AI Agents Quit Early. This Skill Won't Let Them.

What it does

unlazy is a prompt-engineering discipline plus a small set of Node scripts that combat model laziness—the tendency of coding agents to truncate work, ship invisible defects, or declare premature victory. Before writing code, the agent must write concrete acceptance criteria to disk as markdown checklists with runnable CHECK commands. A script verifies each gate by running the command and matching output against an EXPECT string, recording the result as evidence. For Claude Code, an optional hook can mechanically block the agent from ending its turn while gates remain open.

The interesting bit

The method treats agent laziness as a structural problem, not a prompting problem. v1 begged the model to try harder; v2 makes half-finished work structurally visible by forcing the agent to decompose tasks into a depth tree, fix contracts in PLAN.md before fanning out, and verify leaves as fresh subagents. The project’s own controlled test found that frontier models already avoid obvious placeholders, but still ship wrong numbers in final reports—exactly the invisible failure mode the gate-and-evidence ledger targets.

Key highlights

• File-based enforcement: acceptance gates live in GATES.md and gates/*.md, checked by a zero-dependency Node script that matches command output against expectations.
• Depth Tree decomposition: tasks split N layers deep with contracts fixed in PLAN.md before work fans out; leaves run as fresh subagents to dodge long-context erosion.
• Hard stop hook for Claude Code: a millisecond file scan blocks turn completion while gates are unmet, costing zero tokens per check.
• Measured, not guessed: the author’s controlled build-from-scratch tests showed the skill raised effort 1.6–3.9× and caught 4–10 self-found defects pre-delivery.
• Research-backed: cites 2025–2026 papers on LLM underthinking, context anxiety, and premature completion as the grounding for the problem.

Caveats

• Hard enforcement is Claude Code only; other agents get the discipline layer but no mechanical stop hook.
• The depth tree’s original promise that effort multiplies with depth turned out to be fiction—the author’s own testing showed tree 6 cost only 1.0–1.5× tree 3, not the theorized 8×.
• Orchestrated multi-agent mode is deliberately expensive and meant for substantial builds; the docs suggest staying in solo mode for tasks under roughly half an hour.

Verdict

Worth a look if you are tired of AI agents declaring victory while leaving broken math and unfinished logic in their wake. Skip it if you only use agents for trivial one-shot edits where a simple “please double-check” prompt is already enough.