Prompts - Plannotator Docs

Plannotator’s review agents (Claude, Codex, and Code Tour) all shell out to an external CLI. This page maps what those CLIs receive on every invocation: which parts Plannotator controls, and which parts are owned by the CLI’s own agent harness.

Importantly, we don’t invent our own review prompts. The Claude review prompt is derived from Claude Code’s published open-source review prompt, and the Codex review prompt is copied verbatim from codex-rs/core/review_prompt.md. You get the same review behavior those tools ship with. Code Tour is the one exception: it’s a Plannotator-original workflow, so its prompt is ours.

The three layers

Every review call is shaped by three layers:

System prompt. Owned by the CLI (Claude Code or codex-rs). Plannotator never sets or touches this.
User message. What Plannotator sends. Always a single concatenated string of two parts: a static review prompt plus a dynamic user prompt.
Output schema. A JSON schema passed to the CLI as a flag, forcing the final assistant message to match a known shape.

What’s in the user message

The user message Plannotator sends is always:

<review prompt>

---

<user prompt>

Review prompt is a long, static review instruction that lives in the repo as a TypeScript constant. It’s distinct per provider.

User prompt is a short, dynamic line built per call from the diff type (uncommitted, staged, last-commit, branch, PR URL, and so on). The same builder is used for all providers.

Matrix

	Claude review	Codex review	Code Tour (Claude or Codex)
System prompt	Owned by `claude` CLI. We don’t touch it.	Owned by `codex` CLI. We don’t touch it.	Same as whichever engine runs.
Review prompt (static, ours)	`CLAUDE_REVIEW_PROMPT` in `packages/server/claude-review.ts`	`CODEX_REVIEW_SYSTEM_PROMPT` in `packages/server/codex-review.ts` (misnamed; it’s user content)	`TOUR_REVIEW_PROMPT` in `packages/server/tour-review.ts`
User prompt (dynamic, ours)	`buildCodexReviewUserMessage(patch, diffType, …)`	same function	same function
Full user message	`review prompt + "\n\n---\n\n" + user prompt`	same	same
Delivered via	stdin	last positional argv	stdin (Claude engine) or positional argv (Codex engine)
Output schema flag	`--json-schema <inline JSON>`	`--output-schema <file path>`	same as engine
Schema shape	severity findings (`important`, `nit`, `pre_existing`)	priority findings (P0 through P3)	stops plus QA checklist

Why the schema matters

The schema flag is a terminal constraint, not a per-turn one. The agent reasons freely across N turns, reading files, grepping, running tests, and only the final assistant message is forced to deserialize against the schema. Everything upstream is unconstrained exploration.

That’s why this pattern works for review. You get agentic exploration (the whole point of using Claude Code or Codex over a raw LLM call), plus a machine-readable payload the UI can render without any scraping.

What you can tune per job

From the Agents tab in the code-review UI, each provider exposes these settings:

Setting	Claude	Codex	Tour
Model	yes (`--model`)	yes (`-m`)	yes (per engine)
Reasoning effort	yes (`--effort`)	yes (`-c model_reasoning_effort=…`)	yes (per engine)
Fast mode	no	yes (`-c service_tier=fast`)	Codex engine only

None of these change the review prompt or user prompt. They only change how the underlying CLI executes the same user message.

Relationship to code review

See Code Review for the end-to-end flow this feeds into.