Agent Config

Personal configuration for coding agents: skills, subagents, coding rules, hooks, MCP templates, and settings. Clone it anywhere; scoped link scripts wire durable content into Claude Code and GitHub Copilot without putting any tool's runtime directory under Git.

What this is

This repo is the live agent configuration of a data professional working day-to-day in Microsoft Fabric, Azure, and Power BI. It holds skills, a subagent, path-scoped coding rules, hooks, MCP templates, a pre-commit linter, and fixture tests — the same files my agents load when I open a session. Cherry-pick what's useful; the contents track active Microsoft data-platform work, so expect drift as the platform moves.

The repo started life as a Claude Code config cloned directly into ~/.claude. It now lives outside any tool's config directory and is linked in per tool, which keeps runtime state out of the working tree and lets the same skill content serve more than one agent. The content is still Claude-flavored — see Tool support for what each tool consumes.

Status

Active personal configuration. The repo is browseable and cherry-pickable — take what's useful, adapt freely. No support, no semver, no backward-compatibility commitment. Things move; what worked yesterday may have been rewritten. Skills and rules track Microsoft Fabric / Power BI / Azure conventions, which themselves drift; verify against current docs before relying on any specific guidance.

If you find this useful, that's the goal. If you find a bug, an issue report is welcome but not guaranteed a response. Personal config first, public artifact second — see Tags for per-artifact reuse-readiness.

Tool support

The repo is structured so agnostic content and tool-specific wiring stay separable:

Portable content — skills/ (the Agent Skills format is an open spec other tools are adopting), claude/rules/ bodies, docs/, and tests/. The MCP templates are the edge case: the protocol is cross-tool but the config schemas are not, so each template lives with the payload whose schema it is written in.
Claude Code-specific — everything under claude/: settings.json, the hook event wiring in claude/hooks/, the subagent frontmatter in claude/agents/, and the paths: auto-load frontmatter on rules (GitHub Copilot's .instructions.md applyTo: globs are the direct analog). Note the split: a rule's body is portable prose, but the file as it sits on disk is a Claude artifact, which is why it lives under claude/.
One instruction file per scope. Root CLAUDE.md is project scope and never deployed; claude/CLAUDE.md is the user-scope payload. They are not mirrors of each other. A root AGENTS.md was carried alongside them until it went unread — Codex was dropped and VS Code's chat.useAgentsMdFile is off — and its one unique section, the fixture-validation procedure, now lives in root CLAUDE.md. Reinstate it from history if a tool that reads AGENTS.md comes back; the content is not Claude-specific.
Other tools — the repo carried a Codex payload and linker until it went unused; see the history around codex/ if you want it back. Skills use the open Agent Skills format, so any tool that reads SKILL.md can consume skills/ directly.
GitHub Copilot — consumes this repo with almost no wiring and no linker of its own. The VS Code agent surface reads Claude's user-scope paths directly: ~/.claude/rules for instructions, ~/.claude/skills for personal skills (in principle — that one path did not resolve when tested; see below), ~/.claude/settings.json for hooks (it parses Claude Code's hook format), ~/.claude/CLAUDE.md for always-on instructions, and ~/.claude/agents for subagents once chat.agentFilesLocations points at it — all of which link-claude.ps1 already populates. That is why there is no copilot/ payload directory: Copilot is a second consumer of the Claude-format payload, not a separate one. The rest is a settings decision:
```
// Checked 2026-08-29 against microsoft/vscode-docs@28f76f5f.
// Only the agents entry is load-bearing: ~/.claude/agents is not a
// documented default (VS Code's user-profile default is
// ~/.copilot/agents). The other three ~/.claude paths are already
// documented defaults, so listing them is belt-and-braces.
"chat.instructionsFilesLocations": { "~/.claude/rules": true },
"chat.agentFilesLocations":        { "~/.claude/agents": true },
"chat.agentSkillsLocations":       { "~/.claude/skills": true },
"chat.hookFilesLocations":         { "~/.claude/settings.json": true },
// chat.useClaudeMdFile governs ~/.claude/CLAUDE.md and already
// defaults to true — no entry needed.
```
Three traps. chat.instructionsFilesLocations accepts folders only — an entry for ~/.claude/CLAUDE.md is silently ignored, and because chat.useClaudeMdFile loads that file anyway, the dead setting looks like it worked. VS Code also omits any setting left at its default when it writes settings.json, so chat.useClaudeMdFile vanishes from the file when set to true and only appears when set to false — absent means on, not unset. And no copy into ~/.copilot/agents or ~/.agents/skills is needed; both only invite drift. MCP is a workspace .vscode/mcp.json (see .vscode/README.md).

