Documentation and workflow hacks from practitioners

One pattern keeps showing up in developer threads this week: teams are trying to stop one senior person from becoming the human router for every approval, every note, and every undocumented rule. Anthropic’s Claude Code docs now explicitly describe a project-level `CLAUDE.md` plus optional `rules`, `skills`, `commands`, and `agents` folders inside `.claude`, which gives those teams a concrete place to put operating knowledge instead of leaving it in chat history or one person’s head. (code.claude.com) The core file is simple: `CLAUDE.md` is loaded at the start of every Claude Code session, and Anthropic says it should hold project conventions, common commands, and architectural context. The docs also say to keep it under about 200 lines and move task-specific material into a skill or a path-scoped rule, which is the software version of keeping a handbook short and putting the long procedures in labeled binders. (code.claude.com) That structure changes how context is reused. A build command like `npm run build`, a naming rule like “named exports only,” or an application programming interface rule like returning `{ data, error }` can live in a file Claude reads every session, instead of being re-explained in every prompt by every engineer. (code.claude.com) The folders matter because they split permanent rules from occasional procedures. Anthropic’s directory guide shows `rules/` for scoped instructions, `skills/` for reusable workflows, `commands/` for slash-command shortcuts, and `agents/` for specialist subagents, so a team can separate “always do this in this repo” from “run this checklist only when reviewing security.” (code.claude.com) A second idea in these practitioner posts is approval thresholds. The point is not to remove review; it is to stop low-risk actions from waiting in the same line as high-risk ones, which is the same risk-based logic the United States Food and Drug Administration uses in its February 3, 2026 guidance on Computer Software Assurance for production and quality management system software. (fda.gov) That guidance tells companies to focus assurance effort based on risk, not to test everything with the same intensity. In practice, that gives regulated teams cover to say a formatting change, a routine summary, or a low-impact update should not need the same human signoff path as a change that affects safety, quality, or a regulated record. (fda.gov) The regulated-industry angle is why these documentation hacks are getting attention outside software circles. The Food and Drug Administration’s current guidance on artificial intelligence used for regulatory decision-making says sponsors should think about credibility, risk, and documentation when artificial intelligence produces information that supports safety, effectiveness, or quality decisions. (fda.gov) That means a team cannot just say “the model wrote it.” They need a trail that shows what instructions existed, what workflow was followed, and which outputs were reviewed, and a machine-readable folder tree does that better than scattered meeting notes and private direct messages. (fda.gov) (code.claude.com) The third pattern is note cleanup. Microsoft’s prompt engineering guidance says output quality improves when prompts specify the task, the context, and the format, which is why people are turning raw call notes and half-finished bullets into prompts that ask for a clean summary, action items, owners, and open questions in a fixed template. (learn.microsoft.com) OpenAI’s structured outputs guide pushes the same idea from the application side: if you define the shape you want, the model is more likely to return something a workflow can actually use. For overloaded teams, that turns “here are my messy notes” into a repeatable handover document that can be stored, searched, and checked later instead of rewritten from scratch by the next person. (platform.openai.com) Put together, the playbook is pretty concrete. Keep the standing rules in `CLAUDE.md`, move long procedures into `skills` and `rules`, let low-risk work clear under pre-set thresholds, and force meeting notes into a standard summary format, and you get fewer bottlenecks, faster onboarding, and a smaller chance that one absent employee takes half the team’s memory with them. (code.claude.com) (fda.gov) (platform.openai.com)