> ## Documentation Index
> Fetch the complete documentation index at: https://docs.weaveos.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI skills

> Measure how effectively each engineer, team, and your whole organization works with AI coding tools

The AI skills report scores how well engineers work with AI coding agents in Claude Code, Cursor, GitHub Copilot, and Codex, not just how much they use them. Weave reads each captured session, rates seven behaviors such as planning, verification, and ownership from 1 to 10, and explains what raised or lowered each score and how to improve it. Those scores combine with three measures from pull request and cost data (percent AI output, output per dollar, and AI code turnover) into one **Overall AI skills score** from 0 to 10 for each engineer, team, and your organization.

<Note>
  AI skills is being rolled out progressively. If you don't see the AI skills
  section in your reports yet, reach out to us and we'll enable it for your
  organization.
</Note>

## Who can see it?

* The organization-wide AI skills report follows the same access as other organization reports: **Team lead**, **Manager**, and **Admin** roles.
* Team and individual AI skills views follow the [team](/reports/team) and [individual](/reports/individual) report access rules — members can see their own AI skills, managers and admins can see anyone's.
* AI skills is a paid feature. On free plans the page shows a preview behind a paywall.

## Where to find it

The same AI skills breakdown appears at three scopes, each reachable from its report's section navigation:

| Scope | Path | What it shows |
| - | - | - |
| Organization | **Reports → Organization → AI Skills** | Org-wide averages with cross-org benchmark overlays |
| Team | **Team report → AI Skills** | Team averages, compared against your org |
| Individual | **Individual report → AI Skills** | One engineer's skills, compared against team and org |

You can also get there from the **Overview** page: the AI skills radar in the health scores section links to the full report, and clicking a specific axis on that radar deep-links straight into that skill's drilldown.

All views respect the standard [filter bar](/features/filters) (date range, teams, repositories) and the time granularity selector.

## The ten skills

Each axis is normalized to 0–10. Hover any score box or radar axis label to see its description in the app.

### Measured from your code platform

These three axes come from PR data ([GitHub, GitLab, or Bitbucket](/code-platforms/overview)) combined with your [AI tool cost data](/metrics/ai-tool-costs):

| Skill | What it measures |
| - | - |
| **Percent AI output** | The share of [output](/metrics/code-output) attributable to AI tools ([AI output percentage](/metrics/ai-output-percentage)). 100% maps to 10. |
| **Output per dollar** | AI code output per dollar of AI spend, relative to your org average. A score of 5 means exactly the org average; above 5 means more efficient. |
| **AI code turnover** | [Code turnover](/metrics/code-turnover) weighted by AI output share, relative to your org average. Higher scores mean less AI-written code gets reworked shortly after merging. |

### Measured from AI conversations

These seven axes come from scored coding-agent conversations. Weave analyzes each session from the [Weave plugin](/tutorials/weave-plugin-installation) for Claude Code, Cursor, GitHub Copilot, or Codex and scores it 1–10 on every dimension; the report shows the average over the selected period.

| Skill | What it measures |
| - | - |
| **Planning** | How well the engineer scopes and plans work before delegating it to the AI |
| **Verification** | How thoroughly the engineer verifies the correctness of AI output — tests, review, evidence checks |
| **Task handling** | End-to-end task delegation effectiveness: managing context, blockers, and follow-through |
| **Code quality** | The quality of the code produced during AI interactions |
| **Collaboration** | The quality of human-AI teamwork across the session |
| **Ownership** | Engineer engagement versus abdication — higher means the engineer stays critically engaged rather than blindly accepting AI output |
| **Agent parallelization** | Use of subagents and parallel workstreams to speed up delivery |

<Note>
  The overall AI skills score only appears once at least one conversation-based
  skill has data. Without a Claude Code, Cursor, GitHub Copilot, or Codex plugin
  connection, the page shows a connect-your-data empty state instead of scores
  derived from zero conversations.
</Note>

## The overall score and score boxes

The headline **Overall AI skills score** (0–10) is a weighted average of the ten axes. By default all axes weigh equally; admins can retune the weights with the gear icon on the overall-score box (weights can be set organization-wide or per team).

