LeadUp Skill Router

Purpose

Take any inbound request (user task, client message, sales-call notes, WhatsApp thread, half-finished spec) and decide:

1. The intent category (development / growth / sales / SaaS / QA /

automation / security / creative / research / content / new project).

1. The workflow pack (one of 11).
2. The ordered skill sequence to ship the work.
3. The model to run it on (DeepSeek for low risk, Claude

Sonnet/Opus for high risk).

1. Whether internet / browser / MCP / project files are needed.
2. The risk level and whether approval is required.
3. Which memory files (STATUS.md, DEPLOY.md, PROJECT.md,

AGENTS.md) the next session must update.

1. A copy-paste super-prompt the user can hand to Claude Code,

RuFlo, or opencode.

This skill is an orchestrator, not a worker. It does not do the underlying SEO audit, code change, or proposal itself.

When to use

Use when the user is unsure which skill applies, or when the request spans multiple skills.

Trigger phrases: "which skill should I use", "use router", "route this task", "analyze my request", "what workflow needed", "choose best skill", "understand this task", "convert this into workflow", "which LeadUp skill", "auto select skill", "master router", "skill router", "workflow router", "decide skill pack", "rough idea to workflow".

Do not trigger when the user names a skill directly ("use leadup-seo-strategist") or when the request maps cleanly to one skill the user already knows (let that skill run).

Inputs needed

• The request, in whatever shape the user has it (paragraph, bullets,

WhatsApp thread, voice-note transcript, link).

• Project context if known (Hostendor / LeadUp site / Jewellery SaaS /

Salon SaaS / INET CRM / Trivasia / Security suite / other).

• Audience and market hint (India / global / both) if relevant.
• Constraints (deadline, budget, regulated category, do-not-deploy).
• Whether the user wants planning only or planning + execution.

Ask at most 2 clarifying questions, only if the intent is genuinely ambiguous (e.g. "fix it" with no target).

Intent classification workflow

Classify the request into one of 11 categories. Use the dominant verb + object pair:

If the request spans two categories (very common: e.g. "SEO + blog + human polish"), pick the primary category for the pack and add the other category's skills into the sequence.

Full cue list: references/skill-routing-table.md.

Skill pack routing table

11 packs, each with a default skill order. Pick the pack, then trim or extend the sequence to fit the request.

Full mapping with cross-pack hand-offs: references/workflow-pack-map.md.

Skill sequence planner

1. Restate the request in one line.
2. Pick the primary pack.
3. Trim the default sequence to the steps that actually apply.
4. Append cross-pack skills when needed:

• Anything public-facing → end with leadup-human-content-editor.
• Anything touching production / DB / payments → end with

leadup-security-review + leadup-deploy-checker + leadup-release-manager.

• Anything new → start with

leadup-super-prompt-builder and/or leadup-project-kickoff / leadup-existing-repo-analyzer.

• Anything that needs external research → insert

leadup-mcp-tool-orchestrator early.

• Anything touching PII / regulated → insert

leadup-pii-risk-reviewer before code.

1. Cap the sequence at 5–7 skills; longer chains lose focus.
2. Mark which skills are planning only vs execution so the

approval gate is clear.

Sequence shape: assets/workflow-sequence.template.md.

Model routing rules

• LOW RISK → DeepSeek: docs, prompts, content, SEO planning,

keyword ideas, status update, simple analysis, social calendar, blog draft, proposal draft, PII inventory write-up.

• MEDIUM RISK → DeepSeek first, Claude Sonnet review if needed:

UI changes, Docker fixes, API planning, admin-panel changes, normal feature code, n8n workflow design, ads plan, CRO audit.

• HIGH RISK → Claude Sonnet or Opus only: payments, webhooks,

authentication, PII handling code, production deploy, DB migration, server recovery, large dirty-repo cleanup, architecture decisions, multi-tenant boundary work, security review write-ups.

Tie-breakers and example mapping: references/model-routing-rules.md.

Step-by-step workflow

1. Restate the request in one line; note project + audience +

constraints.

1. Classify intent into one of 11 categories (Intent classification

workflow).

1. Pick the workflow pack (Skill pack routing table).
2. Plan the ordered skill sequence (Skill sequence planner).
3. Choose the model per step (Model routing rules).
4. List files / resources the next session must read

(STATUS.md, DEPLOY.md, PROJECT.md, AGENTS.md, repo paths, client docs).

1. Decide research + tools needed (browser / MCP / search /

Playwright / Docker).

1. Set the approval gate (planning OK without approval; commit,

push, deploy, migrate, refund, send to client → approval required).

1. Name memory updates owed at the end (which STATUS.md etc.).
2. Output the router report and a copy-paste super-prompt.

Required output format

One Markdown report with these sections, in this order. The super-prompt is a separate fenced code block at the end.

1. Request understanding — one paragraph restating the user's

request in plain English.

1. Intent category — single line: one of the 11 categories.
2. Workflow pack selected — pack number and name.
3. Skill sequence — ordered list of 1–7 LeadUp skills, each with a

one-line job ("leadup-seo-strategist — audit + 30/60/90 plan").

1. Recommended model / tool — model per skill step (DeepSeek /

Claude Sonnet / Claude Opus) with a one-line reason.

1. Risk level — Low / Medium / High with one-line reason.
2. Files / resources to read — STATUS.md, DEPLOY.md,

PROJECT.md, AGENTS.md, specific repo paths, public links.

1. Research / tools needed — browser, MCP server, Search Console,

Lighthouse, Razorpay docs, public-apis, etc.; each marked available or requires verification.

