https://hub.zoom.us/doc/oPghJFf8RqSVQdr5Cz8v7w?from=voice_notes_AutoStart
AI Fluency Rubric — Baseline
A simple illustration of where individuals, teams, and departments sit. Drawn only from internal source material:
- Chandler/Mike sync (2026-04-06)
- 04-08 voice memo
ai-rubric-canvas.canvas- Message to Owen (2026-04-08)
No external frameworks layered in. This is the starting point before anything else gets added.
Three dimensions
The rubric operates on three scopes. Each has its own progression.
- Individual — how one person uses AI in their own work
- Team — how AI work is shared and built on inside a team
- Cross-department — how AI-readable artifacts flow between departments
Individual
Three tiers of fluency.
Tier 1: Chat fluency
AI is an ad-hoc conversation. Each session starts cold and the platform decides what context is in the window. The mode is “ask and receive,” not “engineer an environment.”
- Uses ChatGPT, Claude, Cursor, Claude Code, or similar as a conversational assistant
- Understands basic prompt engineering
- Knows which model to pick for which task
- Keeps a prompt library or curates a collection of claude code skills
- Context gets loaded automatically or by whatever file happens to be open; the user does not actively manage what sits in the window
- Each session starts from scratch — no compounding across conversations
Tier 2: Context engineer
AI operates inside a workspace you’ve deliberately shaped. You control what it sees, when, and why — every project becomes an environment you engineer, not just a chat you have.
- Runs projects locally in tools like Claude Code or Claude desktop
- Deliberately indicates what gets loaded into context and when
- Uses CLAUDE.md files at multiple sub folder points and skills to discriminately load context
- Uses subagents for context window control
- Uses intentional compaction to clear context at specific artifact creation points
- Can build custom apps on top
Tier 3: Compound / cross-project fluency
Your workspaces compound. Every new project inherits context, structure, and lessons from the ones before it, so you start each new effort miles ahead of a blank new one.
- Projects are housed together and benefit from each other
- New projects or tickets start faster and better because they draw on context and architecture from prior projects
- Uses multiple CLAUDE.md files across project collection folders in a hierarchy of progressive context loading
- You practice intent engineering to optimize for strategic business and customer outcomes
- This is where individual-level impact maxes out
Team
The “Lego brick” framing. When someone produces output (research, analysis, a feature spec, a deliverable), value multiplies when:
- Primary artifacts are visible to teammates — not stuck on one person’s laptop or in a private doc
- Artifacts are machine-readable and shared — structured so an LLM can consume them, not just a human
- The “how you got there” is accessible — the context engineering, inputs, prompts, pipeline, and outputs that produced the deliverable
- These become reusable building blocks — teammates can repurpose them for new work
The behavioral shorthand: a teammate can walk in, pick up the pipeline (not just the output), and continue the work without a meeting.
The opposite: work is locked in one person’s head or in artifacts only a human can parse.
Cross-department
Two directions.
Upstream metabolization (am I consuming?)
- Are you pulling in the primary artifacts from departments upstream of you?
- Are those artifacts being used as inputs in your own context engineering?
- Example: Jordan in marketing pulling the roadmap, PRD and the code that actually shipped, so she can see what was really built and how it differs from what was originally planned.
- Announcement email
- Sales collateral
Downstream enablement (am I producing for others?)
- Are your primary artifacts machine-readable?
- Are they structured so downstream departments can consume them directly?
- Example: A CSM capturing structured customer feedback from every meeting transcript — pain points, feature asks, objections — so product can pull it straight into roadmap planning instead of waiting for it to surface in a meeting or Slack thread.
The failure state: the telephone game
When both directions are broken, information travels through human brains:
- Morgan, Spencer, engineering heads accumulate knowledge of what was built
- Brad gathers it from them, filtered by his comprehension and bandwidth
- Brad translates it for Nick
- Nick translates it for each sales rep
- Each rep translates it for their customers
- Customers translate it for their own teams
Five to six layers of signal degrading. Stale within days. Dependent on each person’s ability to comprehend and articulate.
The target state
- Customer feedback and research is codified into machine-readable form once, then reused
- PRDs and roadmaps run through a translation layer
- Engineers commit the plan and trade-offs alongside the code
- API docs generate from the plan plus committed code, not from retroactive human updates
- Sales conversations get captured as machine-readable text and loop back to product and engineering in near-real-time
- A rep who goes off-script can be corrected in days, not weeks
What the rubric is for
Two purposes:
- Assessment. Score where individuals, teams, and departments actually sit. Make the problem concrete instead of a subjective judgment call.
- Change vehicle. The process of co-creating the rubric with executives forces the alignment conversation. The rubric becomes the objective framework that grounds “you can see how this is bad and you can see how it has to be.”
The forcing function is not mandating AI use. It is mandating the level of quality and the volume of output such that the current workflow cannot hit the bar without AI.
Signals from the canvas
Observable things that separate higher fluency from lower, regardless of tier:
- Clarity of thinking
- Reduction in context switching and ramp-up time
- Has identified the hardest, most valuable part of their work and is pointing AI at it (not at boilerplate, stubs, or tests)
- Can identify why an AI output didn’t work and how to approach it differently
- Modifies the harness and plumbing with lessons learned to eliminate similar problems in future projects
- Tight control of the context window
- Decomposes advanced projects into distinct AI prompt pipelines
- Can manage agents: define what good looks like, write specs, give feedback, iterate