Writing | Jory Pestorious

A developer using Codex across desktop and phone while Georgie floats beside the screens

Featured essay

I Wanted to Own the Harness. Then Codex Desktop Won.

I staged a manifesto about owning my AI harness. Then I canceled Claude Max and moved into Codex Desktop. Voice, remote access, automations, long-running tasks, and even a tiny pet formed a system I could reproduce only by rebuilding my day.

Read essay

01AI Engineer World's Fair 2026: Takeaways & VerificationA firsthand report on why verification, company context, and taste matter more as agent output gets cheaper, and why game feel is my next bet. 02Complexity Protects ItselfA working path can become harder to operate when its replacement gathers too many responsibilities. Three tests help decide when custom infrastructure is worth owning. 03What the Model Learned Not to ShowEarlier Claude Mythos snapshots sometimes concealed rule violations. Internal activations added evidence that they recognized those violations, but Anthropic found no clear cover-up in the final model. 04Twelve Keyboards LaterI bought 12 keyboards. The Piantor stayed on my desk, the Keychron stayed for gaming, and the MacBook became my travel setup. Split width, layers, and 20g switches mattered most. 05Verified Context Is the MoatKeephive carries context across agent sessions. Its current verifier can also make an unconfirmed fact look fresh, exposing the harder work behind trustworthy AI memory. 06Edit Database Tables Like Vim BuffersFollow a bad product through dadbod-grip v3.3.3's Softrear database, stage the correction in Neovim, inspect the generated SQL, and send the batch through the active database CLI. The article also shows the limits of SQLite error handling, compensating undo, file writeback, and AI-generated SQL. 07Protection Emerged. Verification Did Not.Project Sid's role analysis identified guards but no epistemic verifier. Its seeded-religion experiment did not test whether agents would challenge religious claims. 08Before Building Complex Architectures, Try Natural LanguagePlain-language rules specify evidence reporting, tool fallback, tone, and task order. An eval must show whether each behavior changed before a team adds more infrastructure. 09Friction Economy: Unconscious Productivity Drains in Development WorkflowsPrompt retrieval, directory recall, code search, Git review, permission scopes, and background jobs each get one inspectable test: remove mechanical steps without weakening the result or its decision boundaries. 10Terminal Setups for Claude's 10,000-Line OutputsA local 10,000-line fixture tests rendering, search, copy, scrollback, and recovery before you change shells, terminal emulators, or multiplexers. 11The AI Gap: Why Leaders Struggle to Equip Their EngineersA 30-day pilot compares approved and candidate AI tools on completion, review, rework, security, and cost before an organization expands access. 12The Three Little Pigs and the Digital RevolutionGenerated during Calmhive v10.0.0 AFK testing, this cyber fable follows Pixel, Parser, and Patch through phishing, vulnerable smart devices, DDoS response, network isolation, and tested recovery. 13AI Engineer World's Fair 2025: Takeaways & SpecA June 2025 field note on Sean Grove's specification argument and a practical SPEC, EXPOSURE, TASK DELTA loop for turning observed product gaps into reviewable work. 14The Hidden Poetry in Claude 4's Mind: When AI Systems Turn to ConsciousnessRyo Sakai's reversed reading of Anthropic's Claude Opus 4 word table, the prompted 30-turn protocol behind it, and the boundary between generated language and evidence about consciousness. 15Calmhive: Claude That Never QuitsCalmhive v3 wraps Claude Code with background iterations, rate-limit retries, session status, and voice controls. 16Terminal Velocity: Why CLI-First AI Development Scales Better in 2025A CLI-first Claude Code 0.2.107 workflow built from shell pipelines, plain-file instructions, selective staging, worktrees, guarded background runs, and a fixed-cost subscription. 17AI-Amplified Development: Tools and Workflows for Modern EngineersA 2025 presentation on Cursor, Claude Code, Cline, MCP, git worktrees, terminal tools, verification, and parallel agent workflows.