AI-Assisted App Development Workflow with Codec and Ghosty

The goal is to share a practical AI‑assisted coding workflow that actually works. Personal projects such as a presentation app, a spider identifier, and a 2‑dimensional tile editor illustrate the approach. The workflow is broken down into tools, required skills and command‑line interfaces (CLIs), agent usage, issue resolution, and continuous learning.

Tools

Codec is used both from the command line and through its graphical app. Access requires a ChatGPT Pro subscription. The CLI version is prized for speed, the ability to juggle multiple projects and terminals, and the focus it provides. The GUI version adds inline diffs, image previews, and a user‑friendly interface, though it can become sluggish when handling long threads.

Ghosty pairs with the Codec CLI, offering blazing‑fast performance and the capacity to open many tabs and side panes, which keeps the workflow fluid.

Whisper Flow supplies voice input for a range of applications—including emails, ChatGPT, and Codec—allowing the developer to dictate text that is automatically massaged and formatted for the appropriate context.

Skills & CLIs

The App Creator skill is a game‑changer for rapid prototyping. It orchestrates Xcode project scaffolding, provides Xcode Makefiles that define build, test, and run rules, and includes Simple Tasks that can be used to document work, especially when several agents are active.

The Align CLI tool gives agents visual context by running scripts that capture screenshots and identify UI elements, even when GPT 5.4 can interact directly with the screen. This background screenshot capability is non‑intrusive and keeps the agent informed about the current UI state.

Doc Set Query converts local documentation—such as the Dash app’s Apple developer docs—into markdown files. Agents can then query this material with grep or ripgrep, ensuring they have up‑to‑date, precise information beyond their training data.

Working with Agents

Plan Mode in Codec lets the user outline a series of tasks (opened with Shift‑Tab in the CLI). The agent expands the outline, filling in details and surfacing aspects the user might miss. Because the output can contain errors, it must be proofread before execution.

Direct interaction is encouraged: give the agent a single, focused feature, let it iterate, and avoid over‑complicating the request.

Simple Tasks, part of the App Creator skill, serve as a way to document each step, preventing conflicts when multiple agents operate concurrently.

Fixing Issues with Codec

Debugging rarely succeeds in a single pass. The workflow relies on macOS screenshots (Command‑Shift‑4) to capture windows, Xcode build and streaming logs for runtime information, and visual verification tools such as Align or Peekaboo to let the agent “see” the screen output.

The “continue” command can be used sparingly to ask the agent for additional work without losing focus. Periodically asking the agent to summarize its next steps helps keep the process on track.

Continuous Learning

Creating files like learnings.md or best_practices.md records what worked and what didn’t. Before starting a new project, agents read the plan, the learning files, and any relevant documentation, ensuring they begin with the latest context.

Agent configuration files (agents.md) can be synchronized across projects, with global versions for common tech stacks and project‑specific versions for specialized needs. An example from Peter Steinberger demonstrates how to structure these files and incorporate atomic commits.

Demo & Examples

The speaker demonstrates two applications: a slideshow app that adds slides and shows a full‑screen overlay, and a 2‑D tile map editor featuring random maps, elevation, animated elements, a smart drawing brush, and terrain manipulation.

A specific bug involved incorrect shadow rendering on cliff tiles at the water’s edge. The debugging process used screenshots, Whisper Flow for communication, and a comparison of expected tile numbers (water: 21‑24; land cliffs: 17‑20) with actual output. The root cause was a Z‑order issue where the shadow sprite was drawn on top of the cliff tile instead of beneath it.

Fixing the bug required updating learnings.md, plan.md, and the agent file, followed by an atomic commit that ensured a single, non‑conflicting change.

Future Content

The workflow shown for the Codec CLI will be expanded in future videos, and viewers are invited to request a demonstration of the Codec app workflow through comments or likes.