Copilot validates skill frontmatter against its own field list and warns rather than failing. Measured 2026-09-04: an unsupported key leaves the skill loaded and listed, with a diagnostic naming it — unlike the claude.ai upload path, which hard-fails. Supported are argument-hint, compatibility, context, description, disable-model-invocation, license, metadata, name and user-invocable. Every skill in this repo already warns: model: is on all 50, paths: on 27, effort: on 9, allowed-tools: on 5. So the warning is ambient, and adding a Claude-only field costs nothing that was not already being paid — but it also means a real frontmatter mistake is camouflaged by the noise.

paths: being unsupported has a consequence worth stating plainly: a conditional skill is unconditional in Copilot. All 27 load unconditionally there, so the conditional/unconditional listing-cost split that governs when_to_use adoption is a Claude Code fact only, and "free because it is conditional" does not transfer.

~/.claude/skills did not resolve when tested 2026-09-04, while project-scope .claude/skills did — even though only the former is listed in chat.agentSkillsLocations and the latter is at its default. The nine workflow skills were absent from the agent surface and the 41 platform skills junctioned into a client repo's .claude/skills were all present. Junctions are not the cause — Copilot read through one to report a field warning on a file whose only copy lives in this repo. Suspect ~ expansion; an absolute path is the thing to try. Until it is resolved, Copilot sees this payload at project scope only.

Copilot parses the Claude hook format, not its semantics. Matchers are read and ignored, so a matcher-scoped hook fires on every tool call; tool input properties are camelCase (tool_input.filePath, not tool_input.file_path); and tool names differ (create_file, replace_string_in_file). The security-reviewer write guard is matcher-scoped, so under Copilot it runs far wider than it does under Claude Code.

Sessions on Agent Host read user-level instructions and agents from harness-agnostic folders (~/.copilot/instructions, ~/.claude/rules, ~/.copilot/agents) instead of VS Code profile user data — the change that removed the "or your user data" fallback from both location tables in mid-2026.

"Agnostic" here means structured so other tools can consume it — the content is written for and validated with Claude Code first.

Install

Cherry-picking? Browse this repo on GitHub and copy individual files into your own agent config. No install needed. The numbered steps below wire up the whole thing.

Clone anywhere (the link scripts resolve the repo location from their own path):

# HTTPS (default — works for any GitHub user)
git clone https://github.com/wardawgmalvicious/agent-config.git

# SSH (requires GitHub SSH keys configured)
git clone [email protected]:wardawgmalvicious/agent-config.git

Deploy into Claude Code. Copies claude/agents, claude/hooks, claude/rules and claude/mcp into ~/.claude/{agents,hooks,mcp,rules}, and junctions the root skills/ one skill at a time into ~/.claude/skills (no elevation needed). Mirrors claude/CLAUDE.md (→ ~/.claude/CLAUDE.md) and claude/settings.json as plain copies. Back up first if you already have a ~/.claude — the script refuses to replace real directories or drifted files without -Force, but review its warnings before forcing anything.
```
./scripts/link-claude.ps1
```
Using GitHub Copilot too? There is nothing to link and, as of 2026-08-29, little to enable — most of the ~/.claude paths step 2 created are already VS Code defaults. ~/.claude/agents is the exception and needs an explicit entry; see Tool support for the block to paste.
Bootstrap pre-commit hooks (installs pre-commit via uv and runs it once across all files). Requires uv on PATH. Skip this step if you only intend to read, not commit.
```
cd agent-config && scripts/bootstrap-pre-commit
```
(Optional) Add the repo's scripts/ directory to PATH so the helpers (instructions-log, lint-frontmatter.py) are callable by name.

