Skip to content
Back to projects
In progress

Cherry Agents — audited skills and subagents for Claude Code

182 skills and 59 subagents for Claude Code and OpenCode from a single source, with a security scanner, a signed manifest and behavioural evals.

What it is

A catalogue of 182 skills and 59 subagents for coding agents, in two synchronised distributions: Claude Code and OpenCode. Both are generated from the same source, so the content is identical and only the format each platform expects changes. On top sit 16 themed plugins (SDD, security, testing, DevOps, AI engineering, data, GCP…) generated from the catalogue, so you install only the area you need.

What sets it apart from a collection of prompts

The difference is not size: the content is audited and tested.

  • The skill is the source of truth, the agent a thin wrapper. Each subagent preloads its same-named skill and only adds a role, constraints and an output contract. To improve an agent you edit its skill; methodology is never duplicated.
  • Its own security scanner (skill-vet) with eight categories: prompt injection, exfiltration, download-and-execute, credentials, destructive commands, self-persistence, obfuscation and unverifiable dependencies. It also detects invisible Unicode, the vector for poisoned context.
  • Integrity and authenticity of what gets installed. A per-file SHA-256 manifest, signed with a dedicated key (ssh-keygen -Y). The installer aborts if the signature does not match, and the gate fails if the manifest changed without being re-signed.
  • Behavioural evals: real sessions in a sandbox, with deterministic checks on what ends up on disk. They cover the spec-driven development pipeline, skill-vet and two Google Cloud audits; the rest of the catalogue has no eval, and the README says so.
  • The context budget as a design constraint. Claude Code lists skills on every turn under a cap of 1% of the window; the whole catalogue overflows it. That is why descriptions have a length limit, which the gate checks, and why the per-area plugins exist.
  • The gate has tests too: the scripts that validate the catalogue carry their own suite, so a change to them cannot leave it green by accident.

What it covers

The full SDD pipeline (explore, propose, specify, design, tasks, implement with strict TDD, verify, archive), AI engineering (production RAG, LLM evaluation and observability, MCP servers and their security, human-in-the-loop), security, frontend and backend frameworks, data, testing, DevOps and infrastructure, SEO and a virtual executive team.

Where the content comes from

Part of the catalogue is ported from open projects and credited under their licences: 16 skills from addyosmani/agent-skills (MIT) and the executive team from OpenExecutive (Apache-2.0). My work is the catalogue itself and what makes it trustworthy: generating both distributions and the plugins, the scanner, the signed manifest, the evals and the gate.

Stack: Markdown with YAML frontmatter · Python (generators, scanner and tests) · Bash · OpenSSH signing · GitHub Actions.

Private repository · evolving.