better agentic engineering.
The agent-plugin edition of Microsoft's AI-Engineering-Coach.
Analyzes how you actually work with AI coding agents — every tool you use, one report.
Your AI coding tools already write a detailed log of every session to your disk. This plugin reads those logs and tells you how you are actually working: where your prompts are weak, which habits cost you time, how much code the agent writes versus you, and whether your repos are set up for agents at all.
Ask your agent "how am I using AI?" and it answers from your own data. Nothing is uploaded.
It reads Claude Code, Codex, OpenCode, GitHub Copilot (VS Code, CLI, and the desktop app), and Copilot for Xcode.
This is a fork of microsoft/AI-Engineering-Coach, repackaged as an agent plugin. The original is a VS Code extension; this one is not, and its AI-powered features run through Claude Code. See Relationship to upstream.
Add the marketplace, then install the plugin:
/plugin marketplace add hardkoded/ai-engineering-coach-plugin
/plugin install ai-engineering-coach@ai-engineering-coach-plugin
Then build it once, so the coach has something to run:
cd ~/.claude/plugins/marketplaces/ai-engineering-coach-plugin
npm ci && npm run buildThe skill also builds itself on first use if you skip that step — the first report just takes a few minutes longer.
That is it. Ask "how am I using AI?" and the skill activates.
If you would rather not use the plugin system, clone the repo and link the skill:
git clone https://github.com/hardkoded/ai-engineering-coach-plugin.git
cd ai-engineering-coach-plugin
npm ci && npm run build
ln -s "$PWD/skills/ai-engineering-coach" ~/.claude/skills/ai-engineering-coachLink the skill directory, not the files inside it. The launcher finds the plugin root from its own physical location, so a symlinked directory works and a copied script does not.
The repo is also an Agent Plugins package — plugin.json
at the root, Agent Skills under skills/ —
so any client implementing that standard can load it and discover the same skills. Clients
that read skills from a checkout will find them in .claude/skills/ and
.github/skills/, which symlink to the same files.
The generative features need Claude. Skill Finder, the Learning Center and the AI context
review call the claude binary you have already signed in to — there is no API key to set up.
Without claude on your PATH those features hide themselves; everything computed from your
logs still works.
Ask in plain language — the skill description covers these triggers:
how am I using AI? review my last month of Claude Code sessions what are my worst anti-patterns? is this repo set up properly for agents?
Your agent runs the CLI and reads the result. You can also run it yourself:
# a summary you can read
node dist/cli.cjs report
# just your Claude Code sessions, last 30 days
node dist/cli.cjs report --since 30d --harness Claude
# structured output
node dist/cli.cjs report --json | jq .| Flag | Effect |
|---|---|
--since <n>d |
Only the last n days |
--harness <name> |
One tool: Claude, Codex, OpenCode, Local Agent, Xcode, GitHub Copilot CLI, GitHub Copilot App |
--workspace <id> |
A single workspace |
--logs-dir <path> |
An extra VS Code or Xcode log directory (repeatable) |
stdout is the document and nothing else — progress and parse diagnostics go to stderr, so
piping into jq works.
For exploring rather than reading, the same data renders as a local web app:
node dist/cli.cjsIt prints a http://127.0.0.1:<port> URL and opens your browser. --no-open skips that,
--port <n> pins the port.
Screenshots
Session parsing runs in a forked process with its own heap, so a multi-gigabyte
~/.claude/projects will not exhaust memory. The first run takes a couple of minutes; logs
from terminal agents are re-parsed each time rather than cached.
| Skill | What it does |
|---|---|
ai-engineering-coach |
Reports on your sessions — scores, anti-patterns, code output, activity |
update-docs |
Adds or edits a page in the Hugo docs site |
npm test validates plugin.json and every SKILL.md against both specs
(src/plugin/manifest.ts), so a malformed skill fails CI rather
than failing silently in someone's agent.
