tl;dr: I built and shipped a brand-new feature live (in ~1 hour), watch how I approach agentic engineering with codex
Join me for an unfiltered look at building Arena - a live coding session where you can see my development process in action, unscripted. Thanks to Eleanor Berger for motivating me to do this video and for organizing the live event!
What we built
🤖 Heads up! This is an AI-Assisted summary.
- Feature: Arena — a new feature in my upcoming project, to see how well 2–4 users from X match
- Input: Twitter/X user handles
- Pipeline: fetch N tweets per user (shares a 1,000-tweet budget), strip to necessary fields, run profile analysis, then score compatibility (per pair + whole team)
- UX: user picker + “Analyze” button, results table, cached runs selectable under the search box
- Infra touches: DB migration for
arena_cache, long-running job in the background, streaming UI, auth-guarded page
I managed to complete the feature in ~1h, and we got a pair score of 89 for me and @intellectronica.
Stack & Setup
-
Model workflow: codex (OpenAI/GPT-5) for coding sessions; it eagerly reads the codebase and generally “does the right thing” without handheld file lists. I keep a separate Claude-style flow for some web searching, but for repo work codex is the star.
-
Sessions: start fresh for big features, run multiple agent windows in parallel; switch tasks while one is thinking.
-
Branching: work directly on
mainwith atomic commits. Merge conflicts + worktrees cost speed; small, well-scoped commits keep things safe. -
Tooling:
-
Docs ingestion: pull only what’s needed, prefer markdown via a crawler over raw HTML to save tokens.
-
Validation: schema validation on inputs; fail fast and surface helpful messages.
Tactics that mattered
- Keep the agent context clean. Minimize tool noise and only inject docs when needed; markdown > HTML to conserve context.
- Let codex plan on demand. It proposes next steps that are usually solid; green-light them in sequence.
- Cache long tasks early. Add a table + background job queue before you polish UI; this saves you from re-running expensive analysis.
- Copy-paste errors verbatim. Don’t over-explain; just drop logs in and let the agent fix what broke.
- Comments as spec. Ask for clear intent comments near tricky code; they’re for you and for the agent on the next session.
- Write tests afterwards, not first. These agents are great at back-filling tests once the shape exists, and that’s usually enough to catch regressions.
- Work on
main, commit surgically. You still get speed with safety. Backups + Git are your safety net. - Avoid local models for this workload. Context and stability matter more than shaving milliseconds.
- CLIs beat MCPs. A 2-hour CLI wrapper (logs, API pulls) pays for itself and keeps context small.
- Small “proof” projects unblock big features. If streaming or a protocol is hard, build a tiny in-repo example that works, then transplant.
Q&A highlights
- Why Codex over Claude Code? Codex actually reads more of the repo without handholding; requires fewer “look here, then there” hints. Claude still useful for search/web, but for coding Codex felt faster end-to-end.
- Do you branch? Not for features at this stage. Main + disciplined commits is faster and - counterintuitively - safer for rapid iteration.
- Manual approvals for agent actions? No. That becomes “Windows Vista prompts.” Use Git + backups; review diffs; move fast.
- Repo prompts & MCP servers? Nice idea, but they bloat context and introduce fragility. Lean instructions + small CLIs work better.
- Compaction & long sessions? Plan features to fit the context. If you expect many loops (e.g., test repairs), use the flow that compacts well - or split the task.
If you want to read more about my agent workflow, check my AI development post here.
For more advanced prompting techniques with GPT-5, also check out the OpenAI GPT-5 Prompting Guide.