Harness Useful Skills: a workflow written as skills

Hello there!

In September I wrote about Oh My Pi Anvil, where TypeScript owns every transition between planning, implementation, checks and reviews. Harness Useful Skills takes the opposite approach. It has no workflow executor, model router, phase database or stop-hook loop. It is a set of owned us-* skills: the host discovers their descriptions, and the model follows the relevant instructions with the host’s own tools and workers.

The goal is the same, getting from a natural-language request to reviewed, verified work. What changes is where the process lives. Here it lives in instructions that the agent reads, not in code that drives the agent.

From a request to reviewed work

You ask naturally, for example “I want to build a rocket simulator.” Before choosing any stages, us-workflow inspects the current source, relevant callers, acceptance criteria, authorization and risk, then briefly explains which lane it picked.

New or material behavior, architecture changes, uncertain scope and security-sensitive work use the normal lane. It runs research, has the bundled planner draft a plan, and asks for one approval. Implementation then goes to classified workers: frontend, backend, or a general task for work that is genuinely neither. Fresh correctness and security reviews follow. Findings return to repair within the approved scope, and three unsuccessful repair rounds stop the loop with the work kept and the blocker reported.

A clear, bounded, low-risk correction can take the focused repair lane instead. It still requires current source inspection, an authorized scoped fix, verification of the changed path and a truthful report, but it does not restart the whole cycle. Size, urgency or “it was already delivered” alone does not qualify a change for this lane. Skipped stages are reported as skipped, never as complete.

Settings you can switch off

The workflow can be disabled per repository with /useful-skills workflow disabled. That gives a fast lane that still inspects the code, makes the change, exercises it and reports the evidence, but skips the mandatory stages. Six profile-wide switches control the individual stages: research, plan, review, security review, and the two memory stages. For a single request, invoking us-ignore-workflow directly bypasses the stages without changing any saved setting. Quoted or copied text does not count as that invocation.

Besides the development flow, there are focused skills for read-only security, performance and code-quality audits, architecture and test-coverage improvements, and delivery. us-open-pr, us-approve-work and us-ship-backlog-item deliver through a feature branch and a pull request. They need explicit authorization: plan approval does not authorize publication, and none of them pushes directly to the base branch or merges a pull request. A backlog item is marked Done only after its pull request is confirmed merged.

Two kinds of memory

Memory is split on purpose. Facts explain decisions and lessons. A code graph built with Graphify maps relationships between files and symbols. Neither replaces the other. In OMP, one us_memory tool covers both: facts go to OMP’s native memory backend, with Mnemopi recommended for local SQLite storage, and the graph is stored per project outside the checkout.

Graphify setup is lazy. Nothing is installed or built when a session starts, only on an explicit build or query. Its managed dependencies are pinned and checksum-verified, and builds run a scope guard first: a workspace that is /, is or contains a home directory, or contains nested mounts is refused. Fact saves that look like secrets are rejected before they are stored.

Claude Code too

The same repository also works as a Claude Code marketplace, so one project covers both hosts. There, Claude gets 17 namespaced skills such as /useful-skills:us-workflow and seven native agents, from planner to security-reviewer. The models for those roles come from ~/.claude/useful-skills/models.yml. On each session start, a hook generates us-* agents with the configured model and effort, so every role can run on the model that suits it.

Memory is separate in Claude Code: a local MCP server provides memory_status, memory_search, memory_save, graph_build and graph_query. Its facts are stored as private project files. /useful-skills also opens a menu for settings, status and the doctor, which can prepare memory for the project.

Give it a try

With Oh My Pi and Node.js 22 or newer, run these commands in your terminal:

omp plugin marketplace add MSpiechowicz/harness-useful-skills
omp plugin install harness-useful-skills@harness-useful-skills --scope user

For Claude Code, add the same repository as a marketplace and install the plugin, whose ID stays useful-skills@useful-skills-local:

claude plugin marketplace add MSpiechowicz/harness-useful-skills
claude plugin install useful-skills@useful-skills-local --scope user

Restart the host after installation, and that is it. I have tested it on macOS with both OMP and Claude Code. The installation guide also covers project scope and local development.

The project is available under the GPL-3.0 license, and its skills draw on reviewed prompt practices from several open-source projects credited in the repository. If you try it, I would like to hear how the workflow fits your repositories. Issues and contributions are welcome on GitHub.

That is all for now. Stay healthy and keep coding!

Greetings,

Maciej