| Page | Description |
|---|---|
| Dashboard | Practice scores with week-over-week trends, daily activity chart, top workspace stats |
| Timeline | Gantt-style session timeline with per-day drill-down and overlap detection |
| Coding Moments | Screenshot gallery from AI coding sessions with story reels and workspace filtering |
| Page | Description |
|---|---|
| Output | Generated code volume by language, model usage table (token breakdown temporarily hidden) |
| Burndown | Monthly AI token budget progress with projections (temporarily disabled) |
| Patterns | 7×24 activity heatmap and work-life balance signals |
This section appears only when the local GitHub Copilot app is installed.
| Page | Description |
|---|---|
| Productivity | Project sessions with and without issues, pull request and merge conversion, and a seven-day PR merge-ratio trend |
| Issue credits | Rough relative AI spend per GitHub issue across linked workspace, alias, creator, and coordinating sessions |
Issue credit percentages are rough relative estimates, not accurate AI Credit or billing figures.
They normalize the locally recorded total_nano_aiu usage across linked issues to show approximately
where AI usage was spent. The percentages depend on issue links inferred from local workspace and
issue-reference data and must not be used for billing reconciliation. Explicit GitHub issue URLs
pasted in either of the first two session turns also establish the issue link; URLs pasted later are
excluded to avoid treating research links as the session's source issue. All reconciliation and
aggregation is local and read-only. The small organization avatars on this page are loaded directly
from GitHub using the repository owner name; no session content or usage data is included in these
image requests.
| Page | Description |
|---|---|
| Anti-Patterns | Five practice score cards with severity ratings, concrete actions, and example prompts. 45 editable markdown rules plus a coverage heatmap |
| Rule Editor | Create, edit, and tune detection rules visually or as raw markdown. Live-test against your data |
| Rule Playground | Interactive REPL for the rule DSL with field browser, function catalog, and metric list |
| Data Explorer | Browse session fields, view distributions, run ad-hoc filters |
| Skill Finder | Discover repeated prompt patterns and matching community skills from the open-source catalog |
| Context Health | Overall context score, agentic readiness checklist, workspace context map, AI-powered instruction-file review |
| Page | Description |
|---|---|
| Learning Center | Personalized quizzes and code-comparison rounds generated from your actual usage |
| Achievements | XP-based progression with Bronze → Silver → Gold → Diamond tiers |
| Agentic SDLC | How you use AI across the full software-development lifecycle |
| Share | Generate a shareable stat card and export Markdown/JSON summaries |
Skill Finder, Learning Center, and the AI review on Context Health call Claude, and
hide themselves when the claude binary is not on your PATH. Rule Playground and
Data Explorer have no sidebar entry yet — reach them from the Anti-Patterns page.
- Read-only — never modifies your session files
- Local analysis — all parsing and analytics run on your machine
- No telemetry — nothing phones home, and the core analysis paths make no network calls
- Optional AI features — Skill Finder, Learning Center and the context review shell out to
your own
claudebinary, and only when you explicitly invoke them. Everything else runs with no model at all.
Forked from microsoft/AI-Engineering-Coach, an open-source community effort by Microsoft employees. The analysis engine, rules, and dashboard are theirs. This fork adds:
- A standalone CLI, so the dashboard and a stdout report run from a terminal.
- Agent Plugins and Agent Skills packaging, so any compatible agent can use it.
- The generative features rewired from the VS Code language model to the Claude CLI. This fork drops the VS Code extension entirely; upstream is where to go if you want that.
- Fixes for analyzing terminal agents: Claude Code workspaces were dropped from Context Health entirely, and eleven IDE-only rules scored terminal sessions against fields those tools never record.
Upstream is the place for issues about the analysis itself. MIT licensed, like the original.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.
This is an independent fork, maintained by @kblok under the hardkoded organization, and not affiliated with or endorsed by Microsoft. The upstream project it derives from is an open-source community effort by Microsoft employees, itself not an official Microsoft product and not part of any Microsoft service or support offering. Neither this fork nor the original carries any warranty or guarantee; both are provided as-is.