Tool mode
AI completes a bounded task: summarize a source, inspect a repo, restructure a file, draft a pass, chart data, write code, or run checks. I can review the result directly.
My Workflow in 30 Seconds
- Dictate messy intent: goals, constraints, edge cases, examples, worries, and what the output must do.
- Build the prompt with Prompt Compass: objective, context, constraints, steps, output format, success criteria, and a self-check the AI runs on its own output.
- Route the work: Codex for code and files, Claude for thinking and documents, ChatGPT or Gemini for research, specialized MCP compasses for domain-specific checks.
- Verify the output: I own the acceptance criteria. Tests, source checks, diff review, and interpretation all happen here, not inside the chat.
- Remember what is durable: only decisions, caveats, and patterns that matter across future tasks get written back to my second brain.
A finalized prompt, one that specifies what a successful output looks like and how the AI should validate its own response, is what I call a contract. Once it's running, I just call it the prompt.
The Three AI Modes
Oracle mode
This is the danger zone: the model gives a polished answer while silently choosing the method, evidence, exclusions, segments, and caveats.
System mode
The workflow is designed before execution: prompt, routing, source grounding, specialized checks, verification, and memory write-back. The system moves fast because the judgment layer is explicit.
My Workflow
Dictate Intent
I start with rough context: goals, constraints, edge cases, examples, worries, desired outputs, and success criteria.
Build the Prompt
Prompt Compass helps turn that raw intent into a clear prompt with objective, audience, context, constraints, steps, output format, success test, assumptions, tool rules, and stop conditions.
Route by Interface
Codex handles implementation: coding, repo inspection, file shifting, cleanup, testing, and verification. Claude helps clarify what I actually want, refine messy thinking before Codex, create human-facing documents and spreadsheets, and handle visual QA or final polish. ChatGPT is mostly quick inquiry, search, and occasional deep-research packets. Gemini sits in a similar research lane and also shows up through third-party tools like Stitch.
Use MCPs
I also use local AI co-pilots called MCPs: tools connected to my files, scripts, automations, and databases. This includes my second brain, Knowledge Atlas, and specialized compasses like Prompt Compass, Critical Compass, and UX Heuristics Compass.
Ground Context
My second brain is a central AI intelligence layer: a personalized local knowledge system with structured nodes and schemas. It gives AI continuity across projects and can store durable decisions, notes, and caveats. It is an MCP layer, not just a note-taking tool.
Compass Checks
Each compass is a focused AI tool packed with scripts, automations, and domain-specific databases. They improve accuracy and quality while reducing token usage because the AI retrieves targeted context instead of carrying everything in one oversized prompt.
Verify + Remember
I still own the acceptance criteria. For code, that means tests, smoke checks, and diff review. For research, it means sources, uncertainty, caveats, and interpretation. Only durable learning gets written back.
Weak vs. Strong Prompt
Weak prompt
Analyze these survey responses and tell me what users think.
Strong prompt
Before writing findings, inspect the instrument, sample, missingness, unit of analysis, cleaning rules, scale direction, segmentation risks, uncertainty, and the decision the analysis supports. Show what would make the output trustworthy.
My Route
Prompt Compass turns the goal into a clear prompt. Codex can inspect files and produce reproducible analysis code. Research Compass can hold source notes. Critical Compass pressure-tests claims. My second brain stores the final decision and caveats.
Human Judgment
It does not silently decide whether a noisy segment is meaningful, whether respondents generalize to the target population, or whether a correlation supports a product recommendation.
Try It Yourself
Before Delegating
- Where does messy intent become a finalized prompt?
- What would a successful output look like, and how can the AI check that?
- Which interface gets the work: Codex, ChatGPT, Claude, Gemini, or a specialized MCP?
- What memory layer or knowledge base can guide the task?
- Where do specialized compasses, tests, source checks, or visual QA happen?
- Where must uncertainty appear?
- What can AI draft, code, organize, or summarize, but not decide?
- What should be remembered afterward?
Start here: pick one task you're about to hand to AI and write down what a successful output looks like before you begin.
Discussion Questions
- How are you drawing this boundary in your own workflow?
- Do you finalize the prompt before AI starts the work, or does it evolve during the chat?
- What gets routed to a specialized tool instead of one general model?
- Where do you require a human loop-back?
- What gets remembered after the task is done, and where does that memory live?
Next Step
About the author
Jonathan Kyle Hobson is a UX researcher, product strategist, and AI tool builder with 10+ years of experience across design, storytelling, and emerging technology. He holds a master's degree in User Experience from Arizona State University. Kyle built this workflow and its specialized compasses to make AI delegation faster without hiding the human choices behind method, evidence, boundaries, verification, and memory.