Leader — One-Line Ideas to Agent-Ready Briefs

Why this exists

Hand an agent a raw goal (“refactor the auth module, make the tests pass”) and you get the classic failure bouquet: hallucinated commands, quietly skipped tests, a fresh session that redoes finished work, and a final “done” you can’t verify. The failure is rarely the agent — it’s the brief. Leader is the management layer that fixes the brief:

• Research before writing — commands verified by running them, baselines measured, doc-vs-reality gaps documented. Nothing in the brief is made up.
• One round of ≤5 questions — only what can’t be looked up and would change the outcome, each with options and a recommendation. Unanswered decisions are recorded in the brief as “Decisions I Made for You,” defaults visible and overrule-able.
• A ≤4,000-character brief (the /goal hard limit) with measured baselines, a path whitelist, a frozen grading standard, anti-gaming acceptance commands, reverse tests for silent failures, and PROGRESS.md / BLOCKED.md mechanics so any session can resume mid-flight.
• Two brief types — execution briefs (verifiable acceptance commands) and exploration briefs (research / selection), with different completion logic for each. Exploration briefs grade on evidence-backed conclusions, and a proven dead end counts as a pass.
• Grading included — an open rubric inside the brief plus hidden spot-checks the executor never sees, and a ≤5-line plain-English report when the run lands.

What’s in the package

• SKILL.md — the full workflow: research → questions → brief → delivery → grading
• references/anatomy.md — the six-section structural spec every brief follows
• references/style.md — an optional humor corpus (off by default; the briefs stay plain unless asked)