CoFounder Skill Library

NEW

16 curated skills, 3 sub-agents, and 4 slash commands for AI coding agents working in real codebases — spec-driven, guard-enforced, compliance-aware, and cost-routed.

@waymakerai/aicofounder-skillsClaude Code · MCP · CrewAI · LangChainMIT

Why a curated skill library?

Generic skill collections optimize for breadth. CoFounder skills optimize for depth × leverage: each one stands on the existing CoFounder primitives — VibeSpec context, guard policies, compliance presets, sandbox preview, cost router — so a skill is meaningfully better here than the same idea would be elsewhere.

Skills read your VibeSpec at runtime — behavior adapts per project, not per author.
Sensitivity declarations route every skill through the right policies before it runs.
Per-skill model routing keeps Haiku-class tasks on Haiku and Opus-class tasks on Opus.
Destructive skills preview through the sandbox before touching the real workspace.
Telemetry feeds back: which skills work, which are expensive, which keep failing.

Install

npm install @waymakerai/aicofounder-skills

For Claude Code, the skills ship with the cofounder plugin and are auto-registered. For other MCP-capable agents (Cursor, Cline, Windsurf, Claude Desktop), point your MCP client at the CoFounder MCP server — every skill becomes a cofounder.skill.* tool.

Skills

16 skills across six categories. Each declares its sensitivity, model class, and chain position.

Feature flow

(4)

cofounder-feature-newmid

Turns a feature idea into a VibeSpec-compliant spec via clarifying questions and prior-art search. Writes specs/<slug>.spec.yml. Does not write code.

nextSkill: cofounder-feature-implement

cofounder-feature-implementheavy

Reads a feature spec and produces a scoped, branch-isolated draft implementation under VibeSpec rules. Adds tests and runs guardrail checks before handing off.

writes coderuns shell

requires: cofounder-feature-new

nextSkill: cofounder-check

cofounder-checkmid

Runs CoFounder safety gates on the current changeset — secrets, mock data, scope, design system, missing tests. Reports Tier 1 / Tier 2 / Tier 3 findings.

runs shell

cofounder-spec-reviewheavy

Reviews a draft feature spec for completeness, testability, and VibeSpec compliance before implementation. Does not modify the spec.

Codebase intelligence

(1)

cofounder-pattern-scoutlight

Searches for similar functions, components, hooks, models, and routes before agents reinvent them. Returns ranked matches with extend-vs-create recommendations.

nextSkill: cofounder-feature-new

Documentation

(2)

cofounder-diagrammid

Generates Mermaid (or PlantUML) diagrams from real code — actual imports, actual call graph, actual schema. Never invents relationships.

cofounder-readmemid

Generates a README from real exports, real scripts, real config. Every code example must be runnable against the current repo.

Version control

(2)

cofounder-pr-summarymid

Generates a PR description from the actual diff — mental-model framing, blast radius, test coverage. No marketing language, no aspirational "future X".

runs shell

cofounder-commitlight

Writes a Conventional Commits-style message from the staged diff. Detects multi-purpose diffs and suggests splitting. Will never commit without explicit confirmation.

runs shell

Testing

(1)

cofounder-test-planmid

Reads a spec or diff and proposes a tiered test plan — what to assert, at what layer, and what is not worth testing.

nextSkill: cofounder-feature-implement

CoFounder-native

(6)

cofounder-vibespec-authorheavy

Reads the repo to infer existing conventions (design system, file layout, branching style, test framework) and proposes a VibeSpec that codifies them with evidence.

nextSkill: cofounder-feature-new

cofounder-compliance-frameheavy

Reframes a feature spec against HIPAA, GDPR/CCPA, SEC, PCI DSS, FERPA, or SOX requirements before code is written. Wraps the policies and compliance packages.

may touch PII

nextSkill: cofounder-feature-new

cofounder-cost-routelight

Recommends the cheapest model that is still good enough for the task. Also analyzes recent LLM spend and surfaces optimization candidates with dollar estimates.

