Skills are collected intelligence on how to operate a specific tool — whether that is a CLI, an API, or an MCP/ACP server. They codify procedures, protocols, and output formats that personas reference during execution.

• agent-decision/SKILL.md — structured ambiguity escalation with 1-3-1 analysis and FRAME self-review rubric
• agent-memory/SKILL.md — long-term and session memory across sessions
• architect-design-tree/SKILL.md — builds the grill's design tree: decisions, dependencies, recommendations, impact, facts
• architect-impl-grounding/SKILL.md — grounds the grill's artifacts: annotates impl.md per epic, re-grounds after each landing, classifies discoveries
• boot/SKILL.md — session startup — gitignore, auto-update, memory, rules, orient
• browser-inspect/SKILL.md — browser inspection for UI verification (coder)
• code-coherence-review/SKILL.md — logic coherence, correctness, and structural integrity checks
• code-quality-review/SKILL.md — rules-walk procedure for coding standards compliance
• code-sec-review/SKILL.md — static security review — attack surface, CWE/OWASP 2025, MITRE ATT&CK
• coder-self-review/SKILL.md — GRASP self-review rubric — implementation quality gate
• context-maintenance/SKILL.md — how to maintain .context.md files and docs/FEATURE-MAP.md as the project evolves
• contextualizer-self-review/SKILL.md — TRACE self-review rubric — context generation quality gate
• dispatch/SKILL.md — assembles sub-agent prompts with task brief and routes to the correct provider
• grill/SKILL.md — protocol that interviews the user in rounds over the Architect's design tree; three paths control depth
• plan-management/SKILL.md — plan lifecycle: grill entry, grounding, artifact review, per-epic execution, revision
• loop-recovery/SKILL.md — structured recovery and escalation for retry loops
• reviewer-architect-adversarial/SKILL.md — adversarial plan validation on grill artifacts — structural checks and assumption attack before implementation
• reviewer-handoff/SKILL.md — structured review summary format with verdict logic
• review-loop/SKILL.md — two-mode review loop — single reviewer per epic, three reviewers for full branch
• reviewer-self-review/SKILL.md — SHIELD self-review rubric — unified reviewer quality gate
• task-tracking/SKILL.md — file-based to-do tracking for multi-step and multi-session work
• web-search/SKILL.md — web search and page fetch through the TinyFish CLI (all)

Extract a skill when:

• A tool proves difficult enough that a human must step in and write an explicit how-to for the agent to follow.
• A procedure must be standardized across multiple personas, such as a shared protocol or output format.

Do not extract when the procedure is short and intuitive. If a competent agent can work it out without written guidance, a skill file adds overhead without value.

One directory per skill. The directory holds a single SKILL.md. The directory name is lowercase, hyphenated, and must match the name frontmatter field: agent-memory/SKILL.md, dispatch/SKILL.md.

This layout matches the host runtime's skill discovery contract: the runtime loads each skill from its own directory — one SKILL.md per skill — and indexes it by its name.

Shared scripts that support skills live in assets/. They are not skills.

• name (Required) — Skill identifier; lowercase alphanumeric with single hyphens; must match the directory name. Example: agent-memory
• description (Required) — What the skill does in one sentence; the host runtime lists this in the skill picker. Example: Cross-session memory retrieval and storage
• usedBy (Required) — Which personas use this skill. [all] if injected universally via boot. Example: [all] or [maestro]
• relatedTo (Optional) — External tools, CLIs, or APIs this skill wraps or abstracts. Example: [docker, awk] or [anthropic-api]
• version (Required) — Semantic version. Example: 0.1.0
• lastUpdated (Required) — Last modification date. Example: 2026-02-05

• Purpose (Required) — What this skill does and why it exists. One paragraph, no bullet points. Answer "what problem does this solve?" not "what steps does it take."
• Procedure (Required) — Numbered steps for executing the skill. Each step that produces an artifact must describe its output inline — format, structure, and destination. Reference other skills or rules with (uses: path) or (follows: path) as needed.
• Guardrails (Optional) — Skill-specific pitfalls to avoid. Not rules, not procedure repetition. Ask: "what mistake would an agent make when using this skill carelessly?"

Nothing sits between Purpose and Procedure. Definitions the steps depend on (modes, scopes, path rules) fold into Procedure as a preamble. All other sections are reference material and sit after Procedure.