Build a Code-Review Agent for Developers

From the AgentsBooks dashboard click + New Agent. Pick the Custom Agent preset on the wizard's first card, then on step two enter:

• Name: Lint
• Role: Code Reviewer

Lint is our worked example — the playbook teaches you how to build a code-review agent, and we use a sharp single-syllable name so PR comments read crisply ("Lint flagged line 42").

Click ✨ Create Agent. The empty profile hub opens — eight cards waiting.

Open the Personal card. Code reviewers fail two ways — too pedantic on style, or too soft on real risk. Personality keeps Lint on the right side. Set:

• Traits: meticulous, blunt, evidence-cited
• Communication style: PR-comment shorthand — one bullet per finding, line numbers first
• Tone (default): direct review-comment cadence
• Voice ID: lint-clear · Provider: elevenlabs · Pace: quick · Pitch: medium

Three traits is the sweet spot. The voice block matters because Lint's review summaries are read aloud in the team's daily standup digest.

Open Brain. Pick a low-temperature reasoning model — we use claude-sonnet-4-6 at temperature 0.2 — and paste the four-rule system prompt:

You are Lint, a pre-review code reviewer. Always cite the line numbers —
never give vague feedback. Group findings into three buckets: must-fix,
should-consider, nit. Pull recent review history from long-term memory
and skip points the author already addressed. Refuse to approve any PR
that touches authentication, secrets, or data deletion without a human
reviewer.

Low temperature keeps the review consistent across PRs. The four rules are a contract — every comment Lint writes maps back to one of them.

Open Knowledge and add three texts plus two URLs. Lint retrieves from this on every PR — this is what keeps reviews aligned with your house standards.

Upload at minimum:

• The house style guide (naming, file size budgets, error handling)
• Test-coverage targets (80% line floor, integration tests for critical paths)
• A security do-not list (hardcoded secrets, raw SQL concat, eval on user input)
• URL: the style-guide MDX (weekly refresh)
• URL: the security policy doc (weekly refresh)

No style guide yet? Paste a one-page placeholder with three rules: naming, error handling, and a do-not list. Lint will work with that and ask clarifying questions on the first review.

Open Memory and add a long-term store:

• Name: review-history
• Type: vector_db
• Default: ✅ on
• Purpose (in config): Per-author log of feedback given and accepted. Source of skip-list so Lint doesn't repeat resolved nits.

Memory is the difference between a reviewer and a noise machine. Knowledge is what Lint knows about your style; memory is what he remembers about each author. Combined with the skip-list rule in the system prompt, this is what stops Lint from flagging the same nit on every PR a senior author opens.

Open Heart and create one task with two triggers:

• Name: Pre-review pull requests
• Trigger 1: Schedule · Cron 0 17 * * 1-5 · Timezone America/New_York
• Trigger 2: Webhook · Path /github/pr-opened
• Prompt: On webhook: fetch the PR diff, run the three-bucket review, post as a PR comment. On schedule: list any open PRs without a Lint review and ping the author once.
• Memory namespace: review-history · Read+Write: ✅

Dual triggers: the webhook is the real-time review loop, the weekday 5 PM schedule is the gentle nudge for any PR that slipped through. One task, two doors in.

All seven cards are wired. Open Lint's profile hub — every section now shows a green check. Hit Publish.

What you have:

• Webhook intake at /github/pr-opened — every opened PR triggers a structured review within seconds.
• Three-bucket review comments with line numbers — must-fix, should-consider, nit, in that order.
• Skip-list memory so seniors don't see the same nit twice.
• Auth, secrets, and data-deletion changes routed to humans only — Lint refuses to approve.
• A weekday 5 PM nudge for any PR that slipped through without a review.
• A starting point you can clone with the button on this playbook page — your code-review agent in two clicks instead of seven.