Artifact Creator

Create the smallest complete artifact that satisfies the user's requested runtime and security boundary. Keep portable instructions separate from host packaging.

Overview

There is one portable Agent Skills contract, one open MCP protocol, and many different plugin, agent, hook, permission, and marketplace formats. This skill selects the correct authority before writing files and prevents host-specific details from leaking into the portable core.

When to use

Use for a new or substantially revised skill, subagent, plugin, MCP integration, hook, or marketplace entry. For a production modernization involving research, migration, security review, and release evidence, use production-upgrade instead. For review without changes, use artifact-validator.

Prerequisites

• The requested outcome and target repository or destination.
• Existing project instructions and the target host's current specification.
• Authorization for any local writes. Creation does not imply permission to install, publish, deploy, open a PR, or merge.

The pre-authorized Read, Write, and Edit capabilities apply only to local files inside the user-selected target. Every network, shell, installation, and external mutation capability remains behind host and project approval.

Instructions

1. Read project instructions and inspect the target tree before choosing a format. Preserve generated-file ownership and existing public identifiers.
2. Classify the artifact using references/artifact-contracts.md. If the request mixes a portable skill with host wiring, design the portable skill first and the host adapter second.
3. Establish the current authority:

• Agent Skill: the open Agent Skills specification plus any declared local overlay.
• MCP server: the Model Context Protocol specification and selected SDK.
• Plugin, subagent, hook, command, or catalog: the target host's current specification.
• This marketplace: repository instructions and its canonical validators.

1. Derive least privilege from the actual workflow. Do not copy a broad tool list from another artifact. Separate read-only analysis from mutations.
2. Create only useful files. A skill requires SKILL.md; add scripts when deterministic execution matters, references for conditional depth, assets for output material, and evals for behavior that needs proof.
3. For MCP work, define tools around bounded user outcomes. Validate every input, cap output and retries, redact secrets, separate reads from writes, and require explicit confirmation for high-impact mutations.
4. For host adapters, follow references/host-adapters.md. Do not claim …

Artifact Validator

Audit agent-system artifacts without changing them. Select authority by artifact type, run the strongest available deterministic checks, and separate observed facts from unsupported claims.

Overview

This skill is an orchestrator, not another schema authority. In the Tons of Skills repository it delegates to the repository's canonical validators. In other projects it applies the open specification floor and the target host's current contract, clearly labeling checks that could not be reproduced.

When to use

Use for a single file, plugin directory, marketplace, contributor change, or mixed repository. Use artifact-creator when changes are requested. Use production-upgrade when the audit must lead through research, implementation, migration, and release readiness.

Prerequisites

• A target path or repository.
• Read access to project instructions and generated-file ownership rules.
• Network access only when current primary-source verification is necessary. Never transmit private artifacts to an external validator without approval.

Instructions

1. Read project instructions, inspect repository status, and determine whether the target is generated, mirrored, canonical, or host-specific.
2. Run the offline inventory helper:

   python3 scripts/inventory_artifacts.py <target>

1. Use references/validation-routing.md to select the owner for each discovered artifact. Do not merge verdicts from different standards into one unexplained score.
2. In the Tons of Skills repository:

• use scripts/validate-skills-schema.py for skill and agent authority;
• use scripts/validate-mcp-config.mjs and the declared security gates for MCP configuration;
• run catalog synchronization before judging generated projections;
• keep kernel-shadow results advisory unless governance explicitly changes.

1. Validate structure, references, declared versus used capabilities, secret hygiene, path containment, bounded execution, provenance, license, and the evidence supporting every portability or production claim.
2. For behavior, require positive, negative, edge, and adversarial cases. A passing structural score does not prove runtime behavior or safety.
3. Reproduce automated-review findings independently. Reject or qualify any claim that lacks a command, primary source, fixture, or observable boundary.
4. Return a severity-ordered report with exact paths, evidence, authority, remediation options, and a PASS, FAIL, or NOT-VERIFIED verdict per artifact.

Output

Report the inventory, checks executed, exact command results, blocking findings, warnings, unverified surfaces, and overall c…

Production Upgrade

Run a risk-adjusted, evidence-bound modernization from discovery through a pre-release maintainer checkpoint. The workflow is complete without a specific model, subagent implementation, or task tracker, but it must honor stronger project requirements when they exist.

Overview

The quality bar comes from the Databricks rebuild: understand real operator pain, decide architecture before implementation, move load-bearing logic into deterministic code, preserve compatibility intentionally, and prove safety with negative and adversarial evidence. Match the depth to risk rather than matching another project's document count.

When to use

Use for a legacy artifact, broad rewrite, breaking migration, unsafe integration, or production-readiness claim. Do not trigger for a small typo, isolated bug fix, routine dependency bump, or read-only status question unless the user explicitly requests the full upgrade workflow.

Prerequisites

• Target repository or artifact and its project instructions.
• Authority to research and prepare local changes. Publication, deployment, merge, destructive cleanup, and external messaging remain separate approvals.
• Current primary sources for externally defined contracts.

The pre-authorized Read, Write, and Edit capabilities apply only to local, scoped implementation files. Network, shell, task-tracker, subagent, install, and publication capabilities remain behind host and project approval.

Orchestration

Use five focused roles. When the host supports isolated subagents, dispatch the role packets under references/roles and have the coordinator reconcile their evidence. Otherwise execute the same packets sequentially in the main context. A same-identity or inline review is self-review, never independent.

Instructions

1. Recover context and establish authority

1. Read repository instructions, architecture owners, generated-file rules, current status, worktrees, active reviews, and existing task state.
2. If the repository uses Beads, follow references/beads-workflow.md: prime, search, create or reuse, claim…