Summary of "I'm Done With AI Coding"

Concise summary

The creator experimented with “AI coding agents” for several months, then mostly stopped using them inside his editor because the overall workflow cost more time, reduced code quality, and eroded his coding skills. He still uses chat-style AI (Gemini) for occasional questions but no longer relies on in-editor or code-generation agents.

Technical analysis and key findings

Workflow overhead
- Large time investment went into prompt engineering, context provisioning, guardrails, CLIs/hooks and linters to keep agents from producing broken code. He estimated this overhead reached ~70% of his time at one point.
Productivity curve critique
- Initial apparent gains (many generated lines) are misleading. After early gains, time is lost fixing hallucinations, silly bugs, refactors and maintaining generated code. Net productivity can be lower than traditional/manual development.
Quality and maintenance concerns
- AI-generated code frequently contained bugs, poor structure or incorrect behavior, requiring substantial manual review and refactoring. This often turned supposed “help” into extra work.
Skills erosion
- Relying on agents harmed his coding fluency. He spent more time learning prompts and guardrails than core programming concepts, and found his manual coding skills declined.
Competitive landscape / opportunity
- Fewer people are doing deep technical coding; many focus on “prompting.” He sees this as an opportunity to focus on core skills (e.g., Rust, C, protocols).
When he might return to heavy AI coding
- Only if a model appears with dramatically better, near-perfect code generation and no hallucinations — i.e., an exponential improvement over current models.

Tools, models and practices referenced

Guardrails and automation
- Linters in CLI/IDE, cloud hooks, and context engineering to give agents better visibility into codebases.
Specific services/models mentioned
- Gemini (chat) — preferred for occasional Q&A.
- PI CLI (previous tool he used).
- Opus (4.x).
- GPT (5.x) — reported incremental improvements (e.g., 5.1 → 5.4) are negligible for real productivity gains.
Preferred current usage
- Chat interfaces for brainstorming and answers.
- No autocomplete or coding agents in his editor or terminal.

Product announcement — WebDefuel

Purpose

WebDefuel is a learning platform designed to provide fully pre-configured learning environments so students don’t need local setup. It targets learners who struggle with initial environment/configuration.

Key features

No local installation required; includes client, server, databases and dependencies.
Each project/course provides a unique project repo/ID that you can clone locally if desired.
Choice of a web-based editor or using your local editor (Vim, VS Code, etc.).
Full end-to-end projects (examples: Golang backend + HTMX/Temple frontend) with exercises and project tracks rather than only small code snippets.

Early-access / promotion

Buyers of the creator’s previous HTMX + Go course (“HTMX and Go Blueprint”) get one year of the platform free once beta/public release occurs.
Early supporters can sign up for alpha/beta access at webdefuel.com (subtitles sometimes show “webdefield” — likely a typo) to help shape the product. Subscription/yearly plans will be offered after release.

Guides, tutorials and past/future content

Past content referenced
- Previous videos exploring AI coding and the HTMX + Go course/blueprint.
Future content
- He will mostly stop producing AI-coding-agent content and focus on traditional programming tutorials and deeper technical topics (web development, Rust, C, protocols, and possibly machine learning/AI from the creator-side).
- Offer to past course customers: free one-year access to WebDefuel when released.

Decisions & practical recommendations

Don’t assume generated code equals a shipped product — plan for review, testing and refactor time.
If you rely on AI agents, invest in automated guardrails (linters, tests) but expect extra setup cost.
For learning, prefer environments that remove setup friction but still force you to write and understand code manually.

Don’t assume generated code equals shipped product — plan for review, testing and refactor time.

Main speaker / sources

The video’s creator / channel host — a developer and instructor who runs the WebDefuel platform and authored the HTMX + Go Blueprint (name not given in the subtitles).
Mentioned AI providers/models: Gemini (chat), PI CLI, Opus, GPT (GPT-5.x referenced).