Evidence-driven skills for AI coding agents. Each skill supports a specific engineering decision while preserving the user's task, project choices and production-action boundaries. They are independent methods, not a mandatory sequence of design, implementation and operational proof.
List the available skills:
npx skills add Dankosik/agent-skills --listInstall a specific skill:
npx skills add Dankosik/agent-skills --skill postgres-performanceUse the installer's supported destination for your coding agent. This repository contains instructions and references, not credentials or an agent runtime. Review updates before adopting them; record the chosen upstream commit when reproducibility matters. No release/version manifest is introduced by this update.
| Skill | Use it for |
|---|---|
| auth-access-control | Authentication, permission enforcement, tenant scope and revocation |
| cache-engineering | Cache value, key, freshness, fill, invalidation and degraded behavior |
| concurrency-control | Interleavings and arbitration on shared durable state, not runtime memory races |
| distributed-system-design | Composition of component, capacity, state, failure and evolution contracts |
| durable-background-jobs | Durable acceptance, attempts, leases, schedules, effects and recovery |
| external-api-integration | External HTTP/webhook identity, attempts, ambiguous outcomes and reconciliation |
| postgres-performance | PostgreSQL bottleneck attribution and evidence-backed workload improvements |
| postgres-schema-design | Relational identity, invariants, constraints and migration decisions |
| production-diagnosis | Unknown incident/regression localization and causal or recovery assessment |
| reliable-messaging | Broker publish/consume, acknowledgement, ordering and recovery boundaries |
Select by the unresolved decision, not by every technology mentioned in the repository. A job executed from a broker may need both job and messaging guidance, but loading one does not require loading the other. Adjacent skill names are navigation, not automatic delegation or dependencies on separately installed skills. Use only the references for the selected mechanism and reuse complete applicable material already available in context.
Diagnosis and review can finish with a supported explanation or a precise gap without changing code. Design delivers the requested contract and validation plan. Implementation continues through applicable local checks and repair of introduced failures. An explicitly requested runnable test needs source, setup and assertions.
A narrow fix does not reopen an accepted architecture, create a new infrastructure platform or require every test listed in a reference. A full readiness, migration or recovery claim still needs every relevant in-scope guarantee covered. A missing live environment limits the claim; it does not turn mocks into production evidence or block unrelated local work.
Production actions retain exact authorization, preflight and fresh readback. Read-only probes also need appropriate cost and data bounds. Keep observations, inference, proposals, local checks and verified-live results distinct. Comparable measurements are required for an optimization claim, not invented for every fix.
skills/<skill-name>/
├── SKILL.md # decision, scope, invariants and completion
├── references/ # optional mechanism-specific detail, where needed
└── evals/evals.json # existing evaluator material, not agent instructions
evals/instruction-boundaries/ # new neutral prompts and separate grading criteria
docs/instruction-audit.md # audit rationale and limitations
tests/test_skill_resources.py # structural checks only
Do not load evaluation expectations as runtime guidance. The existing per-skill evaluation paths remain unchanged for compatibility; their placement does not make them required reading. New evaluation materials remain outside skills.
Run the standard-library-only resource checks from the repository root:
python3 -m unittest discover -s tests -vThese check inventory, metadata, local links/anchors and evaluator integrity; they do not run PostgreSQL, brokers, production probes or model comparisons. The matching CI job performs the same checks without a packaging/release pipeline. See the audit and evaluation protocol.
The instruction update is informed by OpenAI's skills and prompts article. Its model-specific observations do not establish identical behavior in every host. The earlier collection also drew on the latest-model guide, Matt Pocock's skill glossary and the context-engineering article supplied by the author. Those historical links are attribution, not a claim of fresh verification here.