Short summary
Most AI demos stop when something appears on screen. This one starts there.
Dalton Anderson uses OpenAI Codex to turn fictional customer feedback into a working dashboard, then critiques the first output, steers the redesign, and explains the system around reliable agent work: Plan as the map, Goal as the contract, skills as reusable procedures, and evaluations plus stop conditions for agentic loops.
A practical episode for founders, operators, and builders who want better results from Codex without giving up judgment.
### Mobile-first show notes
Most AI demos stop when something appears on screen. This one starts there.
Dalton gives Codex a fictional customer-feedback dataset and asks for a decision-ready dashboard. The first output works, but it is not good enough. That becomes the real lesson.
This episode shows how to move from a one-shot prompt to a workflow you can inspect, steer, and verify.
You will learn:
- Why Plan is the map and Goal is the contract
- The four parts of a strong goal: outcome, context, constraints, and done-when evidence
- When a repeatable workflow should become a skill
- How clear names keep a growing skill library usable
- Why every agentic loop needs an evaluation and a hard stop
- Where human judgment still matters
The demonstration uses fictional data. No outreach is sent.
## Official OpenAI resources
- [Codex use cases](https://developers.openai.com/codex/use-cases)
- [Build skills](https://learn.chatgpt.com/docs/build-skills)
- [Build plugins](https://learn.chatgpt.com/docs/build-plugins)
- [Subagents](https://learn.chatgpt.com/docs/agent-configuration/subagents)
- [AGENTS.md](https://learn.chatgpt.com/docs/agent-configuration/agents-md)
- [Follow a goal](https://learn.chatgpt.com/use-cases/follow-goals)
- [Scheduled tasks](https://learn.chatgpt.com/docs/automations)
- [Git worktrees](https://learn.chatgpt.com/docs/environments/git-worktrees)
- [OpenAI Codex repository](https://github.com/openai/codex)
- [OpenAI Plugins repository](https://github.com/openai/plugins)
Note: the older [openai/skills repository](https://github.com/openai/skills) is deprecated and now directs readers to OpenAI Plugins.
## Skill repositories worth exploring
- [Anthropic Skills](https://github.com/anthropics/skills): official Claude skill examples and templates
- [Superpowers](https://github.com/obra/superpowers): a cross-agent software development workflow and skill collection
- [Microsoft Skills](https://github.com/microsoft/skills): skills and custom agents for Microsoft developer workflows
- [Microsoft Learn Agent Skills](https://github.com/MicrosoftDocs/Agent-Skills): Microsoft and Azure skills grounded in Learn documentation
- [Gemini CLI](https://github.com/google-gemini/gemini-cli): Google's open-source coding agent with Agent Skills support
- [Gemini CLI Agent Skills guide](https://geminicli.com/docs/cli/using-agent-skills/)
- [Agent Skills specification](https://github.com/agentskills/agentskills): the open format behind portable skills
- [Vercel Skills](https://github.com/vercel-labs/skills): a cross-agent CLI for discovering, installing, and sharing skills
Install selectively. Read a skill before trusting it, understand the tools and permissions it can use, and test it on bounded work first.
## Chapters
00:00 Why this episode exists
01:45 Turning fictional feedback into a dashboard
02:29 Commands, context, compact, goals, and Plan
04:16 What the first plan is doing
07:05 Steering the build with butter yellow
09:03 Reviewing the first dashboard
10:37 Plan is the map, Goal drives the work
15:54 An honest review of the redesign
17:23 Building a Goal and the four-part prompt formula
21:08 Commands, skills, and reusable workflows
24:57 Naming skills so they stay usable
26:34 Broad threads and focused projects
28:27 A simple context-and-constraints analogy
30:52 Keeping agentic loops safe
32:21 Closing
