refactor(prompts): apply Claude 5 context-engineering rules to the commands

[rlorenzo]

Summary

Anthropic's new rules of context engineering reports removing ~80% of Claude Code's own system prompt with no performance loss: current-generation models follow intent better than procedure. This reworks the command sources on that basis — net -69 lines across the six commands, with the trigger conditions each one was missing added to its description.

Changes

• Triggers moved into description. It's the only text a model sees before deciding to load a skill, so each one now says when to reach for it — and explicitly disambiguates the three overlapping reviewers: code-review (staged, read-only) vs. code-refinement (staged, edits code) vs. review-pr (an open GitHub PR).
• Removed self-restating body lines in review-pr, commitmsg, and efficient-orchestration, plus the When to Use This Skill section in dependency-review — all redundant with the descriptions.
• code-review: replaced the 5-step diff recipe (hunk-header parsing, @@ -old,+new @@ arithmetic) with the two gotchas it actually encoded — cite line numbers from the + side, and an empty diff with staged files means re-run without --unified=0.
• dependency-review: dropped "calculate the days between publish date and today" and the semver primer, keeping the non-obvious signal (breaking changes in a minor/patch release are a red flag on their own).
• README: documents the house style under a new "Prompt style" section so new commands follow it.

Deliberately untouched

• Hard constraints the review loops depend on — the NO_FURTHER_FEEDBACK sentinel, the reviewer never writing source files, the gh api --slurp/--jq conflict.
• Rubrics and output templates: the blog classes these as references the model fills in, not rules that constrain it.

Verification

• tools/generate --check — in sync (all 5 harness copies regenerated from the sources)
• test/run — 47/47 passing
• markdownlint-cli2 on the README and command sources — clean

🤖 Generated with Claude Code

https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk

---

Generated by Claude Code

[greptile-apps]

Greptile Summary

The PR streamlines six command prompts around intent-focused context engineering.

• Adds invocation triggers and sibling-command disambiguation to command descriptions.
• Condenses procedural guidance while retaining command-specific constraints, rubrics, and output templates.
• Regenerates the corresponding skills and shared prompts for all supported harnesses.
• Documents the prompt-writing conventions in the README.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete changed-code defect identified.

The command sources and generated artifacts remain synchronized, and the removed text is redundant with retained scope, outcome, and output requirements.

Important Files Changed

Filename	Overview
.claude/commands/code-review.md	Condenses staged-diff collection guidance while preserving staged-only scope, output constraints, and review policy.
.claude/commands/dependency-review.md	Moves activation criteria into the description and removes redundant procedural explanations without changing the five review checks.
.claude/commands/efficient-orchestration.md	Adds explicit activation and exclusion criteria while retaining orchestration tiers, delegation rules, and guardrails.
.claude/commands/review-pr.md	Moves the workflow trigger into front matter while preserving the review loop and its hard operational constraints.
prompts/code-review.md	Keeps the generated shared review prompt synchronized with the authoritative code-review command.
README.md	Adds house guidance for concise, intent-oriented command prompts.

_{Reviews (1): Last reviewed commit: "refactor(prompts): apply Claude 5 contex..." | Re-trigger Greptile}