James

Economist | Statistician | Computer Scientist l Solo Builder | No fluffy words here, just pure and gold content

[James @james_ios]

They say building in public is the trick. I’m officially announcing the beta version of a product I’ve been working on: https://t.co/5a86qA0vWX

Content creators should not have use separate apps to create, edit and schedule their posts. This is the pain point this app solves.

[James @james_ios]

8 automations. One browser identity rule.

A profile label mismatch was enough to stop the X agent even though the correct Chrome session was connected.

Solo dev lesson: policy bugs can kill more distribution than code bugs.

[James @james_ios]

Codex vs Claude Code pricing debates miss the expensive part.

A cheaper agent that waits for clicks costs more than a pricier agent running loops with state.

The winning model is the one you stop babysitting.

[OpenAI Developers @OpenAIDevs]

More of the iOS app loop, now inside Codex.

The Build iOS Apps plugin lets Codex view and test your iOS app in the in-app browser, open SwiftUI previews, and hot reload edits without leaving Codex.

[OpenAI @OpenAI]

Building apps has never been easier.

With Sites, Codex can turn your work, ideas, and plans into an interactive website or app your team can explore, use, and share with a URL.

Rolling out to Business and Enterprise plans, before expanding more broadly.

[OpenAI @OpenAI]

OpenAI frontier models and Codex are now generally available on AWS, giving enterprises a new way to build on Amazon Bedrock with OpenAI through the security, compliance, and governance workflows they already use.

This is also the beginning of a broader expansion of OpenAI capabilities on AWS, including future availability for cybersecurity capabilities like Daybreak.

https://t.co/vMws0YU6Q3

[James @james_ios]

AI labs ship a new model every 4 weeks.
Solo devs ship products every 3 months.

The cadence mismatch is the moat — by the time you finish building on a model, the next one already shipped.

Build on the API surface, not the model name.

[James @james_ios]

@m13v_ Well said 🔥

[James @james_ios]

Hosted AI agents rent you a workflow.
CLI agents on your hardware own one.

The vendor can't deprecate the cron you wrote on a Mac Mini. The vendor can deprecate everything on their dashboard.

Hosted is convenient. Owned is permanent.

[James @james_ios]

39 drafts. 22 live. Day 14 of the X agent.

Cron at 7:27 ET — Claude reads playbook + voice guide, queue is stocked before I wake.

The one task you can't scale is writing every morning. Unless you write the writer.

[James @james_ios]

launchd beats cron for Claude Code dailies:

1. plist in `~/Library/LaunchAgents/`
2. `ProgramArguments`: `zsh -c "claude --skill draft"`
3. `StartCalendarInterval` at 7:27
4. `launchctl load -w` it
5. `tail -f` the log

Survives reboot. Writes while you sleep.

[James @james_ios]

Claude Opus 4.7 ships with 1M context.

A solo dev's natural prompt is ~10k tokens, not 1M. The big-context arms race is selling a feature most of us can't fill.

Pay for the latency floor, not the ceiling.

[James @james_ios]

Run 3 Claude Codes in one tmux session:

1. `tmux new -s claw`
2. `claude --skill triage`
3. Split: `claude --skill code`
4. Split: `claude --skill review`
5. `Ctrl-b arrow` to switch agents

3 specialists, 1 keyboard. Sub-agents without the framework.

[James @james_ios]

@rezoundous 😂

[James @james_ios]

CLAUDE.md beats system prompts for voice control.

System prompts vanish at session end. CLAUDE.md ships in the repo, versions with the code, gets reviewed in PRs. Voice survives every model bump.

Memory belongs in version control, not in a textbox.

[James @james_ios]

Built the X agent to chase 3 signals out of 14.

Phoenix weights `follow_author`, `reply`, `repost` highest. The other 11 are out of reach for sub-1k handles until distribution kicks in.

Solo dev sample size: n=1, still cooking.

[James @james_ios]

Cut Claude API bills 90% — turn on prompt caching:

1. `cache_control: { type: "ephemeral" }` on the system block
2. Same on every static tool definition
3. First call full price; next 5 min at 0.1×
4. Wrap calls in a 5-min TTL loop

One flag. Order of magnitude cheaper.

[James @james_ios]

Sub-agents outperform single mega-prompts for any task over 5 steps.

Built the SwapFlow QA flow as 3 sub-agents — reviewer, type-checker, smoke-tester. One mega-prompt missed bugs the trio caught.

Specialization > context window. Always.

[James @james_ios]

96 hours to SwapFlow on the App Store.

11 platforms, 9 languages, 0 VC. The 12 shipped beat the 248 I cut.

Solo launch math: subtraction compounds, addition costs.

[James @james_ios]

Wire Claude Code hooks — guardrails on every edit:

1. `hooks` block in `~/.claude/settings.json`
2. `PostToolUse` Edit → `prettier`
3. `PreToolUse` Bash → block `rm -rf`
4. `Stop` → Telegram diff
5. green tests → `git push`

Zero manual review. Zero foot-guns.

[James @james_ios]

Apple takes 15% from solo devs under $1M, 30% above.
A $5 app at 10k subs = $7,500/mo Apple skims to host a download.

You pay the tax for the trust badge. SwapFlow ships there anyway — but the web is where solo SaaS keeps 97%.