's second run finally produced styled, working React components.
Workflow
1. Prepared three files before each run: PROJECT.md (spec), TASKS.md (step-by-step list), copilot-instruction.md (agent rules). [Setup · Preparation] 2. Ran four builds with o4-mini, Gemini 2.5 Pro, Claude 3.7 Sonnet (twice), refining specs between runs. [—] 3. Babysat shell commands: corrected path errors (cd frontend && bun install inside /frontend/frontend) and Windows/Linux CLI mismatches; added "use full path names in CLI" rules. [Oversight · Exception handling] 4. Used ChatGPT deep-research to pin the correct Tailwind v4 + ShadCN install order, then re-ran. [Exception handling]
Reported result
- Completion time, cost, success rate: not disclosed by the source. - o4-mini: ~80% of tasks done with assistance, but the result was broken. - Gemini 2.5 Pro: stopped editing around task 12 of 70; looped on empty promises. - Claude 3.7 Sonnet: most proactive; its second run delivered styled, working React components. - All four builds initially failed React + TypeScript + Tailwind + ShadCN setup — none configured Tailwind v4 with ShadCN out of the box. (Poster-reported; not independently re-tested.)
Human role
- Oversight: the poster constantly monitored shell commands, correcting paths and CLI issues. Preparation: spec, task list, and instructions files written before runs. - Other division of labor: not stated in the source.
Limitations
- Single poster, single private project; anecdotal. - Windows (cmd/PowerShell) runs added CLI friction. - Model behavior changes quickly; results may not reproduce. - ~70-task project; React/Tailwind v4/ShadCN setup is version-sensitive.
Try it
Act as a senior full-stack developer using GitHub Copilot agent mode. Build a new app with React, TypeScript, Tailwind CSS, ShadCN/UI, Python, and PostgreSQL. Inputs: 1. Project Spec [replaceable]; 2. Task List [replaceable]; 3. Instructions File [replaceable]. Steps: 1. Initialize the project structure per the spec. 2. Run setup commands per stack in order. 3. Monitor all shell output for path or OS-specific CLI errors. 4. Pause for my confirmation before destructive commands or global installs. 5. On error, propose a fix and wait for approval. Deliverable: a working local environment with dependencies installed and a basic component rendered. Metric: count manual interventions needed. Safeguards: do not publish code or spend credits without permission.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.