Notes:

Plugins, credentials, sessions, memory, caches, and other runtime state remain in each tool's real home directory, not in this repo. That separation is the point of scoped linking.
This config is Windows-targeted (per-skill junctions, PowerShell link scripts, Git Bash for the shell hooks). Hook commands resolve via $HOME so they're portable across users on Windows, but Linux / macOS users will need to symlink manually and adjust paths — no promise it works elsewhere out of the box.

Ongoing workflow

Edit files in place and commit like any other repo. skills/ is the only payload that needs no deploy step — each skill is junctioned, so the tools read the same file. Everything else is a copy and needs a scripts/link-claude.ps1 re-run to reach the live payload: claude/agents/, claude/hooks/, claude/rules/ and claude/mcp/ take repo content on a plain run, while claude/CLAUDE.md and claude/settings.json need -Force. The script verifies everything and exits non-zero if any link or copy needs attention — including after moving or renaming the repo folder, which it repairs automatically.

Those four were junctions until 2026-09-02. The deploy step is the point of the change: none of them is hot-reloaded, so a fresh session was needed either way and the junction bought no immediacy, while making every uncommitted save — and every git switch, stash and rebase, including pre-commit's own stash/restore around a commit — live for every session on the machine. Hooks were the sharp end, because they execute.

Needing no deploy step is not the same as being picked up by a running session, and the two differ by payload type. Skills hot-reload: Claude Code watches skill directories and re-reads them in-session, and that works through these junctions — verified on 2.1.251 for skill add, skill removal, and skillOverrides (upstream fixed in-session skill reload in 2.1.216; this file previously said the opposite). Editing a description in place is the one case not confirmed here, and a description is the trigger, so restart before trusting a changed trigger. Subagents, commands, and rules are not watched — restart after editing those.

One caveat that only bites during a repo reorganization: moving a payload directory leaves ~/.claude/<name> pointing at the vacated path until scripts/link-claude.ps1 runs again. Files are simply absent in that window, and a hook that cannot be found fails open rather than blocking — so re-link before relying on an enforcement hook again. This is not hypothetical: the claude/ regrouping produced exactly four such failures before the re-link.

Handoff discipline

Skills and subagents in this repo are authored via a brief-before-draft pattern: a structured handoff brief — frontmatter specs, body outline, portability caveats, post-draft validation steps — is written and settled before any artifact is drafted from it.

Which surface writes the brief is not part of the pattern. It was originally split across two, a chat session proposing structure and Claude Code drafting from it, because only the latter had filesystem access and only the former could drill sources at length. Claude Code does both now, and the two steps routinely happen in one session. What survives the collapse is the artifact: the brief is the record of what was decided and what was deliberately left out, and the excluded set is what bounds the draft.

The pattern earns its place when an artifact has non-trivial behavioral contracts — refusal patterns, severity rubrics, scope-enforced read-only or destructive guards. For pure reference skills (canonical-answer content), the pattern is overkill; real-use validation suffices.

Templates and worked examples live in docs/handoffs/:

templates/skill-handoff.md — fill-in template for new skills
templates/subagent-handoff.md — fill-in template for new subagents
examples/ — reference briefs derived from validated artifacts

The templates are internal tooling. Consumers cherry-picking from this repo don't need to adopt the brief pattern; the templates are included in case the discipline is useful elsewhere.

License & security

License: MIT — see LICENSE.
Security: see SECURITY.md for reporting issues.

Agent Config

At a glance

README

Agent Config

What this is

Status

Tags

Contents

Shared content

claude/ — Claude Code payload

Mechanism and supporting material

Tool support

Install

Ongoing workflow

Handoff discipline

License & security

Comments (0)