◆ Methodology
Your overall score starts from equal base weights within your source, applies modest
role-based adjustments, and rewards balanced strength across dimensions. A small quality
adjustment based on correction rate applies when enough messages are available. Reference
values shown below are engineer baselines and scale with inferred role. Each dimension
explains what it measures, why it matters, how it's computed, and what to do to grow it.
Methodology version 4
Cross-tool dimensions
These dimensions score across more than one tool, wherever the underlying signal exists. Coverage is listed per dimension.
Customization
What it measures.
Whether you configure your AI environment — instruction files (CLAUDE.md, AGENTS.md), MCP configs, custom subagents, and custom slash commands.
Why it matters.
The biggest leap from "user" to "operator" is when you start shaping the tool to your work, not the other way around.
How it's measured.
Two ways, depending on which tool and which plugin version produced the data.
Claude Code and Cowork, plugin 0.3.26 and later
— mostly a state measurement rather than an activity one. The plugin counts the configuration surfaces present on your machine (config_surfaces_built)
and the subset it saw you exercise in the window (config_surfaces_used), then scores breadth — log-normalized to a reference of 100 — scaled by that liveness ratio. Config you no longer touch keeps half credit rather than dropping to zero. A quarter of the dimension is still a bonus for editing config inside the rolling 30 days. Only the two counts leave your machine, never a path, project name, or file body.
Everything else
— Codex, Cursor, OpenCode, and any device on an older plugin — scores the original way: the sum of
custom_mcp_config_writes
and claude_md_writes
(or agents_md_writes
for Codex) over the rolling 30 days, log-normalized to a reference of 10.
Until 26 October 2026
Claude Code and Cowork take whichever of the two is higher, so updating your plugin can only help. From that date the write count retires for those two tools and only the surface measurement counts — a device that has not updated reports no surfaces and keeps just the write-count portion, so its Customization drops. Codex, Cursor, and OpenCode are not affected by that date and keep the write count, because no plugin version lets them report configuration yet.
Skill file authorship is tracked separately under Custom Skills.
How to grow it.
Add a CLAUDE.md to your project that captures your conventions. Add one MCP server you'd use weekly. On Claude Code and Cowork with a current plugin, a setup you built months ago still counts, as long as you are still working in it.
Sources: Claude · Codex · Cowork · OpenCode · Cursor · Pi · Hermes · OpenClaw · NanoClaw
Parallel Agents
What it measures.
Whether you're running multiple agents in parallel within a session or creating
isolated worktrees for cross-process fan-out.
Why it matters.
Parallel orchestration is how senior operators compress hours into minutes — branching work across subagents and converging the results.
How it's measured.
Weighted blend of breadth (sessions with orchestration, log-normalized to a reference of 50), depth (parallel agent turns), and peak (max parallel agents in any session). Each signal uses its own role-scaled normalization. A small capped bonus is added when the
scanner observes git worktree add
commands. Worktree creation is treated
as supporting evidence, not proof that another agent ran.
How to grow it.
Use the Agent tool to fan out independent subtasks. When two pieces of work don't depend on each other, launch them at the same time.
Sources: Claude · Codex · Cowork · OpenCode · Cursor · Pi · Hermes · OpenClaw · NanoClaw
Background Work
(was Context Leverage)
What it measures.
Whether you delegate work to run unattended — Cron, Task, or Monitor primitives.
Why it matters.
Senior operators don't sit and wait. They queue work, walk away, and come back to results.
How it's measured.
Sessions in which you used context-leverage tools (TaskCreate, ScheduleWakeup, CronCreate, Monitor, etc.), log-normalized to a reference of 80.
How to grow it.
The next time a subtask will take a while, schedule it via Task and switch contexts. Use Cron for nightly jobs you'd otherwise babysit.
Sources: Claude
Tool Breadth
What it measures. How broadly you engage with skills and MCP servers.
Why it matters.
