ayghri/i-have-adhd

The premise here is simple and the frustration it addresses is real: coding assistants default to a style that buries the actionable part under preamble, tangents, and hollow closers. This repo is a Claude Code skill — a structured prompt file — that enforces a different contract between the assistant and the reader.

The ten rules are the whole product. Lead with the next action. Number multi-step tasks. End with one concrete next step. Cap lists at five items. Suppress tangents. No preamble, no recap, no closers. The before/after example in the README makes the problem vivid: the "before" response is a paragraph of throat-clearing that eventually arrives at actionable advice; the "after" opens with the exact command to run and a numbered three-step sequence. The difference isn't intelligence — it's structure.

The framing around ADHD is honest about its own scope. The rules are loosely adapted from a clinical workbook on adult ADHD, but the README explicitly notes the adaptation is for how an LLM should respond, not how a human should organize their day. That's a meaningful distinction. These rules aren't a cognitive accessibility accommodation in the clinical sense; they're a response-formatting discipline that happens to match what anyone — with or without an ADHD diagnosis, as the tagline says — actually wants from a coding assistant under time pressure.

Installation is a single copy-paste into the CLI prompt, which points the assistant at the repo's AGENTS.md. The skill lives in a markdown file, so forking and editing it is the intended customization path. The tuning instructions show how to swap in a personal fork via the Claude plugin system, which means the rules are fully user-modifiable.

What's worth noticing is rule five: restate state every turn. That's not just a formatting preference — it's a recognition that conversational context degrades across turns and that an assistant restating where things stand is doing real work, not padding. Rule six — specific time estimates in minutes, not vague qualifiers — is similarly pointed. Vague hedges like "this might take a bit" are useless; a number, even an approximate one, is actionable.

The repo won't change how the underlying model reasons. It changes what the model chooses to surface first and what it chooses to omit. For anyone who has watched a capable assistant bury a one-line fix inside three paragraphs of context they didn't ask for, that's a meaningful intervention.