Skip to content

Arcadi4/skills / skill-fails

skill-fails

Verified

Use when an agent violates a loaded skill despite it being present in context — a gap exists between what the skill says and what the agent did. Also use when reviewing skill effectiveness after a violation, or when deciding whether a skill needs enhancement vs the failure is tolerable.

Identify Skill Failures

Overview

When an agent violates a loaded skill, a gap exists between the skill's instructions and the agent's behavior. This skill provides a 4-step post-mortem to close that gap with a targeted enhancement — or decide the failure is tolerable.

When to Use

  • Agent violated a skill that was loaded in context
  • Mismatch between skill recommendation and agent action
  • Deciding: enhance the skill, or tolerate?

Do NOT use when: skill wasn't loaded (teach, don't fix), agent followed skill but got wrong result (skill is wrong, not unenforced), or context couldn't contain the skill at all.

Core Pattern — 4-Step Post-Mortem

StepQuestion
1. Capture violationWhat did the agent do vs what does the skill say? Quote both.
2. Classify gapWhy did the instruction fail to prevent the behavior? Use taxonomy below.
3. Assess tolerabilityContext window % at failure × severity. Fix or tolerate?
4. Propose enhancementOne targeted edit to close that specific gap.

Gap Taxonomy

GapSymptomFix
Passive languageRule stated but no enforcement gate. Agent reads, doesn't internalize.Add a concrete check.
Missing anti-patternSkill shows correct behavior, never wrong. Agent violates without recognizing it.Add counter-example.
Rationalization loopholeAgent's excuse not covered in prevention table.Add the excuse + counter.
Cognitive-load bypassRule buried in prose. Agent under pressure skims past it.Move to scannable position (red flag, bold, gate).
Contradictory guidanceTwo rules conflict. Agent picks the easier one.Resolve contradiction.
Context-window attenuationSkill loaded early, details lost by action time.Rarely a skill fix. Move critical rules earlier or add gate.

Tolerability

Context window fullness at failure time is the primary toleration factor.

Context %Minor severitySevere severity
< 50%Fix — agent had roomFix — no excuse
50–80%Judgment callFix — severity demands it
> 80%Tolerate — context explains itFix anyway

Severe = security, data loss, irreversible, user-visible breakage. Minor = style, verbosity, suboptimal structure.

Enhancement Format

One edit, not a rewrite:

Gap: [type]
Skill says: "[quote]"
Agent did: "[quote]"
Proposed fix: [one edit — gate, counter-example, rationalization row, or reorder]

Example: Checklist Body (skill commit)

Violation: Agent wrote bullet-point checklist as commit body: "Merged: - A + B - C..."

Skill said: "Body tells a story, not a checklist."

Gap: Passive language. No enforcement — just "read aloud."

Context: ~60%. Fix needed.

Fix: Added concrete body gate rejecting bullets/lists/headers. Added to Rationalizations and Common Mistakes.

Common Mistakes

MistakeFix
Rewriting the skill instead of a targeted editOne edit, one gap.
Blaming the agentSkills prevent failures. If it didn't, the skill has a gap.
Fixing context-window failures in the skillLong context is environmental. Only fix if severe.
Adding a rule without a gateRules without enforcement are passive.

Validation

After proposing an enhancement:

  1. Would this have prevented the specific violation?
  2. Does the fix address the gap type, or just restate the rule?
  3. Did you avoid rewriting unrelated sections?
  4. If context >80%, is this a skill fix or a loading/positioning fix?
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall