Skill Packs
Curated groups of skills, tools, hooks, prompt surfaces, and scaffold assets
A pack turns repository evidence into a curated operating bundle.
The repo receives the right context without forcing every operator to assemble an AI setup by hand.
Pack Resolution
hi scaffold detects repo facts, maps those facts to packs, asks for confirmation, then writes the scaffold output. The CLI selects packs, not raw skills.
For this domain model, the important repo facts are simple: dependencies and file extensions. Dependencies show which frameworks, libraries, and toolchains are present. File extensions show which languages and source surfaces exist.
The CLI mechanics belong in the CLI domain. The Harness concept is that pack selection follows repo evidence.
Pack Types
The high-level model has two pack types:
| Pack Type | Selected By | Adds |
|---|---|---|
| Agnostic | Every harnessed repository. | Requirements, planning, review, docs, debug, quality, research, subagent workflow, and delivery workflow context. |
| Language | File extensions and language toolchain evidence. | Language-specific skills, prompts, examples, and validation expectations. |
Specific surfaces get their own pack pages only when they need durable conceptual guidance. Resolver details belong in the CLI domain.
The mandatory agnostic pack groups include debug, docs, misc, planning, quality, research, requirements, and subagents. They carry phase wrappers such as delivery-phase, review-phase, debugging-phase, docs-ingest-phase, and requirements-phase, plus the human-only business-finder and functional-finder profiles over one finder-phase engine. Finder is optional intake; Requirements Phase is the sole route through Requirements Grill, Create Spec, and delivery-depth Write Backlog projection. Research also carries prototype for throwaway logic/UI prototypes, review for Standards versus Spec review framing, and autoreview as the structured review runtime. The docs pack distributes writing-fragments, writing-beats, writing-shape, and writing-for-agents; the latter replaces writing-great-skills for skills, prompts, handoffs, and agent instructions. The misc pack installs show-me and wait-what into shared .agents scope without propagating them to application workspaces. show-me adds compact visual explanations where the owning workflow requires them; handoff does not use it. Phase wrappers stay global orchestration entrypoints rather than scoped primary skills. Finder entrypoints additionally require explicit human invocation.
The default security pack includes architect-pipeline for provider-neutral CI/CD and infrastructure-deployment orchestration. It covers release identity, artifact creation, publication, retry, and promotion theory, while audit-cicd-security remains the route for explicit pipeline security audits and security-best-practices remains the route for application security. Pulumi-specific APIs and resource, state, registry, authentication, ESC, and Automation API details remain owned by the Pulumi pack; architect-pipeline defines only the orchestration boundary and deployment gates.
The quality pack separates production change from test-portfolio work. tdd drives production behavior from an observed RED test to minimal GREEN implementation. make-tsuite audits automated software test signal, overlap, ownership, placement, and execution authority, then applies test-only changes when the operator authorizes them. It keeps production behavior immutable and discovers the repository's own language, framework, execution mechanism, tiers, and fixtures, including required delivery gates when present. TSuite means test suite; the skill is language-, framework-, and toolchain-agnostic.
The requirements pack includes the imported domain-modeling skill beside requirements-grill. During a grill, the active glossary is its working persistence surface. Accepted closure uses requirements-grill's wiki synthesis to promote accepted terminology into the canonical routed per-context glossary. Downstream specification, planning, implementation, backlog, backend/frontend structure, architecture, and simplified-explanation work reads that glossary. Terminology changes return to requirements-grill; downstream consumers do not rewrite glossary authority. The imported skill remains the sole authority for domain-modeling behavior; Harness owns invocation and persistence routing.
wait-what runs only when the user invokes it after an explanation does not land. It asks for restored context, ASD-STE100 Simplified Technical English, and stable repository terms from GLOSSARY.md. It does not rewrite every response automatically.
write-backlog is an explicit lifecycle exception and the sole provider writer. It reads wiki authority and fresh provider state, reconciles stable provider and durable wiki identities, then previews material Product Area, Initiative, Epic, Story, Task, V*, blocker, metadata, or view changes before writing. Existing coherent structure is reused first. Provider-native relationships, exact readback, and textual ticket fields remain authoritative.
Pack Output
A selected pack can contribute several kinds of scaffold output:
| Output | Purpose |
|---|---|
| Skills | Install the execution contracts the repo should have available under .agents/skills. |
| Prompt guidance | Shape generated root and scoped prompt surfaces without turning packs into local product truth. |
| Subagent templates | Create scoped worker contracts that match the repo surfaces detected by the pack. |
| Hooks and tools | Attach required guardrails and tool expectations. Validation Gates, External Tool Contract, and Hooks explain those mechanics. |
| Source examples | Point agents toward useful upstream or package source when the surface benefits from comparison. |
Packs distribute operating context. Product requirements, business logic, and repo-specific source truth remain in the repository that owns them.
For a practical lookup of available skills and source SKILL.md files, use Skill Reference. This page stays focused on the pack model; the reference page lists individual skills.
For individual skills that deserve more operator judgment than a reference row can carry, use Atomic Skills.