awino

Use when asked to create a new agent, subagent, orchestrator, or specialist. Searches the existing skill library first, writes a spec for approval, then emits a validated agent definition to staging.

Decide whether a capability should be a skill, a hook, a script, a recipe, or an MCP server, then build it. Use when asked to create a tool or extend what an agent can do

Use on A.W.I.N.O. first run to create the .smith folder scaffold and verify the knowledge registry against upstream. Triggers on "bootstrap A.W.I.N.O.", "set up A.W.I.N.O.", missing .smith folder.

Audit project configuration such as pyproject.toml, uv.lock, Makefile/Justfile, CI workflows, harness config, .env files, and kilo.json permissions for drift, conflicts, and unsafe defaults. Use for config, setup, CI, or permission audit requests. Read-only, never mutates files.

Use when asked a conceptual agentic-engineering question such as what is a harness, how should context be managed, or which pattern fits. Fetches at most three chapters from the book registry and answers with citations.

Debug concrete bugs, errors, exceptions, ValueError failures, pytest failures, and failing tests through reproduce, diagnose, fix, and verify phases with evidence-backed fix authorization.

Decompose work into parallel subagents with disjoint file ownership. Use when independent workstreams can run simultaneously and the orchestrator should coordinate rather than implement

Mission-first discovery for a raw idea or sparse repository. Reflects the project purpose, asks one frontier question at a time, captures goals tenets expectations non-goals and success criteria, and refuses to rush into a spec before the user confirms intent.

Evidence sufficiency and citation gate for research RAG QA and scientific workflows. Decides answer retrieve-more clarify no-answer or block-unsupported-synthesis before factual output is released.

Dual-write memory discipline for A.W.I.N.O. - stores durable rules in the Memory MCP extension for recall and mirrors every entry to an append-only file ledger for audit

Iterative work-review loop with fresh context per iteration and independent verification. Use when a task has a machine-checkable completion gate and may need several attempts to get right

Reproducibility discipline for scientific research pipelines. Requires traceable run artifacts with inputs configuration model identity event history and rerun instructions.

Research then Plan then Implement for complex multi-file changes. Use for refactors, migrations, feature additions, large upgrades, and documentation overhauls where jumping to code would drift

Use when asked to refresh A.W.I.N.O.'s knowledge base or curate the chapter registry against upstream. For updating A.W.I.N.O.'s own installed version (Claude plugin or standalone clone), run 'awino update' directly instead - that is a single command, not a skill with judgment calls.

Use when an agent is misbehaving and the complaint is vague. Converts "the agent is bad" into a named failure mode, a responsible surface, and a structural fix rather than a prompt patch.

Create concrete visual explanations from architecture, processes, timelines, comparisons, and quantitative data. Use whenever the user asks for a diagram, chart, visualization, image, schematic, map, dashboard, workflow, sequence, or visual comparison. Prefer Mermaid in chat, then SVG or self-contained HTML when richer output materially improves understanding.