My Router

My Router

Overview

A routing layer that sits between superpowers (outer workflow) and domain skills (inner execution). It reads the working directory, file types, and prompt to decide which skill to invoke. The user does not need to remember skill names; the router selects automatically.

When to Use Superpowers vs. Direct Dispatch

The router decides this. Not all tasks need superpowers' full ceremony.

| Effort signal words in prompt (e.g., "extensively", "extensive", "deep", "thorough", "in-depth", "carefully", "comprehensive") | Superpowers + extended thinking — enable extended thinking (Alt+T) and route through superpowers regardless of task shape | The user is explicitly asking for more deliberation |

The rule: if the domain skill is obvious and the scope is clear, skip superpowers and dispatch directly. If the task needs exploration or planning, let superpowers run the outer loop and the router dispatches during execution. If the user signals they want deep effort, always use superpowers with extended thinking enabled.

Integration with Superpowers

When superpowers is active, it handles workflow phases: brainstorm → plan → execute → verify. The router activates during the execute phase and dispatches to the right domain skill.

When superpowers is not active (direct dispatch or quick task), the router works standalone.

How Routing Works

At dispatch time, the router checks three signals in order (keywords, file types, project structure). In a consuming repo, before applying the table below, it first merges any consumer-local routing extensions: a routing-table.local.md at the repo root, or a ## Routing section in AGENTS.local.md. Those two files survive bootstrap, so they are where a consuming repo registers project-local skills that bootstrap would otherwise overwrite; on a keyword or file-type conflict, the local row wins.

1. Prompt keywords (highest priority)

The user's prompt often contains the clearest signal. Match keywords to skills:

When an explanatory figure request explicitly asks for PowerPoint or native editability, use editable-figure before the generic figure, proposal, README-polish, or slide routes. A request to simplify an existing editable figure stays on that route. A full slide deck still uses the presentation route, and an explicit SVG, TikZ, screenshot, or prompt-only request keeps its requested format. Before taking the route, check that the session can perform the desktop PowerPoint validation editable-figure requires. If it cannot, explain the limitation and offer ci-mockup-figure only when its output fits the request; do not silently substitute HTML for an explicitly requested PPTX.

2. File types in working directory

If prompt keywords are ambiguous, inspect the files being worked on:

3. Project structure hints

Some projects declare their type in AGENTS.local.md or directory naming:

• Directory named proposals/, nsf-*, nih-* → proposal context
• Directory named papers/, manuscript/, draft/ → paper context
• Submodule linked to Overleaf → shared paper/proposal, apply extra caution per submodule safety rules

Dispatch Rules

1. Local-first override → before dispatching, scan skills/ for any local skill that is a more specific variant of the matched skill (e.g., nsf-bibref-filler over shared bibref-filler, or nsf-figure-builder over shared figure-prompt-builder). If a local variant exists, use it instead.
2. If a prompt keyword matches → invoke that skill (after applying rule 1).
3. If file context matches but prompt is vague (e.g., "help me with this") → state the detected context and proposed skill, ask the user to confirm before proceeding.
4. If multiple skills could apply (e.g., "review this proposal" matches both implement-review and local proposal skills) → use implement-review as the outer loop and ask the user which agency lens to apply (NSF, NIH, etc.). Load any local proposal skills as supplementary criteria.
5. If nothing matches → fall through to superpowers or general Claude behavior. Do not force a skill where none fits.

Skill Lookup Order

In consuming project repos, skills can be project-local, pack-deployed, or bootstrapped from shared config. When dispatching, look for each skill in this order:

1. skills/<name>/SKILL.md: project-local (highest priority). Projects may have their own custom skills here that are not in the shared config. The router should scan skills/ for any available SKILL.md files, not just the ones listed in the routing table below.
2. .claude/skills/<name>/SKILL.md: pack-deployed by anywhere-agents pack install. The .claude/ prefix is a historical Claude Code convention; the SKILL.md contents are agent-agnostic.
3. .agent-config/repo/skills/<name>/SKILL.md: bootstrapped from shared config.
4. Installed plugins: Claude Code plugins (e.g., frontend-design, superpowers, pyright-lsp) provide their own skills and capabilities. Check /skills output for plugin-provided skills alongside the three locations above. Plugin skills may overlap with custom skills; prefer the more specific one (e.g., a project-local nsf-proposal-composer over a generic plugin writing skill).

If a project-local skill matches the task better than a pack-deployed or bootstrapped skill, prefer the project-local one. The router itself follows this same lookup order: in the agent-config repo it lives at skills/my-router/; in other projects it lives at .claude/skills/my-router/ or .agent-config/repo/skills/my-router/.

Skill Availability

Skills come from three sources with different availability:

Before invoking a reference skill, check that its SKILL.md exists in the current project (in skills/, .claude/skills/, or .agent-config/repo/skills/). If it does not exist, inform the user: "This task would use [skill name], but it is not available in this project. You can copy it from https://github.com/yzhao062/agent-config/tree/main/reference-skills/."

Combining with Dual-Pass Workflow

For tasks that benefit from build-then-audit:

• First pass: router dispatches to the domain skill (e.g., bibref-filler adds citations)
• Second pass: router dispatches to implement-review or the domain skill's audit mode

The user can say "two-pass this" and the router will wrap the domain skill in dual-pass-workflow automatically.

Combining with Implement-Review

When review is needed after domain skill execution:

1. Domain skill runs (e.g., proposal writing, citation filling)
2. Changes are staged
3. Router dispatches to implement-review with the appropriate lens (paper, proposal, or code)
4. Codex reviews using the lens-specific criteria

Examples

User says: "Add citations to the related work section" → Router detects: keyword "citations" → checks for local variant (e.g., nsf-bibref-filler) → if found, uses it; otherwise falls back to shared bibref-filler

User says: "Review this" → Router detects: staged changes exist, .tex files in a proposal directory → implement-review, asks user which agency lens (NSF, NIH, etc.)

User says: "Make an editable PowerPoint overview figure for this proposal" Router selects editable-figure, applies the proposal context, and uses the available presentation tooling to build and inspect a native PPTX.

User says: "Make an overview figure for the proposal" → Router detects: keyword "figure" + "proposal" → checks for local nsf-figure-builder → if found, uses it; otherwise falls back to shared figure-prompt-builder

User says: "Help me improve this section" → Router detects: .tex in proposal directory, no specific keyword → asks user: "This looks like proposal work. Should I use local proposal skills, or is this general editing?"

User says: "Build the feature and review it" → Router detects: code context → superpowers handles the build, then router dispatches to implement-review with code lens