swing-skills

Swing — AI Agent Guardrails for Claude Code

English · 한국어 (Korean)

Your AI is not swinging at all.

It picks the safe default, agrees when it should push back, cites no sources, hides its reasoning, and assumes everything will work. Six cognitive failures. Six firewalls. Make it swing.

npx skills add thestack-ai/swing-skills --all

Built for Claude Code | Compatible with AI agents via Agent Skills open standard Created by @thestack_ai

---

See the Difference

Before: Default AI Research

You: Is SQLite viable for an app with 1000 concurrent users?

AI: No, SQLite is not suitable for high-concurrency applications. Use PostgreSQL or MySQL instead for production workloads.

After: With swing-research

You: /swing-research Is SQLite viable for an app with 1000 concurrent users?

AI: Yes, with caveats. 1000 concurrent users ≠ 1000 concurrent writes. Typical web apps generate ~30 concurrent write transactions at this scale. SQLite handles ~120 write transactions/sec in WAL mode (verified: sqlite.org benchmarks + Expensify production case study serving 10M+ users on SQLite). The real constraint is write contention above ~100 concurrent writers, not user count.

Sources: 🏛️ sqlite.org WAL documentation, 🛡️ Expensify engineering blog (2024), ⚠️ HN discussion (corroborated)

The baseline gave a wrong answer confidently. The skill forced source verification and reached the opposite — correct — conclusion.

Before: Default AI Code Review

You: Review this JWT auth module

AI: The code looks well-structured. Consider adding more comments and using TypeScript for better type safety. Variable naming could be improved in some places.

After: With swing-review

You: /swing-review Review this JWT auth module

AI: Steel-man: JWT-based stateless auth was chosen for horizontal scaling. The 15-min access + 7-day refresh split balances security and UX.

🔴 Critical: No refresh token rotation. A stolen token grants persistent 7-day access with no detection. Fix: implement token family chain rotation.

🟠 Major: Rate limiting uses in-memory Map(). Resets on restart, bypassed in multi-instance deploys. Fix: move to Redis with TTL.

Verdict: PASS WITH CONDITIONS — must fix token rotation before deploy.

The baseline found nothing. The skill found a Critical security vulnerability.

---

Six Failures, Six Firewalls

Cognitive Failure	What Happens	Firewall	What It Forces
Premature closure	AI rushes to execute on ambiguous requests	swing-clarify	Structured clarification before execution begins
Hallucination	AI states claims without verification	swing-research	Source-traced, cross-verified claims with S/A/B/C tier grading
Anchoring bias	AI locks onto the first "obvious" answer	swing-options	5 probability-weighted options including unconventional alternatives
Confirmation bias	AI agrees with you instead of challenging	swing-review	Steel-man then 3-vector attack. "Looks good" is structurally banned
Black-box reasoning	AI gives conclusions without showing why	swing-trace	Assumption inventory and weakest-link analysis
Optimism bias	AI assumes the plan will work	swing-mortem	Assumes failure, identifies 5 specific failure scenarios with circuit breakers

---

Skills

swing-clarify

Prevents premature execution on ambiguous requests.

Request → 5W1H Decomposition → Ambiguity Score → Clarify or Proceed

• Breaks every request into 6 dimensions: What, Who, Where, When, Why, How
• Scores ambiguity (0-6) — clear requests get a green light, ambiguous ones get up to 3 questions
• Every question includes multiple choice options and a stated default assumption
• Designed to run FIRST — before any other skill in the chain

Try: /swing-clarify Build me an auth system

swing-research

4-stage verified research pipeline with anti-hallucination safeguards.

Deconstruct → Search & Collect → Cross-Verify → Synthesize

• Every claim must be traced to a specific, citable source — or labeled Unverified
• Source tiering: S (academic/specs), A (official docs), B (community — flagged), C (general)
• Cross-verification: key claims require 2+ independent sources (same-origin rewrites don't count)
• Adaptive depth: narrow questions get 2-3 queries, broad landscape gets 8+

Try: /swing-research Is gRPC better than REST for mobile backends?

swing-options

Probability-weighted option generator that breaks anchoring bias.

• 5 options by default, each assigned a typicality zone (Conventional → Wild card)
• At least 1 option from the unconventional or wild card zone — ideas your AI normally suppresses
• Hidden Assumptions section exposes why the obvious answer seems obvious
• Decision matrix with criteria derived from your stated constraints + 1 hidden criterion you didn't consider

Try: /swing-options Which database for a real-time leaderboard?

swing-review

Structured Devil's Advocate that finds real problems, not nitpicks.

• Steel-man first: articulates the strongest case FOR the current approach before attacking
• 3 independent vectors: Logical Soundness / Edge Case Assault / Structural Integrity
• Severity classification: Critical (must fix) / Major (should fix) / Minor / Note
• Every Critical and Major finding includes a counter-proposal with trade-off analysis
• Explicit verdict thresholds: any unmitigated Critical = FAIL

Try: /swing-review We chose Kubernetes for our 3-person startup

swing-trace

Makes AI reasoning visible, auditable, and decomposable.

• Assumption inventory: every assumption numbered, rated for criticality and verifiability
• **Decision