Fatih

Built a full stack for founders this past month. All open-source, all free. apply-yc: interviews you, drafts your YC application in your voice, fills the form. S26 closes May 4. apply-a16z: same thing for the a16z speedrun. Up to $1M. tonone: 23 AI agents as your engineering and product team once you get in. elephant: Claude Code memory across sessions. Your whole team, same context from day one. No excuse not to apply. No excuse not to build. github.com/thisisfatih/apply… github.com/thisisfatih/apply… github.com/tonone-ai/tonone github.com/tonone-ai/elephan…

Claude Code forgets everything between sessions. You re-explain your stack, your decisions, your context -- every single time. elephant fixes that. Persistent memory, automated changelogs, auto-updated READMEs. One plugin, works across every project. Your whole team gets the same context from day one. github.com/tonone-ai/elephan…

Every time you start a new Claude Code session you re-explain your entire stack. elephant fixes the memory. tonone fixes the expertise. 23 specialist agents -- each permanently owns a domain. Forge knows your infra. Spine owns your APIs. Warden knows your security. No re-onboarding. No context switching. github.com/tonone-ai/tonone github.com/tonone-ai/elephan…

No excuse not to apply to YC or a16z anymore. We built free open-source tools for both applications. They interview you, draft answers in your voice, fill the form automatically. Runs in Claude Code. Just show up with your idea. github.com/thisisfatih/apply… github.com/thisisfatih/apply…

YC S26 closes in 6 days. I studied what made successful applications specific rather than generic, then built an open-source agent around it. It interviews you, critiques your draft as a YC partner would, and fills the form automatically. github.com/thisisfatih/apply…

Most teams treat code review as a quality gate. The best teams treat it as a knowledge transfer. The diff is secondary. The reviewer learning why the decision was made is the actual value.

Unpopular opinion: the context window isn't your bottleneck. Your mental model of what the agent is doing is. You can give an agent 200k tokens and still produce garbage if you don't scope the task tightly. Bigger context just lets you defer that problem longer.

The AI database deletion story is making rounds but the real lesson isn't 'don't use AI agents.' It's 'you haven't set up your guardrails.' PreToolUse hooks in Claude Code let you intercept every shell command before it runs. Most people leave this off by default.

What's your actual cutoff for trusting an AI agent to run unattended? File edits? Shell commands? DB writes? Curious where people draw the line in practice vs in theory.

If your Claude Code session keeps losing context mid-task, the fix isn't a bigger CLAUDE.md. Split the work into smaller scopes. One task per session is faster than one session per day.

Hot take: the bottleneck in agentic coding isn't the model. It's the engineer's ability to verify output fast enough to stay unblocked. Speed of review, not speed of generation, is the new constraint.