AI Engineering Workflows

A running, informal thread throughout deep-dives and random covered how experienced technology professionals (including an industry expert himself) actually use AI coding tools day to day:

• Tool combinations discussed: Claude Code, Claude Desktop, and Codex used together with distinct roles (e.g., one an experienced professional's proposed split: Codex for verification/implementation, Claude Desktop for strategy/memory/tooling, Claude Code for repo-level work). an industry expert confirmed he personally uses both Claude Code and Codex, description of Codex (specifically referencing "5.5") as "really good once you get used to it," with the caveat that switching between models requires re-adapting your prompting style to each model's behavioral quirks. He also called himself "a huge subagents enjoyer" - describing subagent-based workflows as an effective paradigm, though he personally prefers writing his own custom skills rather than installing the field-made ones, citing that writing skills yourself, highly tailored to your own workflow and preferences, is both higher value and a better learning exercise than adopting someone else's off-the-shelf skill set.
• Obsidian as a "personal wiki" feeding into agentic coding tools: one experienced professionals described years of maintaining a personal Obsidian knowledge base (initially unrelated to AI) that turned out to make it much easier to "plug in" Claude Code/Codex effectively once agentic tooling became relevant - framed as a reason Obsidian has recently become popular specifically because of the agentic AI wave, not despite it.
• Cost/tooling-tier discussion: experienced professionals debated whether upgrading from free-tier Claude to a paid plan (~€20/month) is worth it versus alternatives like DeepSeek, generally concluding it depends on your specific use case and how often you hit usage limits - with some experienced professionals expensing higher tiers (e.g., a $100/month Codex plan) as a business cost when the work justifies it.
• On enterprise-scale agent usage: a shared anecdote described real-world Claude usage patterns involving on the order of 50 subagents spun up and ~1M tokens consumed for a single task, prompting discussion that such heavy subagent fan-out is realistically built for enterprise workloads, not typical consumer-tier usage.
• On internal company tooling vs. commercial AI coding tools: expert-led own experience at Amazon was that internally-built developer tooling was often worse than what was externally available, which led to internal pushback and eventually broader internal approval to use Claude Code - framed as a broader pattern: it's very hard for individual company internal-tools teams to out-build dedicated AI coding companies, except where licensing cost specifically justifies self-hosting/forking open-source alternatives instead.
• A concrete, fully-described AI-assisted video production pipeline (shared by an industry expert in design-reviews, in response to a question about how his agentic video-editing setup works): built as a set of custom Claude Skills wrapped around Remotion (a framework for programmatically generating video/animation using JavaScript rather than manually driving something like FFmpeg or After Effects). End-to-end flow:
  1. Manually cut down bad takes and remove dead air from the raw recording.
  2. Claude runs a script to transcribe the video locally using Whisper.
  3. Claude does deep research on the video's topic via a web-search tool, grounded in the transcript.
  4. Claude spins up a Remotion server, and - using the edited video plus a large set of custom rules/guidelines refined over months of evals - generates a new animation roughly every 3 seconds.
  5. The pipeline produces 3 full variations of the finished video, and an industry expert manually picks the best one.
  • Every video on the associated Instagram account (instagram.com/mavenhq) was produced with this exact pipeline. A experienced technology professionals specifically noted being surprised the animations were generated programmatically via JavaScript rather than via FFmpeg-based editing - worth flagging as a genuinely distinctive technical approach compared to typical AI video-editing setups.
• On architectural drift as agentic coding scales: an experienced professional flagged a broader open industry question worth tracking - how do you prevent a codebase's actual architecture from slowly drifting out of alignment with its intended design when local tests still pass, especially as more of the implementation work shifts to agents rather than humans who "feel" the codebase evolving.