Seven independent entry points, from research questions to recurring maintenance

HumanLayer Skills

Make Engineering Work Easier to Review

Seven engineering skills from HumanLayer for research and design, visual explanations, clearer PRs, project instructions, tighter React props, and recurring maintenance built around a defined task.

By HumanLayer

HumanLayer Skills: Make Engineering Work Easier to Review project image

Entry points used here

  • Claude Code: install individually and invoke /skill-name
  • RPI in Codex: target codex during installation, then invoke $rpi
GitHub stars
5,504
Checked
Categories
Engineering

Start with research questions for one small change

Install only rpi in the target repository. Supply a bounded change and real source entry points, then review its saved request and questions before starting research in the same task directory. That gives you a concrete first result to inspect.

Pick a task and get started

Run the selected task’s installation command at the target project root, choosing that skill and Claude Code. Open the repository in a session, confirm the entry point and supporting files, then send its prompt. The first RPI task saves the request and questions; research follows only after review.

Choose a deliverable

Create a question set for a bounded change, then investigate the current behavior with file-level evidence.

  1. Install the skill for this task

    Run at the target project root and confirm only the selected skill for Claude Code. Without -g, installation is project-scoped. The pinned skills 1.7.4 installer requires Node.js 22.20+; skill content still comes from the repository at install time. Keep all supporting files and check any same-name overwrite prompt.

    npx [email protected] add humanlayer/skills --skill rpi -a claude-code
    Read the author’s setup guide

    Check the entry point and its reference directory

    For the first task, confirm that .claude/skills/rpi/SKILL.md and references/ are readable; the installer may link to a shared copy. Start a new session and find rpi in the / menu. Keep the whole directory: RPI’s processes, templates and five agent instructions live under references/. Installing one skill does not register all named agents from the Claude plugin. RPI includes fallbacks that read or pass along those instructions. If its document directory is blocked, choose a writable location or grant the required scope before continuing.

    Read RPI’s directory and host handling
  2. Give your AI this first task

    These seven prompts are editorial examples, not executed results. Replace the inputs with your own. The draft-only stopping points are explicit limits in these examples; they are not the default behavior of every skill.

    /rpi create research questions .agents/artifacts/task-retries; flow: tdd I want to improve retry behavior for failed background tasks. I will provide the real worker, task storage and test paths. Save the original request, do a lightweight code pass, and draft questions about retry counts, timing, persistence, restart recovery and existing test coverage. Do not design the solution yet. Complete only the questions phase. Use the next available NN- filenames and preserve exact source pointers and repository commits. State the chosen task directory. Ask if an existing directory is ambiguous; if writing is blocked, wait for my choice of access or location instead of silently moving it. Return the request file, question file and next research command, then pause. After approval I will run /rpi research .agents/artifacts/task-retries; flow: tdd. That phase should answer each question from code and tests, distinguish tests read from checks actually run, and retain open questions. Neither phase may edit application code, commit or push. End research with choices for a TDD, an outline or direct implementation; do not begin implementation automatically.

  3. Check what you receive

    • Separate request and question documents with clear numbering, source pointers and repository commits.
    • After approval, a current-system research document with evidence, testing coverage and unknowns; design and implementation have not begun.
    Read the seven original entry points
More about this skill

HumanLayer Skills contains seven skills you can install separately. RPI carries research, design and implementation between sessions through local documents. show-me selects a compact visual or code sketch to explain a question, while visual-pr turns a change into a reviewable PR. The remaining skills handle project instructions, React prop contracts, automation for a known maintenance job, and control loops that first need a measurement and selection policy.

RPI makes decisions inspectable. Research questions define what to investigate; research documents describe the existing system; design documents settle the intended change; an outline divides implementation into outcomes and checks. Each phase normally starts in a fresh context, although you can continue in the same session. TDD means technical design document here, not test-driven development. After research you can choose a design discussion, a PRD, a TDD, or skip design.

The entry points have different effects. show-me may return a call tree or save an HTML file. visual-pr can commit, push, and create or update a PR. The two automation skills write workflows that later invoke coding agents and produce PRs. Choose the deliverable and stopping point before choosing the command.

This page starts with one skill installed for Claude Code and gives a separate editorial example for each of the seven entry points. The author also documents a Codex invocation for RPI. The guidance comes from a pinned source review; the tasks and their expected benefits were not tested here.

Sources and verification

Sources checked

The pinned main skills, marketplace and plugin manifests, selected processes and helper code, installer documentation and official GitHub guidance were checked. This is not an installation, execution or publication test. Download and reading coverage are recorded separately.