Create Target

Create a new target implementation for universal-ai-config. A target maps universal template frontmatter and hooks to a specific AI coding assistant's configuration format.

Phase 1 — Research & Approval

1. Research the target's config format — find up-to-date documentation online for how the target AI assistant expects its configuration files. Look for:

• File locations and naming conventions (e.g., .cursor/rules/*.mdc, .github/instructions/*.md)
• Frontmatter format for each file type
• Hook configuration format (JSON structure, event names)
• MCP tool reference syntax — how the target references MCP tools in hook matchers, agent tools: frontmatter, and any allow/deny lists. Find the exact string format (e.g. mcp__server__tool, server/tool, MCP:tool) and the wildcard convention for "all tools on a server"

Look through docs for all supported configuration types:

• Instructions/rules/"system prompt"
• Skills/Commands
• Agents
• Hooks

1. Present findings for approval — before writing any code, present a structured summary of your research to the user. For each supported configuration type, include:

• What it maps to: the universal template type it corresponds to (instructions, skills, agents, hooks)
• File location & naming: where the target expects these files and what extensions/naming conventions it uses
• Frontmatter/metadata format: the exact fields, keys, and structure the target uses (with examples from docs)
• Hook format (if applicable): event names, JSON structure, how handlers are defined
• Source links: direct URLs to the official documentation pages you referenced

Format as a clear table or grouped list so the user can review each mapping. Ask the user to approve or flag any corrections before proceeding. Do not continue until the user explicitly approves.

Phase 2 — Plan

After the user approves the research findings, enter plan mode to design the full implementation before writing any code. The plan should cover:

1. Review existing targets for patterns — read through at least one existing target implementation (e.g., src/targets/claude/index.ts) and the shared types to understand the exact interfaces and conventions you need to follow.

1. Draft the implementation plan — produce a detailed, step-by-step plan that includes:

• Target definition: the full defineTarget() structure — name, outputDir, supportedTypes, and for each supported type:
• frontmatterMap — every universal key and what it maps to (string rename or function transform), with reasoning
• getOutputPath — the exact path logic, including how alwaysApply or other flags affect routing
• Hooks (if supported): transform function logic mapping universal event names → target events, outputPath, and mergeKey if the hook config nests inside a larger file
• Registration: changes needed in src/targets/index.ts and the Target type in src/types.ts
• Config schema: any changes needed in src/config/schema.ts or src/config/defaults.ts
• Tests: which fixture(s) to use or create, and what assertions to write
• Docs: any updates needed to README, meta-instruction templates, or seed files
• Upstream validation reference: draft the full content of a new 2.X section for [.universal-ai-config/skills/report-feature-changes/upstream-validation-reference.md](../report-feature-changes/upstream-validation-reference.md) (following the Part 4 template in that file) as part of the plan — include it inline so the user can review and correct the URLs and per-feature mappings before anything is written. The draft must cover:
• At least one Tier 1 authoritative reference URL (docs home, full reference, source repo if open).
• At least one Tier 2 changelog / release-notes / blog URL so future drift can be dated.
• Tier 3 community/adjacent resources if any are notable.
• Per-feature index rows for every feature category from Part 1 that the target supports (file locations, frontmatter per template type, hooks, MCP, tool system, permission modes, memory scopes, model identifiers). Leave a row blank with a — note for categories the target doesn't support.
• Any known beta/preview features or documentation quirks worth pinning in Part 5.
• If the target has structural differences from other targets (e.g. a different config format, a many-to-one instruction model, files outside the normal output directory), include an "architectural notes" block before the section (see the Codex section for an example).
• If any features exist that users can hand-author but uac doesn't manage, list them in a "Deferred uac coverage" subsection.

This is mandatory — without it the report-feature-changes skill has no reference URLs and the target will silently drift.

Reference the defineTarget() pattern:

```typescript import { defineTarget } from "../define-target.js"; import type { UniversalFrontmatter, UniversalHookHandler } from "../../types.js";

export default defineTarget({ name: "", outputDir: ".", supportedTypes: ["instructions", "skills", "agents", "hooks"], instructions: { frontmatterMap: { description: "description", globs: "", alwaysApply: "", }, getOutputPath: (name, fm) => rules/${name}., }, // skills, agents — similar pattern // hooks — implement transform function }); ```

Key decisions to address in the plan:

• frontmatterMap: maps universal keys → target keys. Use a string for direct rename, or a function (value, fm) => Record for transforms
• getOutputPath: determines output file path from template name and frontmatter
• consolidate: if the target maps multiple templates of a type into one or more output files (e.g., many instructions → one AGENTS.md), use the consolidate function on TemplateTypeConfig instead of getOutputPath. Consolidate-based types can also emit files at root-relative paths outside outputDir. See the Codex target for a reference implementation.
• Hook transform: converts universal event names and handler format to target format
• Hook mergeKey: if the target's hook config is nested inside a larger settings file (like Claude's settings.json which has a hooks key)
• mcpToolRef update: add a case "" branch to mcpToolRef() in src/core/parser.ts with both the specific-tool form (e.g. server/tool) and the wildcard form (e.g. server/*). Derive these from the MCP tool reference syntax you found in Phase 1.
• New universal fields: audit whether the target needs new fields in UniversalFrontmatter, UniversalMCPServer, or UniversalHookHandler for lossless round-trip fidelity. If so, plan additions to src/types.ts and the meta-instruction uac-template-guide.md field tables. Common triggers: target-specific auth options, timeout knobs, per-server tool allow/denylists, agent sandbox modes.
• Pipeline extensions: if the target needs a non-JSON output format (e.g., TOML for hooks/MCP) or a new writer primitive for merging into a shared user-managed file, plan those explicitly. Examples: format: "toml" on HooksTypeConfig/MCPTypeConfig, a mergeTomlKey writer primitive analogous to the existing mergeJsonKey.

1. Get plan approval — present the plan to the user for review. Do not start implementing until the plan is approved.

Phase 3 — Implementation

Once the plan is approved, execute it:

1. Create the target directory — create src/targets//index.ts

1. Implement defineTarget() — follow the approved plan to build the full target definition

1. Register the target — add to src/targets/index.ts and the Target type in src/types.ts

1. Add mcpToolRef case — add the new case "" branch to mcpToolRef() in src/core/parser.ts. Add corresponding test cases to tests/unit/parser.test.ts (specific-tool and wildcard forms for the new target).

1. Add new universal type fields — if the plan identified new fields for UniversalFrontmatter, UniversalMCPServer, or UniversalHookHandler, add them to src/types.ts now. Also update the uac-template-guide.md field tables to document the new fields with their per-target support annotations.

1. Add tests — create integration tests using existing fixtures that generate for the new target, verifying output paths and content

1. Update the upstream validation reference — copy the draft section from the approved plan into [.universal-ai-config/skills/report-feature-changes/upstream-validation-reference.md](../report-feature-changes/upstream-validation-reference.md), applying any corrections the user made during plan review. After inserting, skim Part 1 of that file and confirm every feature category your target supports has a corresponding row in the per-feature index.

1. Run pnpm check to verify everything passes

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fabis94
• Source: fabis94/universal-ai-config
• License: MIT

Install and usage instructions live in the source repository linked above.