cofounder-context-fitmid

Distills long inputs (huge files, long histories, large RAG returns) into the minimum context that preserves the answer. Wraps the context-optimizer package.

cofounder-sandbox-previewmid

Runs destructive operations (migrations, mass renames, third-party API calls with side effects) in the sandbox package and reports what would happen — without touching reality.

runs shell

cofounder-second-opinionheavy

Routes the same prompt through a different model or provider and diffs the answers. Surfaces substance disagreements (different recommendations), not surface differences (word choice).

network

Sub-agents

Spawn-on-demand variants for tasks that benefit from a fresh, isolated context window.

cofounder-reviewer

Reviews a diff or PR against the repo VibeSpecs. Reports only high-confidence, high-priority issues across three tiers.

cofounder-scout

Spawnable sub-agent variant of pattern-scout. Use when you want a separate isolated context for a deep prior-art search before writing new code.

cofounder-spec-writer

Asks 3–5 clarifying questions, searches for prior art, and writes a VibeSpec-compliant feature spec. Confirms before writing.

Slash commands

Quick triggers for the most common flows. Each delegates to the corresponding skill or agent.

/cofounder

Root command — routes to feature-new, feature-implement, check, or review.

/spec

Turn a feature idea into a VibeSpec-compliant spec file.

/check

Run CoFounder safety gates on the current changeset.

/explain-changes

Mental-model walk-through of the current branch — not a file list.

Platform moves

Seven architectural features that make every skill stronger — not just more skills.

FlowSpec skill chaining

Skills declare nextSkill (suggested follow-up) and requires (hard prerequisites). The runtime computes a topologically-sorted chain, detects cycles, and surfaces missing dependencies before execution.

Golden-fixture replay

Each skill ships fixtures/*.json with input + expected behavior. The CI scanner replays them on every PR so skill regressions get caught — substring assertions, refused-file lists, latency budgets.

MCP re-exposure

Every skill is auto-published as an MCP tool (cofounder.skill.*) so non-Claude-Code agents (Cursor, Cline, Windsurf, Claude Desktop) can invoke the same skills through the standard protocol.

Skill telemetry

Skills emit structured events (skill.run, skill.complete, skill.fail, plus skill-specific events). Default JSONL transport writes locally; the dashboard package consumes them to show what works, what is expensive, and what keeps failing.

Sensitivity declarations + guard mapping

Every skill declares its data class (mayTouchPII, writesCode, runsShell, network). The runtime maps sensitivity to policy presets from the policies package and runs preflight checks before invocation.

Per-skill model routing

Skills declare modelClass: light | mid | heavy. The router picks Haiku, Sonnet, or Opus, bumps up for large context, and refuses to downgrade quality-critical skills (spec review, compliance, second-opinion) for cost reasons.

Sandbox preview for destructive skills

Skills tagged sensitivity.runsShell or writesCode are eligible for sandbox preview. Migrations, mass renames, and side-effecting API calls run in the sandbox package first — three outcomes: safe-to-run, needs-review, blocked.

Skill schema

Skills are markdown files with YAML frontmatter. The CoFounder extensions are additive — a skill without them works as a plain Anthropic Agent Skill.

---
name: cofounder-feature-implement
description: Use when the user points to a CoFounder feature spec...
sensitivity:
  mayTouchPII: false
  writesCode: true
  runsShell: true
modelClass: heavy           # light | mid | heavy
nextSkill: cofounder-check
requires: [cofounder-feature-new]
emits: [skill.run, branch.created, files.written]
---

# Skill body in markdown — instructions, process, output format.

Required fields

name — slug used to invoke
description — when-to-use hint

CoFounder extensions

sensitivity — drives policies & sandbox
modelClass — drives the router
nextSkill · requires — drive chaining
emits — drives telemetry

Packages Agents