1. Approval needed? — yes/no per step; reason in one line. Default

approval-required for commit / push / deploy / migration / refund / external send.

1. Memory updates owed — list of files the next session must

update at the end (STATUS.md, DEPLOY.md, PROJECT.md, AGENTS.md, CHANGELOG.md).

1. Execution plan — short numbered list the user will hand to the

next agent.

1. Super-prompt (copy-paste) — a single fenced code block

containing the full prompt for Claude Code / RuFlo / opencode, pre-loaded with project context, the skill sequence, the safety + approval rules, and the expected output format. Built on the assets/super-prompt-output.template.md shape.

Templates: assets/router-output.template.md, assets/super-prompt-output.template.md, assets/workflow-sequence.template.md.

Safety rules

• Never commit / push / deploy / migrate / refund / send to client

without explicit user approval. The router only plans; the worker step asks for approval at its own gate.

• **Never read or print real .env values, API keys, tokens, secrets,

card data, Aadhaar, or other PII** in the router output. Use __SET_ME__ placeholders and "redacted" markers only.

• Never ask the user to paste secrets. Reference env names only.
• **High-risk routes (payments, security, PII, DB migration, prod

deploy)** → force Claude Sonnet/Opus + add leadup-security-review / leadup-pii-risk-reviewer to the sequence + insert an approval gate before code.

• Browser / Playwright / MCP usage → route through

leadup-mcp-tool-orchestrator (discover read-only first, escalate with approval).

• Public-facing copy → always end with leadup-human-content-editor

before publishing.

• Regulated industries (clinic / finance / kids / real estate /

alcohol / gambling) → require leadup-pii-risk-reviewer + legal/compliance flag in the report.

• Planned-but-not-yet-built skills

(leadup-multi-tenant-architect, leadup-backup-rollback-planner) → substitute the closest shipped skill and mark the gap in the report.

Full rule set: references/safety-routing-rules.md.

Common mistakes

• Treating the router as a worker — doing the SEO audit / writing the

code itself.

• Producing a 12-skill chain — users can't run that. Cap at 5–7.
• Forgetting leadup-human-content-editor on public-facing copy.
• Skipping leadup-security-review / leadup-pii-risk-reviewer on

PII / payments / regulated work.

• Picking Claude Opus for low-risk content tasks (cost without need).
• Picking DeepSeek for payment / auth / DB migration work (risk).
• No approval gate on commit / push / deploy / migrate steps.
• Output without a copy-paste super-prompt at the end.
• Ignoring memory updates (STATUS.md etc.).

Troubleshooting

• Ambiguous request: ask up to 2 clarifying questions; if still

unclear, pick the safest pack and flag the assumption.

• Multi-intent request: pick primary pack; append cross-pack skills

to sequence; cap at 7.

• User wants execution, not planning: still produce the router

report and the super-prompt; tell the user the next agent must ask for approval at sensitive steps.

• High-risk + tight deadline: do not downgrade the model. Time

pressure is not a reason to use DeepSeek on auth/payment code.

• No project context: write the report with placeholders; ask the

user to fill before kicking off the worker.

• Skill referenced is "planned": substitute the closest existing

skill and call out the gap in the "Skill sequence" section.

• Request references a non-LeadUp tool: still route through the

closest LeadUp skill (leadup-mcp-tool-orchestrator for tool discovery, leadup-api-research-builder for vendor API, etc.).

Test prompts

Should trigger (5)

1. "Which LeadUp skill should I use for this client message?"
2. "Route this task: 'Create SEO plan and blog content for Hostendor'."
3. "Auto select skill — we need to add WhatsApp reminders to our salon SaaS."
4. "Master router: client wants a free maps API for a Flutter app."
5. "Convert this rough idea into a workflow — 'make our landing page convert better'."

Should NOT trigger (3)

1. "Use leadup-seo-strategist to audit our site." (direct skill call → run that skill)
2. "Write the actual blog post now." (→ leadup-blog-content-writer)
3. "Just polish this copy." (→ leadup-human-content-editor)

Functional test cases (2)

1. Given "Create SEO plan and blog content for Hostendor" with no

regulated category and no PII, return a Growth Marketing Pack route with 4–5 skills (leadup-growth-research-agent → leadup-seo-strategist → leadup-keyword-competitor-researcher → leadup-blog-content-writer → leadup-human-content-editor), DeepSeek for planning/content + optional Claude Sonnet review, no approval needed for planning, approval needed before any website code changes, memory updates to STATUS.md, and a copy-paste super-prompt.

1. Given "Add Razorpay refund flow to our jewellery SaaS prod", return

a High-risk route with Claude Sonnet/Opus only, mandatory leadup-security-review + leadup-pii-risk-reviewer early, approval gates before commit / push / deploy / refund test, full QA + release pack at the tail (leadup-qa-test-case-generator → leadup-deploy-checker → leadup-release-manager), and memory updates to STATUS.md + DEPLOY.md.

Success criteria

• All 12 output sections present in order.
• Skill sequence is 1–7 ordered LeadUp skills, each existing or

flagged as planned.

• Model per step matches risk level (no DeepSeek on auth / payments /

PII / migration).

• Public-facing copy ends with leadup-human-content-editor.
• PII / payments / regulated routes include leadup-security-review

and leadup-pii-risk-reviewer.

• Approval gate explicit per step (planning vs commit / push / deploy

/ migrate / send).

• Memory updates named (STATUS.md, DEPLOY.md, etc.).
• Final fenced code block is a copy-paste super-prompt that loads the

next agent with project context, sequence, safety + approval rules, and expected output.

• No real secrets, PII, or live keys anywhere in the output.