ayghri/i-have-adhd

The premise here is blunt and accurate: coding assistants default to a response style that buries the actionable content under preamble, tangents, and cheerful closers. This repo is a Claude Code skill — a structured prompt ruleset — that retrains that default behavior for any session where you install it.

The mechanism is ten rules encoded in a SKILL.md file. They are not vague style suggestions. They specify concrete behaviors: lead with the next action, number multi-step tasks, cap lists at five items, give time estimates in minutes rather than qualitative hedges, end every response with exactly one concrete next step. The before/after example in the README makes the gap vivid. The "before" response meanders through context, names the relevant file and lines, floats a possible approach, suggests a tangential audit of dependency versions, and closes with "Hope this helps!" The "after" response opens with the exact command to run, numbers three steps with precise file locations, and ends with a conditional instruction for the failure case. Same information, radically different cognitive load.

The framing draws loosely on a clinical text about adult ADHD — Ramsay and Rostain's toolkit — but the README is honest that the adaptation is for LLM output behavior, not human self-management. That distinction matters. The rules aren't about accessibility in the sense of accommodating a disability; they're about eliminating a class of LLM verbosity that wastes everyone's time regardless of neurotype. The name is a provocation, not a diagnostic claim.

Installation is a single paste into a CLI prompt, which tells Claude Code to fetch and apply the skill from the repo's AGENTS.md. Customization is a fork-and-swap: edit the skill file, uninstall the upstream copy, reinstall from your fork. The tooling is Claude Code-specific, so this won't port directly to other agent environments without manual adaptation of the plugin installation steps.

What's genuinely useful here is the specificity. Most "be more concise" prompts degrade over a conversation as the model reverts to its defaults. A persistent skill file that re-invokes these constraints per session is a more durable intervention. Rule five — restate state every turn — is particularly well-considered for multi-step debugging sessions where context drift is the actual failure mode.

The repo is small and the concept is simple. But the ten rules are well-chosen, the example is honest about what changes, and the fork-to-customize path means the constraints are actually yours to own rather than a black box.