Field guide / sub-domain
9
sources in this field
Updated July 15, 2026
Current thesis
The shortest path to orientation.
Voice input captures rough thought before polished compression, preserving uncertainty, emphasis, and unfinished lines of thought. Raw meeting transcripts and dictated planning notes outperform short summaries as Claude source material because the model fills transcription gaps and works from messy input. Monologue (from Every) or WhisperFlow pipes speech into any focused app; a gooseneck microphone is recommended. Mac Mini with Telegram integration enables mobile Claude Code access—send commands like /ce:plan fix the timeout issue from anywhere, with tmux sessions surviving bad WiFi. Voice modeling is a reusable skill workflow: analyze several weeks of substantive, audience-separated Slack and email to infer what the person challenges, notices, and ignores; require example-supported patterns; show drafts before saving; update from gaps between draft and actual sends. Claude Sonnet 4.5 replicates individual writing voices with 94% accuracy in blind tests across literary, essay, and professional styles, enabling executives to delegate communication while maintaining authentic voice. Build reusable skills (e.g., $tibo-voice) from these patterns. Voice's value is fidelity, not speed: the agent receives unedited, messy thinking, giving it richer context than polished typed prompts.
Evidence board
Claims worth carrying forward01Pinned megathreads (one per workstream: Chief of Staff, Agents SDK, Twitter monitor, etc.) accumulate history and preferences across months. Access via Command-1 through Command-9. Tradeoff: long threads likely fall out of cache, incurring higher cost than fresh short threads—continuity is worth it for important workstreams.
02Heartbeats are thread-local automations that schedule recurring checks without human presence. Example: Chief of Staff thread runs every 30 min to scan Slack and Gmail, drafts replies but never sends them. A single loop can cross tool boundaries—Slack feedback → Remotion render → @computer for file upload—without manual intervention between steps.
03Agent memory should live as files in an Obsidian vault (kept as a GitHub repo), not just as conversation history. AGENTS.md at vault root instructs agents to update people/, projects/, and notes/ pages as they learn. GitHub diffs become a review surface showing what the agent judged important enough to persist—preventing silent accumulation of 'vibes' in chat history.
04Voice input's value is not speed but fidelity: the agent receives the unedited, messy version of your thinking. Combining Wispr Flow (system-wide dictation) with Codex's built-in voice, and piping call transcripts via Granola, gives the model richer context than polished typed prompts ever would.
05Steering lets you inject the next instruction after each tool call while the agent is still working, building a queue of intent before walking away. Heartbeats then monitor long-running outcomes (PR state, Slack replies) after you leave—converting 'one prompt, one answer' into a persistent operating loop.
06The side panel is where Codex becomes a work surface rather than a chat app: it renders Markdown (commentable), spreadsheets (formula-aware, cell-editable), CSVs, PDFs (LaTeX-ready), and slides. The in-app $browser lets the agent see and control the same web surface you are annotating—closing the inspect-edit loop without switching tools.
07Goals need a verifiable oracle, not just an intent. Weak: 'implement the plan in this Markdown.' Strong: 'migrate Python Rich to Rust and pass all original unit tests.' The test suite acts as a continuous success criterion the agent can push against autonomously—ambition without verification is just a wish.
08index.html is preferred over Markdown as Codex output format: it is a self-contained application, requires no server, opens instantly in the side panel, and can be updated by Heartbeats so a fresh artifact is waiting on thread return. Use Vite only when heavier tooling is unavoidable—it requires a persistent server.
Adjacent fields
Latest evidence
jason
Jason Liu's workflow guide for getting maximum leverage from OpenAI Codex by treating it as a persistent work OS rather than a one-shot chat tool. Key primitives: durable pinned threads (Command-1 through Command-9), voice input for unedited thinking, steering to queue intent mid-execution, an Obsidian vault as shared agent memory, Heartbeats for recurring loops, and the side panel for artifact inspection and annotation.
jason
Codex has matured from a coding assistant into a general work OS. Core patterns: durable pinned threads (Cmd+1–9 shortcuts), voice input for vague prompts, steering (mid-task interruption) vs. queuing (next-task scheduling), tool layers ($browser/$chrome/@computer), thread automations on a schedule, Goals with explicit verifiers (test suites, benchmarks), side panel for in-place artifact review, and shared memory via Obsidian vault with AGENTS.md routing rules.
Samantha Trimble
> another codex hack: for internal facing comms, make it read the last 3 weeks of slacks from someone at your company you trust and admire. ask it to understand what he/she pushes back on, what piques their interest, what they ignore, how they communicate with their team, etc. create a skill that yo
Nick Spisak
> https://t.co/ykuoaf5dhH - Use the [[claude/workflow]] capability of the Claude AI agent to streamline your coding processes and boost productivity — leverages [[claude]], [[vibe-coding]] - Integrate the [[claude/voice]] feature of Claude to narrate your coding steps, providing hands-free assistanc
rLLM
> https://t.co/Dl67Q7wMgx - The rLLM project has released a new version, which includes updates to the [[claude]] architecture, design, and ecosystem - The new release introduces support for [[voice-tools]] and [[physical-ai]] integrations, enabling new use cases for the [[claude]] language model
Nick Khami
> https://t.co/eSZGpkEGU4 - Use the [[claude]] plugin for Obsidian to seamlessly integrate with the Claude AI agent, enabling natural language interactions and advanced AI-powered capabilities within your Obsidian workflow - Configure the [[claude/settings]] to customize the Claude agent's behavior,
Griffin Hilly
> https://t.co/HWyMFZK7Ob - Leveraging [[claude/workflow]] to automate research, writing, and ideation within the [[.obsidian]] note-taking environment - Integrating [[claude/voice]] for hands-free note capture and command execution during the research process
Manthan Gupta
> https://t.co/em8mMn8bhj - The tweet provides a link to a detailed overview of Claude, an AI agent with a diverse set of capabilities including [[claude/architecture]], [[claude/design]], [[claude/ecosystem]], [[claude/settings]], [[claude/voice]], and [[claude/workflow]]. - The overview discusses
Shiv
> Lots of companies are now building primitives for an economy where AI agents are the primary users instead of humans. > > They're betting on an economy of AI coworkers.