Skills and MCPs are how Claude Code reaches outside the conversation — into your tools, services, and data.
How it's measured.
Distinct count of skill_counts
+ mcp_server_counts
in the 30-day window, log-normalized to a reference of 100.
How to grow it.
Install one MCP server you'd actually use weekly. Try a built-in skill you've been ignoring.
Sources: Claude · Codex · Cowork · OpenCode · Cursor · Pi · Hermes · OpenClaw · NanoClaw
Planning
What it measures.
Structural follow-through on planning — a planning workflow followed by successful implementation, or a durable plan artifact — not every way you might plan.
Why it matters.
Plans cut rework. Power users front-load thinking and then act on it.
How it's measured.
A session earns credit when a planning signal — Claude's Plan
agent, ExitPlanMode, Codex's native Plan Mode, or update_plan
— is followed by a successful Write, Edit, or apply-patch in the same session and local date, or when the session writes a recognized plan artifact: a Markdown file under any
plans/
directory at any depth (including .claude/plans/), or a
PLAN.md
or *-plan.md
file. Qualifying sessions are log-normalized to a reference of 10. We never read plan content or judge plan quality, and we never upload artifact paths or transcript text — only the structural sequence is checked locally. Conversational planning without follow-through or an artifact is intentionally undercounted.
How to grow it.
Follow every plan with implementation in the same session, or write it down as a durable artifact under any
plans/
directory (including .claude/plans/), or use PLAN.md
or *-plan.md.
Sources: Claude · Codex
Repetition NEW
What it measures.
When you use a skill, do you go deep — or just try once?
Why it matters.
Repeat usage signals that a skill has become real practice, not a one-off experiment.
How it's measured.
Total invocations across skill_counts
in the 30-day window, log-normalized to a reference of 20.
How to grow it.
Pick one or two skills and use them repeatedly across a week, instead of trying twenty different things once each.
Sources: Claude · Codex · Cowork · OpenCode · Pi · Hermes · OpenClaw · NanoClaw
Custom Skills NEW
What it measures.
How many of your own custom skills you've authored and actively practice.
Why it matters.
Authoring a skill is one thing; making it part of your daily flow is the real signal. Writing your own automation and then using it repeatedly shows both investment and payoff.
How it's measured.
Blend of two signals, log-normalized to a reference of 8: the set of skill names you authored (detected via Write/Edit/apply_patch calls on
~/.claude/skills/<name>/...
or ~/.codex/skills/<name>/...
paths in your transcripts) intersected with your skill_counts
keys, plus the summed practice counts for those authored skills.
Namespace suffix-match applies: a plugin-installed skill named
<plugin>:<name>
still credits you if <name>
matches a skill you authored locally and shipped back.
How to grow it.
Author a skill that automates a recurring task. Use it in real workflows. Share it back to your company plugin — the namespace suffix-match means you still get credit. Author another.
Sources: Claude · Codex · Cowork · OpenCode · Pi · Hermes · OpenClaw · NanoClaw
Multi-Tasking NEW
What it measures.
How many concurrent sessions you run at peak — multi-tasking across distinct workstreams.
Why it matters.
Operators who run several workstreams at once are squeezing more value from the tool than serial users.
How it's measured.
Peak max_concurrent_sessions
over the 30-day window, log-normalized to a reference of 15.
How to grow it.
Try Cowork or run two terminals. The next time you're blocked on one task, start the next.
Sources: Claude · Codex · Cowork · OpenCode · Cursor · Pi · Hermes · OpenClaw · NanoClaw
Tracked but no longer scored
We removed these from the score because they were too easily saturated, didn't separate users in practice, or measured the wrong thing. We still capture the raw data so we can revisit if the user base diversifies.
-
Prompt Efficiency — was constant 100 across users.
-
Tool Diversity — saturated at the top of the cohort.
-
Conversation Depth — saturated similarly.
-
Basic Chat — measured the floor, not the ceiling.
