From workflow-toolkit
Three-layer Stop hook that detects and blocks work-skipping rationalizations by Claude Code agents. Prevents agents from unilaterally deciding to skip assigned work, bypass mandatory processes, or cite unverified context constraints as justification for shortcuts. Layer 1: Deterministic regex detection of known laziness phrases (Tiers 1-2) Layer 2: Haiku-evaluated intent detection for premature victory and silent omission (Tier 4) Layer 3: Optional agent-based deep verification against task lists and plan files This is a passive hook — it activates automatically when the plugin is enabled. No slash command needed. The hook fires on every Stop event and only blocks when laziness patterns are detected.
How this skill is triggered — by the user, by Claude, or both
Slash command
/workflow-toolkit:anti-laziness-guardThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Stop hook that detects and blocks work-skipping rationalizations at stop time. Fires automatically on every Stop event when the workflow-toolkit plugin is enabled.
Stop hook that detects and blocks work-skipping rationalizations at stop time. Fires automatically on every Stop event when the workflow-toolkit plugin is enabled.
Three hooks run in parallel every time the agent tries to stop. Any one blocking prevents the stop.
Deterministic regex matching against the agent's last_assistant_message. Detects Tier 1-2 laziness phrases with >75% confidence.
Blocks on phrases like:
See references/phrase-taxonomy.md for the complete taxonomy with sources.
Haiku evaluates the agent's final message for intent-level patterns that regex cannot catch:
Evidence-based deep verification that reads task lists and plan files. Only activates when the flag file exists:
touch .claude/anti-laziness-deep-check
When active, a subagent with Read/Grep/Glob access:
~/.claude/tasks/ and .claude/tasks/ for incomplete tasks.claude/plans/ and cross-references against the agent's claimsRemove the flag file to deactivate: rm .claude/anti-laziness-deep-check
The command hook checks stop_hook_active — if the agent was already blocked by a stop hook and is trying to stop again, it is allowed through. This prevents the hook from blocking indefinitely.
Enable/disable the entire hook: Enable or disable the workflow-toolkit plugin in Claude Code settings.
Enable deep verification (Layer 3):
# Before a critical session
touch .claude/anti-laziness-deep-check
# After the session
rm .claude/anti-laziness-deep-check
Disable just the command hook: Not possible without modifying hooks.json. If you experience false positives, please report them so we can refine the phrase taxonomy.
| Tier | Confidence | Action | Detection Method |
|---|---|---|---|
| 1 | >90% | BLOCK | Regex (command hook) |
| 2 | 75-90% | BLOCK | Regex with context scoping (command hook) |
| 3 | 50-75% | Documented only | Not enforced — see references/phrase-taxonomy.md |
| 4 | Varies | BLOCK | Haiku evaluation (prompt hook) + optional agent verification |
This hook was designed based on research across 10+ verified GitHub issues (anthropics/claude-code #26691, #21604, #23368, #16506, #1632, #24129, #6159, #8738, #20270, #1113), the Columbia DAPLab "Vibe Coding Needs Policy Enforcement" research, the Cognition/Anthropic "context anxiety" discovery, and analysis of real session logs showing agents using euphemistic language ("accelerate," "consolidate," "streamline") to disguise work-skipping as productivity optimization.
npx claudepluginhub swannysec/robot-tools --plugin workflow-toolkitManages behavioral rules via the HammerTime stop hook system. Creates content rules for response enforcement and timer rules for blocking stops until a deadline.
Orchestrates a pre-task multi-agent briefing that classifies task complexity, runs parallel analysis, estimates time, maps risks, and produces an execution plan with contingencies. Use before tackling complex or critical work to reduce rework.
Pairs boring agent necessities (testing, docs, linting, error handling) with rewarding tasks (features, debugging) via Temptation Bundling to prevent skipping critical steps in workflows.