Below it, one score box per skill shows that axis's current value. A box shows **—** when no data is flowing for that axis (for example, no cost data connected for output per dollar). Every box is clickable:

* The **seven conversation skills** open the conversation drilldown focused on that skill.
* **Percent AI output** opens the [AI output percentage](/metrics/ai-output-percentage) PR-level drilldown.
* **Output per dollar** opens a drilldown of PRs with their code output and attributed AI cost.
* **AI code turnover** opens the [code turnover](/metrics/code-turnover) drilldown.
* The **overall score** opens the drilldown across all scored conversations in the date range.

## Charts

The main row shows three charts side by side:

1. **AI output %** — the adoption trend over time
2. **AI skills radar** — all ten axes as a radar polygon, with comparison overlays
3. **AI skills** — the overall composite score trend. Click any point to open the conversations scored in that period.

### Radar comparison overlays

The radar draws dashed comparison lines that adapt to the scope you're viewing:

| Viewing | Overlays shown |
| - | - |
| Organization | **Median orgs** and **Top 10% orgs** (cross-org benchmarks) |
| A team | **Your org** and **Median orgs** |
| An individual | **Team** and **Your org** |

The cross-org lines respect the dashboard **Benchmarks** display toggles (Median and Top 10%), the same ones that govern [benchmark lines](/annotations/benchmarks) on other charts.

Below the chart row, the report also includes a per-skill trend chart for each conversation axis, so you can see how planning, verification, and the rest move over time.

## AI skills insights

On individual and team views, an **AI skills insights** panel sits above the charts. It synthesizes the period's AI sessions into a plain-English weekly summary plus actionable insight cards:

* A **week selector** switches between weeks (most recent shown by default).
* The week's **session count** is clickable and opens the conversation drilldown for that week.
* Each **insight card** shows a title, detail, and the skill it relates to. Cards tagged with a skill are clickable and open the by-skill drilldown scoped to that week, so you land directly in the sessions that drove the insight.

Insights appear once enough AI sessions have been analyzed for the person or team.

## Conversation drilldown

Clicking a conversation skill, the overall score, a composite chart point, or an insight card opens the conversation drilldown — a table of individual scored AI sessions. Every row is one conversation, with:

* **Title** — a short, auto-generated description of what the session was about
* **Author** and **Source** — who ran the session and in which tool (Claude Code, Cursor, or Codex)
* **Activity** — when the session took place
* **Scores** — when focused on one skill, that skill's score plus **Why this score** and **How to improve** explanations; when opened from the overall score, the **Overall** score, all skill scores, the full **Rationale**, and the **Strongest positive signal** / **Strongest negative signal**

The explanations are generated per conversation during scoring, so you can see exactly which behaviors in a session raised or lowered each score — and what to do differently.

### Exporting a conversation

Each drilldown row has a download button that exports the conversation as a JSON file containing all scores, the full rationale with per-dimension explanations, and the turn-by-turn transcript (user prompts, agent messages, and tool calls; very large tool inputs are summarized rather than included verbatim).

Access is enforced server-side with the same rules as viewing that member's data in the app.

## AI skills in the individual report table

On the [individual report](/reports/individual), the members table includes an expandable **AI skills** column showing each person's overall score. Expanding the group adds one column per axis, so you can sort and compare any specific skill across engineers.

