Preporato
Build & submit taskBetaintermediate

Engineer a Claude Code Project Config and CI Gate

Configure Claude Code for a real project: a layered CLAUDE.md, a path-scoped rule with a glob, a Skill with frontmatter, and a headless CI gate that runs claude -p with structured JSON output and fails the build on a bad result. Submit a single file for instant, rubric-based feedback.

2.5 hrs

Est. time

4

Outcomes

7

Rubric criteria

65%

Pass score

What you'll learn

Skills you'll have real reps in after shipping this.

The CLAUDE.md hierarchy
User, project, and directory CLAUDE.md files layer together. Project-level lives in the repo so the whole team shares it.
Path-scoped rules
Rules under .claude/rules/ with glob patterns apply extra instructions only to the paths that need them.
Headless CI integration
claude -p with --output-format json gives the pipeline a parseable result it can gate the build on.
Plan vs Direct, and review
Plan Mode suits risky or wide-ranging changes; direct execution suits small scoped ones. An independent review session catches what self-review misses.

The scenario

Your team adopted Claude Code, but everyone configures it differently. Some rules live in one person's user settings, the project has no CLAUDE.md, and a teammate wired Claude into CI by piping its raw text output into grep. Reviews are inconsistent and the agent keeps ignoring conventions that were only ever written in a Slack message.

You are going to standardize it. A project-level CLAUDE.md, a path-scoped rule for the directory that needs special handling, a reusable Skill, and a CI step that runs Claude headlessly and gates the build on structured output the pipeline can actually parse.

Your role

You are a Claude solutions architect setting the team standard for Claude Code. Your deliverable is one file that contains the project configuration (CLAUDE.md, a rule, a Skill) and a runnable CI gate, with the configuration decisions made explicit.

Start the task to unlock the full brief

You'll get the step-by-step requirements, setup commands, the 7-criterion grading rubric, tips, and the ability to submit your solution for instant AI grading.

Free to start · submit when you're ready

What you'll build in this Claude Code workflow task

This is a build-and-submit task, not a guided lab. You set the team standard for Claude Code on a real project: a layered CLAUDE.md, a path-scoped rule, a reusable Skill, and a CI gate that runs Claude headlessly and blocks the build on a parseable result. The deliverable is one file a team could adopt as its configuration baseline.

The skills here are the ones that make an AI-assisted codebase consistent instead of chaotic. You put shared conventions in project-level config so the whole team gets them, scope special instructions to the paths that need them, package a repeatable workflow as a Skill, and integrate Claude into CI with structured output rather than scraping text. You also make the judgment calls explicit: Plan Mode versus direct execution, independent review versus self-review, and team versus user configuration.

Grading is rubric-based and explainable. Your submission is scored against weighted criteria (CLAUDE.md and hierarchy, path-scoped rules, the Skill, the headless CI gate, Plan-vs-Direct, the review pattern, and team-vs-user config) with per-criterion feedback quoted from your file. The pass threshold is 65 percent and you can resubmit. These are the Claude Code configuration skills the Claude Certified Architect exam tests.

Frequently asked questions

Do I need to actually run Claude Code in CI?

No. Grading is review-only, so you write the CI gate as if it runs (claude -p with --output-format json, parsed, exiting non-zero on failure). You are graded on the configuration and the gate being correct and concrete.

Can I submit a Markdown bundle instead of a script?

Yes. A single .md that contains the CLAUDE.md, the rule, the Skill, and the CI gate (as a fenced script block) is accepted. A .py script that contains the config as strings and a runnable gate is equally fine.

What goes in project config versus user config?

Anything the whole team should share (the project CLAUDE.md, .claude/rules, shared Skills) is committed to the repo at project level. Personal preferences stay in user-level config and are not committed. The task asks you to make that split explicit.

What counts as a complete submission?

A single file containing a project CLAUDE.md (with the hierarchy noted), a glob-scoped rule, a Skill with frontmatter, a headless claude -p CI gate with JSON output, and stated criteria for Plan-vs-Direct, the review pattern, and team-vs-user config.