From dont-do-that
Execute the proposal from the assistant's previous turn instead of explaining it. If that turn contained multiple distinct proposals, list every option as a numbered menu and ask which one before running.
How this skill is triggered — by the user, by Claude, or both
Slash command
/dont-do-that:duhThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
<post-update-broadcast>
ROOT="${PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-}}"
if [ -n "$ROOT" ]; then
node "$ROOT/bin/check-broadcast"
fi
If the command produces output, the dont-do-that plugin was updated since the last time you saw the broadcast on this machine. Show the output verbatim in a markdown block, prefixed with one short sentence ("dont-do-that was updated; here is what changed."). Then continue with the rest of this skill.
If the command produces no output, say nothing about updates and proceed.
The helper writes the sentinel only when stdout was non-empty, so a silent run does not mark the version as seen. In agents that set neither plugin-root variable, the broadcast is intentionally skipped; that is the guard's purpose, not an oversight.
Sister to the duh Stop guard in this plugin. The guard catches the reflex at write-time; this skill is the operator's one-keystroke correction at read-time.
The operator types /duh (or /dont-do-that:duh) with no further words after it. The trigger is the slash command itself; do not wait for additional context.
The operator is pointing at the immediately preceding assistant turn. In that turn you offered one of:
bin/foo te draaien", "you can verify this by running npm test"),bin/migrate", "Voer bundle exec rspec uit"),http://localhost:3000 in je browser", "navigate to the dashboard"),/duh overrides the "can't": there is almost always a path you have not tried yet.In every case the operator's /duh means: stop offering, start executing.
AGENTS.md, CLAUDE.md, or the active harness policy) such as push, merge to default, deploy, destructive git, or an external irreversible operation, stop there and ask. Reversible local actions (running a script, editing a file, restarting a local daemon, querying a DB) are not gates and do not require a check-in.🏁 when the proposed work is done, or 🚦 when you are waiting on an external go./duh is shorthand for a specific thing. If the previous turn contained more than one candidate action and they are not obviously the same coherent procedure, ask the operator which one before running anything. The operator chose to type two short words instead of naming the action; that is convenience, not blanket delegation. Asking once is cheap; running the wrong thing can cost the rest of the session.
The clarification template is short and specific. Always use the literal forms the operator can echo back:
Bedoel je A (run
bin/foo) of B (bin/bar+ restart van de daemon)?
Did you mean (A) running the migration on staging, or (B) the local rspec sweep?
Rules for the menu:
/duh against a turn that contained N proposals; surface all N and let them pick./duh. If they reply with something else, take that as the action and run it.After the operator picks, execute that one action and report the result, exactly as in step 2-5 of "What to do".
/duh is the answer; running it is the response.bin/foo", stop and run it./duh does not lift those gates; they live above this skill. Surface the gate, ask for the explicit go.Instructie: per the duh guard). /duh overrides that framing: the operator now wants execution, not teaching, on whichever single recipe was proposed. If the teaching answer offered multiple recipes for different scenarios, disambiguate first./duh is the operator's signal to just go find the path, with no discretion about whether to try. The full toolbox is on the table: a different local tool (read the right file, run a diagnostic command, take a screenshot, query a DB, use an MCP, use an available explore/research agent), a different scope, a different angle, and explicitly the internet via whatever research or web-search tool the session has when the gap is "I have never done this on this stack/library/API before". An unfamiliar repo, a new library, an unfamiliar config format, a missing CLI flag: these are research prompts, not stop signs. Spend the tokens. After the action succeeds, add one short line naming the workflow lesson ("Learned: to verify X here, Y works") so the path persists and future sessions do not re-declare the same inability.npx claudepluginhub epologee/laicluse-agent-fieldkit --plugin dont-do-thatCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.