## How scoring works

<Steps>
  <Step title="Sessions are captured">
    The [Weave plugin](/tutorials/weave-plugin-installation) for Claude Code,
    Cursor, GitHub Copilot, or Codex reports each coding session — prompts,
    agent messages, and tool calls.
  </Step>

  <Step title="Conversations are scored">
    Weave's scoring model analyzes each conversation and assigns a 1–10 score
    per dimension, along with a written rationale, per-dimension explanations,
    and improvement suggestions. New conversations are typically scored within
    about an hour.
  </Step>

  <Step title="Scores roll up">
    Session scores average into the per-skill trends and radar for the selected
    date range and scope, and combine with the PR-based axes into the weighted
    overall score.
  </Step>
</Steps>

## Prerequisites

<AccordionGroup>
  <Accordion title="For the conversation-based skills (most of the report)">
    Engineers need the [Weave plugin](/tutorials/weave-plugin-installation)
    for Claude Code, Cursor, GitHub Copilot, or Codex installed and reporting.
    These capture the per-session signals (prompts, agent messages, tool calls)
    that conversation scoring requires — org-level telemetry alone is not enough.

    For Codex, agent messages come from the Weave hooks the install command
    registers, and those hooks need a one-time trust approval on each machine
    before they run. See [hook trust prompt](/ai-tools/codex-plugin#hook-trust-prompt).
  </Accordion>

  <Accordion title="For percent AI output">
    A connected code platform ([GitHub](/code-platforms/github),
    [GitLab](/code-platforms/gitlab), or
    [Bitbucket](/code-platforms/bitbucket)) plus [AI tool
    attribution](/analysis/ai-output-attribution) data.
  </Accordion>

  <Accordion title="For output per dollar">
    AI spend data — auto-detected tool subscriptions, entries under
    **Settings → Financial**, or ingested cost telemetry. See [AI tool
    costs](/metrics/ai-tool-costs).
  </Accordion>

  <Accordion title="For AI code turnover">
    A connected code platform with enough merged PR history to compute [code
    turnover](/metrics/code-turnover).
  </Accordion>
</AccordionGroup>

## Improving your scores

1. **Plan before you delegate** — scope the task, state constraints and success criteria up front (raises **Planning** and **Task handling**)
2. **Verify AI output** — run the code, run the tests, review the diff instead of accepting on faith (raises **Verification** and **Code quality**)
3. **Stay engaged** — question suspicious output and steer the agent rather than rubber-stamping (raises **Ownership**)
4. **Iterate constructively** — give specific feedback when the AI misses, instead of restarting or fighting it (raises **Collaboration**)
5. **Parallelize where it fits** — use subagents for independent workstreams on larger tasks (raises **Agent parallelization**)
6. **Read your own drilldowns** — each scored conversation includes "Why this score" and "How to improve" written for that specific session

## Public API

Available from [`GET /metrics`](/api-reference/metrics/get-metrics) as `ai_skills`: the overall 0–10 composite score. Supports `group_by=account` and weekly granularity. This is not [`GET /skill-usage`](/api-reference/skills/get-skill-usage), which returns skill invocation counts.

## Common questions

### Which AI coding habits raise or lower a score?

Open a conversation skill's drilldown. Each scored session shows **Why this score** and **How to improve**, written for that session, so you can see which behaviors raised or lowered it. The [improving your scores](#improving-your-scores) list above maps common habits to the skills they move.

### Which AI coding tools does AI skills cover?

Sessions from the [Weave plugin](/tutorials/weave-plugin-installation) for Claude Code, Cursor, GitHub Copilot, or Codex, plus Claude Code session transcripts from the Compliance API on Claude Enterprise. Organization-level telemetry alone isn't enough, because scoring reads each session's prompts, agent messages, and tool calls.

### Does AI skills look at cost and output, or only conversations?

Both. Percent AI output, output per dollar, and AI code turnover come from pull request and AI cost data, and the other seven skills come from scored conversations.

### Can I compare our AI skills with other companies?

At the organization scope, the radar draws **Median orgs** and **Top 10% orgs** lines from cross-organization benchmarks. A team's radar shows **Your org** and **Median orgs**, and an individual's shows **Team** and **Your org**.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.