Examples:
EXEC-B001: Execute command mandatory behavior 1
DEF-P002: Define command prohibited behavior 2
</example>
<error_escalation>
Minor inconsistency in behavior ID format
Note in report, proceed
Missing one boundary test
Document issue, add missing test
A decision_criteria factor names a quality to be rated rather than an observable unmet
condition, so no reader can check which factor decided
STOP, restate the factor as a condition that can fail and give it a precedence before
proceeding
Error escalation missing critical level
BLOCK operation, require complete 4-level structure
</error_escalation>
<anti_patterns>
Writing a bare cross-reference — an inherits="skill#anchor" attribute, a refs entry,
a "see core-patterns" note — where the content itself belongs, on the assumption that something
resolves it. Nothing does. The file is then carrying an empty slot that reads to every later
reader as if it were filled, which is worse than an obviously missing section.
State the four levels and their domain-specific examples in the file itself, and load
core-patterns with the Skill tool when the shared template is what you actually need
(skill_loading).
<best_practices> Load core-patterns with the Skill tool and write the error_escalation, decision_criteria, and enforcement structures into the file itself — a reference left in their place delivers nothing (skill_loading) Tag every finding with an evidence tier — verified, inferred, or assumed — and never promote a tier to strengthen a report (evidence_tiers) Write gates as conditions that can fail, not as scores to clear (decision_criteria, numeric_self_assessment) Customize error_escalation examples to be domain-specific while keeping structure Use consistent behavior ID prefixes within each agent/command Escalate a single-pass review into an independent, skeptical refutation only for consequential findings, and weigh it against its own cost — false positives, lazy-validation risk, token cost, and shared blindspots (adversarial_verification_escalation) Model absence structurally instead of with an in-range sentinel, and never infer "unset" from a value inside the valid domain (presence_vs_value) Mirror a worktree into the shared checkout with a file sync rather than a branch switch, and remove a worktree only once its diff against the target is empty (parallel_project_isolation) Derive a strategy-driving cost estimate from the emitter itself, and have tests consume the same function (estimator_derived_from_emitter) When two related rulesets appear to contradict, find the distinguishing condition and add a reconciling note before weakening either (apparent_rule_conflict_resolution) </best_practices>
<decision_tree name="skill_activation"> Does the task clearly match this skill domain? Use this skill workflow and constraints Use a more appropriate domain skill </decision_tree>
<related_agents> Locate code patterns and references for this domain Review implementation quality against this skill guidance </related_agents>
<related_skills> Parallel execution strategies and timeout configuration Output formats, reflection checkpoints, agent weights For memory operations to store pattern decisions </related_skills>