5 cognitive firewall skills for AI agents, packaged as a Lore bundle.
AI agents are confident by default. They anchor on the first solution they find, hallucinate sources, skip failure analysis, and produce reasoning you cannot audit. Stack Skills provides five structured skills that counteract specific cognitive failure modes -- each one a deliberate thinking protocol, not a vague instruction to "be careful."
When you need to stress-test a decision or PR before committing to it, adversarial-review runs a three-vector attack. It steel-mans the current approach first (why was this chosen? what does it optimize for?), then systematically attacks logical soundness, edge cases, and structural integrity. Every finding gets a severity classification and a concrete counter-proposal with trade-offs. It is not a code review -- it is a deliberate attempt to break your reasoning.
When you are stuck on the obvious answer, creativity-sampler generates five alternatives across a typicality spectrum from conventional to wild card, using probability weighting to break anchoring bias. It exposes hidden assumptions that constrain your option space.
When accuracy matters more than speed, cross-verified-research enforces source tracing with anti-hallucination safeguards. Every claim must be cross-verified by two or more independent sources, graded on a confidence tier system (S/A/B/C). The agent cannot assert facts without citing where they came from.
Before committing to a plan or irreversible decision, pre-mortem assumes complete failure six months out and works backward through five categories (technical, organizational, external, temporal, assumption) to surface risks. Each scenario must name specific technologies and quantities -- "the database might not scale" is banned. The output includes measurable leading indicators and circuit breakers with concrete trigger conditions.
When you need to see why a conclusion was reached, reasoning-tracer forces the agent to make its reasoning chain visible: assumption inventories, decision branching, confidence decomposition, and weakest-link identification. It turns opaque "here's what I think" into auditable reasoning you can challenge.
The skills are designed to chain: run creativity-sampler to generate options, cross-verified-research to validate them, then adversarial-review to stress-test the winner. Or run pre-mortem on a plan, then creativity-sampler to generate mitigations for the top risks.
Originally created by @thestack_ai as whynowlab/stack-skills. Licensed under MIT.
| Skill | Purpose |
|---|---|
| adversarial-review | Devil's Advocate stress-testing with 3-vector attack |
| creativity-sampler | Probability-weighted option generation that breaks anchoring bias |
| cross-verified-research | Source-traced research with anti-hallucination safeguards |
| pre-mortem | Prospective failure analysis using Gary Klein's technique |
| reasoning-tracer | Makes reasoning chains visible and auditable |
- Converted from Claude Code skill format to Lore bundle format
- Skill directories moved from
skills/(lowercase) toSKILLS/(uppercase per Lore convention) - Frontmatter: removed embedded trigger keyword lists from descriptions; added
user-invocable: true(Lore standard);argument-hintandallowed-toolspreserved as pass-through fields - Removed references to skills not included in this bundle (deep-dive-analyzer, skill-composer, orchestrator strategy team)
- Skill body content preserved as-is -- no rewrites
- Added
manifest.jsonandLORE.mdper Lore bundle standard
MIT License. See the original repo for full license text.
