Prevent common coding mistakes with battle-tested agent skills

nodumbmode is a collection of agent skills built from real failures in real projects - a new skill is added only after a concrete failure, never from theoretical best practices. It ships six skills: nodumb, a common-sense manifest and rule set against typical agent mistakes - understand the task first, check its scope, gather facts, then write code - meant to run before nontrivial code, a mass edit, or a second debugging attempt, and outputting briefly what is known, where the risk is, what to do next, and how to verify the result; edge-hunt, which finds corner cases hard to spot by reading a task linearly, by crossing states, data, order of actions, failures, and parallel changes, then turns findings into expected behavior and testable scenarios; changelog-discipline, which records every completed code change - mechanical edits briefly, decisions with their reason, prior state, and rejected alternatives; system-feedback, which checks whether the real system state matches what a human understands from the interface, designing feedback only for genuine uncertainty rather than adding a toast or spinner to every action; ask-nodumb, a design consultant for product and UX tasks that helps surface a hidden assumption and separate the product's image, model, and technology; and nodumb-loop, which adaptively routes a long task between the other skills, choosing only the check that can actually change the next step, fetching an available fact first, and returning not to the start of the process but to the affected level - useful when a task crosses several solution levels, changes mode mid-work, or continues after a failed check, but unnecessary for a local, well-understood edit.

Use nodumb before writing nontrivial code, a mass edit, or a second debugging attempt; edge-hunt when a task's corner cases will not surface from a linear read; changelog-discipline after any completed change worth a record; system-feedback when it is unclear whether the UI actually reflects system state; ask-nodumb for product/UX design questions; and nodumb-loop specifically for tasks that cross multiple solution levels or change mode along the way - not for a small, locally understood edit, where the loop adds overhead without benefit.

Input is the task at hand, handed to the relevant skill; output depends on the skill - nodumb returns a short summary of known facts, risk, next step, and verification method; edge-hunt returns expected behaviors and testable scenarios; changelog-discipline returns a change record; system-feedback returns a feedback design; ask-nodumb returns design guidance; nodumb-loop returns routing decisions between skills as the task progresses.

npx skills@latest add hanumatori/nodumbmode

Requires Node.js. The CLI asks whether to install globally or into the current project, detects installed agents (Claude Code, Codex, Cursor, Windsurf, Continue, and dozens more supported equally), and lets you pick where to install. For non-interactive installs (CI or scripts), name everything explicitly: npx skills@latest add hanumatori/nodumbmode -g -y -a claude-code -a codex --skill '*' (-g global, -y no prompts, --skill '*' all skills, -a target agent) - --all is not recommended, since it targets every known agent including ones without global-install support and errors on each. Skills are written once to ~/.agents/skills/, with symlinks placed in each agent's directory, so an update reaches everywhere at once with no duplication. Update named skills explicitly (npx skills@latest update nodumb-loop nodumb edge-hunt ask-nodumb changelog-discipline system-feedback) rather than bare update, which updates every installed skill including ones from other authors. If a skill does not appear for an agent, the CLI may have copied files without linking them - link manually with ln -sfn ~/.agents/skills/<skill> ~/.codex/skills/<skill> (symlinks, not copies, so future updates still apply automatically).

Developers who want agent behavior corrections grounded in actual project failures rather than generic best-practice advice, applied before risky edits, debugging retries, or long multi-stage tasks.