Career Copilot

What It Is

A private thinking partner for exploring the next chapter of my career, using a folder of markdown files (values, strengths, reflections, hypotheses, learnings) that I work through with Claude Code (Opus) running inside Cursor. The AI coach pushes back when my reasoning drifts from my stated values, surfaces patterns across months of journaling and conversations, and helps me clarify what I want next. After more than a decade running Wonderstorm, I needed to re-ask questions I hadn't asked since my early thirties, and leveraging AI to synthesize the journey has been enlightening.

What I Learned

A coach is only useful if it has a backbone — I wrote pushback into the system prompt ("be honest and direct, push back when needed"), and the most valuable sessions are the ones where Claude reminds me I asked for brutal honesty and then names the trap I'm falling into: getting distracted by compensation, prestige, or an exciting problem when the role may otherwise contain things that conflict with what I stated I wanted 6 months ago. The pushback is how I pressure-test what I'll actually compromise on and what I won't.
Pattern recognition across a growing knowledge base is the part you can't get from a single conversation — One example Claude surfaced from my various reflections: every meaningful career move I've made (PE to gaming, Riot to Wonderstorm, and now toward AI) was driven by searching for purpose over status or promotion. I hadn't connected those three transitions until Claude tracked them across separate files and named the underlying pattern, then told me to trust it.
The harness around the model matters as much as the model — Running Opus inside Claude Code's terminal (with plan mode, ultrathink, ask-the-user, and subagents available, plus the file context, persistent memory, and skills I've layered on top) produced richer discussion than the same model accessed through Cursor's built-in agent. Some of it is that Claude's personality comes through more directly when there are fewer layers in between. The takeaway has generalized: across every project on this site, the harness (context, tools, memory, saved workflows) does more for output quality than swapping models.
Context is what compounds across sessions — Model intelligence resets every session unless you build durable scaffolding around it. ChatGPT and Claude have started carrying memory across chats, which helps, but the most leveraged version I've found is still a markdown store the AI can read, search, and edit in place across months. The longer this project runs, the better Claude gets at this specific corner of my life.
An index file for knowledge is worth building — Without one, I noticed by observing tool calls that Claude only reads certain files, and its selection can feel haphazard and leave gaps. The fix is to have AI summarize a large body of source docs into a synthesis index with short descriptions pointing to the full files, similar to how CC skills work. Claude reads the index first, then goes deeper where it needs to.
ChatGPT Projects work right up until the knowledge base needs to be alive — I started with a dozen documents uploaded to a ChatGPT Project, but every new insight required me to download, edit, re-upload, and delete the old version, and the friction killed the practice. Moving everything to markdown in Cursor (synced via Obsidian and iCloud) turned a static reference into a living document the AI can edit in place.

How It Works

A markdown-first knowledge base across folders (values, strengths, reflections, hypotheses, learnings, role funeral, scorecard) so every artifact is searchable, diffable, and editable in place by Claude
Claude Code with Opus inside Cursor's terminal for the deeper reflection sessions, and Cursor's own agent for quick file edits
Custom agents.md and CLAUDE.md files that wire in the pushback behavior (be a thought partner, not a yes-man) and tell the model to surface contradictions across files proactively
Symlinked to an Obsidian vault on iCloud so I can capture on mobile (Granola notes from a coffee chat, dictated thoughts on a walk) and reflect on the laptop later, with no friction between capture and synthesis
A small markitdown (Python) pipeline for converting source PDFs and Word docs (reflection exercises, StrengthsFinder reports, journaling exports) into markdown the model can read and edit
A weighted scorecard (CAREER-SCORECARD.md) with must-haves, major priorities, and hard-no's that Claude references whenever I evaluate a new opportunity, so the criteria stay consistent

Built With

Started as a ChatGPT Project with a system prompt that asked it to be a career coach and thought partner (with the backbone to push back when warranted), then migrated to Cursor once the knowledge base needed to live as files I (and the AI) could edit in place. Today the building blocks are Cursor for the file system and quick edits, Claude Code with Opus for the deeper reflection sessions, Obsidian and iCloud for capture from anywhere, and a small markitdown pipeline for converting source PDFs and Word docs into markdown.

Loading content...

What It Is

What I Learned

A coach is only useful if it has a backbone — I wrote pushback into the system prompt ("be honest and direct, push back when needed"), and the most valuable sessions are the ones where Claude reminds me I asked for brutal honesty and then names the trap I'm falling into: getting distracted by compensation, prestige, or an exciting problem when the role may otherwise contain things that conflict with what I stated I wanted 6 months ago. The pushback is how I pressure-test what I'll actually compromise on and what I won't.

Pattern recognition across a growing knowledge base is the part you can't get from a single conversation — One example Claude surfaced from my various reflections: every meaningful career move I've made (PE to gaming, Riot to Wonderstorm, and now toward AI) was driven by searching for purpose over status or promotion. I hadn't connected those three transitions until Claude tracked them across separate files and named the underlying pattern, then told me to trust it.

The harness around the model matters as much as the model — Running Opus inside Claude Code's terminal (with plan mode, ultrathink, ask-the-user, and subagents available, plus the file context, persistent memory, and skills I've layered on top) produced richer discussion than the same model accessed through Cursor's built-in agent. Some of it is that Claude's personality comes through more directly when there are fewer layers in between. The takeaway has generalized: across every project on this site, the harness (context, tools, memory, saved workflows) does more for output quality than swapping models.

Context is what compounds across sessions — Model intelligence resets every session unless you build durable scaffolding around it. ChatGPT and Claude have started carrying memory across chats, which helps, but the most leveraged version I've found is still a markdown store the AI can read, search, and edit in place across months. The longer this project runs, the better Claude gets at this specific corner of my life.

An index file for knowledge is worth building — Without one, I noticed by observing tool calls that Claude only reads certain files, and its selection can feel haphazard and leave gaps. The fix is to have AI summarize a large body of source docs into a synthesis index with short descriptions pointing to the full files, similar to how CC skills work. Claude reads the index first, then goes deeper where it needs to.

ChatGPT Projects work right up until the knowledge base needs to be alive — I started with a dozen documents uploaded to a ChatGPT Project, but every new insight required me to download, edit, re-upload, and delete the old version, and the friction killed the practice. Moving everything to markdown in Cursor (synced via Obsidian and iCloud) turned a static reference into a living document the AI can edit in place.

How It Works

A markdown-first knowledge base across folders (values, strengths, reflections, hypotheses, learnings, role funeral, scorecard) so every artifact is searchable, diffable, and editable in place by Claude

Claude Code with Opus inside Cursor's terminal for the deeper reflection sessions, and Cursor's own agent for quick file edits

Custom agents.md and CLAUDE.md files that wire in the pushback behavior (be a thought partner, not a yes-man) and tell the model to surface contradictions across files proactively

Symlinked to an Obsidian vault on iCloud so I can capture on mobile (Granola notes from a coffee chat, dictated thoughts on a walk) and reflect on the laptop later, with no friction between capture and synthesis

A small markitdown (Python) pipeline for converting source PDFs and Word docs (reflection exercises, StrengthsFinder reports, journaling exports) into markdown the model can read and edit

A weighted scorecard (CAREER-SCORECARD.md) with must-haves, major priorities, and hard-no's that Claude references whenever I evaluate a new opportunity, so the criteria stay consistent

Built With