Decision 017: Skill Enhancements — Failure Modes, Gates, MVE, I/O Contracts¶
Date: 2026-03-28 Status: Accepted
Context¶
Skills tell agents what to do but not when they're doing it wrong or what to cut under pressure. Observed failure patterns:
- Goal: Make skills self-correcting (agents recognize their own mistakes) and gracefully degradable (agents know which steps are load-bearing).
- Constraints: Must not break existing skill contract; must keep SKILL.md under ~500 lines; additive only.
- Anti-goals: Not adding new skills. Not restructuring existing workflows.
- Boundary + time horizon: All 17 skills, near-term (single batch of changes).
- Actors + incentives: AI agents following skill workflows — incentivized to complete quickly, which causes step-skipping.
Options considered¶
| Option | Optimizes for | Knowingly worsens | Reversibility |
|---|---|---|---|
| A: Add inline failure modes + gates + MVE + I/O to each SKILL.md | Agent self-correction, composability, context efficiency | Slightly longer SKILL.md files (~24 lines each) | Fully reversible (remove sections) |
| B: Add failure modes to shared reference file | DRY, single update point | Loses per-skill specificity; agents must load extra file | Fully reversible |
| C: No change (rely on guardrails) | Simplicity | Agents continue failing in predictable, undocumented ways | N/A |
Decision¶
Option A: Add four enhancements directly to each SKILL.md as SHOULD-have sections in the skill contract.
- Inputs / Outputs — what a skill requires from upstream and produces for downstream (placed after Chooser).
- Common failure modes — 3-5 bullets documenting how agents actually fail, serving as anti-anchors (placed after Guardrails).
- Minimum viable execution — 3-4 load-bearing steps vs nice-to-have (placed after Workflow).
- GATE annotations — inline blockquote annotations at critical workflow steps where agents commonly skip or execute superficially (inline within Workflow, not a separate section).
Additionally, add proportionality boundary examples to the workflow skill's Step 0 to help agents classify ambiguous changes.
Why not Option B: failure modes are only useful when they're specific to the skill being executed. A shared file loses that specificity and adds a file-load step agents might skip.
Kill criteria / reversal trigger¶
- If agents consistently ignore the new sections (no measurable improvement in step completion), remove them to reduce context load.
- If SKILL.md files exceed ~500 lines after additions, move content to references/.
Measurement + review ritual¶
- Leading indicators (early): Agents cite failure modes when self-correcting; gate conditions appear in agent reasoning.
- Lagging outcomes: Fewer instances of known failure patterns (e.g., forecast default-association trap, review self-agreement).
- Instrumentation source: Manual review of agent transcripts.
- Owner + cadence + action trigger: Repo maintainer, quarterly, triggered if failure patterns persist unchanged.
Consequences¶
- Positive: Skills become self-correcting; composition is explicit; agents can degrade gracefully under context pressure.
- Trade-offs: Each SKILL.md grows ~24 lines; maintenance burden increases slightly (failure modes must stay current).
- Compatibility: Fully backward-compatible. New sections are SHOULD-have, not MUST-have. Validation scripts unchanged.
Review date¶
2026-06-28