Drafts verifiable /goal objectives with binary done-when checks, guardrails, and budgets.
npx skills add https://github.com/QwenLM/qwen-code --skill goal-draft
Please help me install this Agent Skill. Skill: goal-draft Source: https://github.com/QwenLM/qwen-code/tree/main/packages/core/src/skills/bundled/goal-draft Command: npx skills add https://github.com/QwenLM/qwen-code --skill goal-draft
Vague goals like "make it faster" or "clean up the module" cannot be judged by an automated verifier, so agent loops run until they hit a limit without ever finishing. This Skill turns a fuzzy intention into a structured /goal objective that an independent verifier can actually judge from transcript evidence.
npm run bench:checkout, and drafts an objective requiring the benchmark to exit 0 with p95 below a threshold, forbidding changes to the benchmark itself, and stopping as blocked after 20 turns.Ask the agent to run /goal-draft followed by what you want done, for example turning "keep going until the tests pass" into a verifiable goal objective.
High-intent search queries and answers about installing and using this skill.
Define one Outcome plus numbered binary "Done when" checks where at least one is tool-observable, such as a command exit code, an output line, or a grep match. Add Must-not guardrails, a turn or time Budget, and an On-block protocol so the loop stops cleanly instead of running forever.
Run /goal-draft followed by what you want done, or pass an existing goal to tighten it. The skill inspects the workspace for real check commands, asks at most one round of clarifying questions, then hands you a /goal set line or proposes the goal for approval.
The verifier sees only transcript evidence: visible assistant output and tool results. Claims about tests passing or files changing need an actual tool result, and claims about user approval need a real user message, so objectives must force the agent to paste decisive output lines.
Skip goal-drafting for normal one-shot tasks, requests needing design or product judgement, or anything with no way to be checked from the agent's own output. A goal that cannot be checked is a prompt, not a goal, and will just run until a limit is hit.
No. The skill only drafts the objective text and the /goal set line; it never edits files, runs the checks, or starts the task. Its allowed tools are restricted to read-only operations: get_goal, read_file, glob, and grep_search.
LerianStudioRequire fresh verification command output before claiming task completion.
patleemanMaintain a single active objective across turns with automatic scheduling.
Cortexa-LLCMonitor spawned AI agents and intervene with corrective guidance during multi-agent orchestration.
breakingcircuits1337Verify command outputs before claiming task completion.