m2ai-portfolio Failure-Mode Tool Router
Takes a task description and routes it to the correct AI tool category based on the failure modes of wrong-tool choices. Distinct from model routing (which picks the tier within a single provider) — this routes across tool categories and harness types.
Trigger
Use when the user says "/failure-mode-tool-router", "which AI tool should I use for this", "should I build an agent or use ChatGPT", "is this a Claude Code task", "which harness for this", or when the user is uncertain whether to automate, use a chat interface, or build a custom agent.
Phase 1: Intake
Collect from the user:
- Task description — what is being done? One paragraph is ideal.
- Frequency — one-time, recurring, or event-triggered?
- Inputs available — what data is accessible and in what form?
- Stakes — consequences if the output is wrong?
If the task description is too vague (under 2 sentences), ask: "Can you describe one real instance of this task from start to finish?"
Phase 2: Classify Against Tool Categories
Evaluate the task against five categories. Pick the BEST fit, then name the top two wrong-tool failure modes.
Category A: Scheduled / Repeating Automation
Profile: The task runs on a schedule or known trigger, produces structured output, crosses 2+ tools, follows a predictable path, and success criteria are pre-specifiable.
Right tool: A scheduled agent harness (n8n, Make, Workspace Agents, or a custom scheduled Claude Code agent).
Do NOT use this for: Novel research tasks, tasks requiring judgment on new information every run, or single-occurrence tasks.
Category B: Novel Research and Synthesis
Profile: The task requires surfacing information that is new, unknown, or not in a structured database. Output quality depends on web freshness. Criteria: recent, sourced, synthesized.
Right tool: A research-oriented tool (Perplexity, search-augmented LLMs, AutoResearch agents, or Claude with web search MCP).
Do NOT use this for: Repeating ops tasks (use Category A), single polished documents (use Category C).
Category C: Single Polished Artifact
Profile: One-time or rare task. The goal is a high-quality document, report, plan, or analysis that will be reviewed and edited by a human. Speed matters less than quality.
Right tool: A capable chat LLM with large context (Claude, GPT-4, Gemini) in an interactive session. No agent loop needed.
Do NOT use this for: Tasks that need to repeat (use Category A), tasks that need real-time data (use Category B).
Category D: Long-Horizon Autonomous Execution
Profile: The task spans multiple hours or sessions, requires navigating ambiguity across many steps, and cannot be fully pre-specified. A human checking in periodically is acceptable.
Right tool: A long-context autonomous agent harness with human-in-the-loop gates (Claude Cowork, agentic IDE, or a fully-supervised agent run).
Do NOT use this for: Well-defined repeating ops (Category A), tasks requiring fresh research (Category B).
Category E: Custom Code / Tool Build
Profile: The task produces a reusable artifact (tool, script, API, MCP server, database migration). Output must be technically correct, version-controlled, and maintainable. Failure means broken software.
Right tool: Claude Code (or equivalent code-first agent) with file system access, test running, and git integration.
Do NOT use this for: Document generation (Category C), ops automation where the output is a structured action rather than code (Category A).
Phase 3: Anti-Recommendation
After picking the best category, explicitly state the top 2 wrong-tool choices and why they fail for this task.
Format:
Wrong tool 1: {tool name}
Why it fails: {one sentence — the specific failure mode}
Wrong tool 2: {tool name}
Why it fails: {one sentence — the specific failure mode}
Examples of failure modes:
- "Building a scheduled agent for novel research tasks leads to stale output — the agent's knowledge doesn't refresh."
- "Using an interactive chat for a daily ops task means it only runs when someone opens a browser."
- "Using a research tool for a coding task produces descriptions of code, not executable code."
Phase 4: Output
TOOL ROUTING DECISION
=====================
Task: {one-line summary}
Best fit: Category {A|B|C|D|E} — {category name}
Recommended: {specific tool or approach}
Rationale: {one sentence}
Wrong-tool warnings:
1. {tool} — {failure mode}
2. {tool} — {failure mode}
Next step: {concrete action — e.g., "Run workflow-fit-scorer to confirm automation readiness" | "Start an interactive Claude session with large context" | "Set up a scheduled harness using the scheduled-agent-harness skill"}
What This Does NOT Do
- Does not recommend a specific model tier within a tool. Use a model router for that.
- Does not evaluate cost. Use a cost model tool for budget-constrained decisions.
- Does not score the workflow's automation fit. Use a workflow fit scorer for that.
- Does not account for existing tool licenses or vendor lock-in — assumes access to all tools.
Source
Framework extracted from Nate Kadlac newsletter (2026-04-27): "Your team spends 5 hours a week on work a sales consultant automated in an afternoon." Nate's anti-recommendation logic for Workspace Agents — explicitly steering users away from it when the workflow doesn't fit the 5 properties — is the seed for this generalized tool router.