hse-agent

What the toolkit is

A skill is a reusable set of instructions for a particular task. In hse-agent, each skill describes when it applies, the context it needs, the artifact it should produce, and the checks a practitioner should make before using the result.

My contribution is how these practices are organized into task instructions, guidance for choosing skills, and review expectations. Context profiles are short descriptions of your field, users, working constraints, and design system. The AI assistant uses those instructions and the source material you supply to draft its response.

The toolkit is intended for practitioners who already bring context and judgment to the work. It does not provide access to research, operational data, or connected systems unless those inputs are separately available in the environment.

How hse-agent, PEB, and BDX fit together

Behavior-Driven Experience (BDX) provides a way to investigate behavior and identify experience improvements. The Product Experience Blueprint (PEB) connects evidence about people and work to coherent product decisions, scope, implementation, and validation. hse-agent contains skills that help an AI assistant apply these approaches and produce related artifacts.

You can use PEB and BDX directly with a team or use the toolkit to help prepare their artifacts. In either case, review the evidence and reasoning before treating the output as product guidance.

Replace my mining context with your own

My experiments are grounded in mining and operational technology. My workplace toolkit therefore includes a context-adaptation skill for roles, shift handoffs, field conditions, site variation, operational risk, and subject-matter review. That context helps in my work; it should not become an assumption about yours.

Replace that layer with a description of your own users, tasks, decisions, vocabulary, working conditions, data definitions, and authority. Identify the domain experts who can resolve gaps. Keep research synthesis, workflow analysis, blueprinting, backlog development, and validation skills available alongside it.

My workplace edition also includes an MDS 2.0 Flutter skill. The public edition replaces those company design-system references with a profile for your approved components, tokens, accessibility requirements, and delivery conventions. Review examples in the other skills for assumptions that may also need adaptation.

A practical workflow

1. Choose the task. State whether you need research synthesis, a workflow analysis, a blueprint, a backlog, or a validation plan, and name the decision the output should support.
2. Supply context. Provide relevant source material, users, goals, constraints, data availability, and any existing decisions. Identify which sources are authoritative and where information is incomplete.
3. Ask for traceable reasoning. Request separate observations, interpretations, assumptions, and recommendations. Require source references where available and explicit labels for unknowns.
4. Review the result. Check claims against the inputs, test whether the workflow makes sense, and resolve disagreements with the relevant practitioners or specialists.
5. Carry learning forward. Update the artifact and its linked requirements or backlog as prototype review, data investigation, and user research change the direction.

Inspect a skill in use

This authored demonstration uses the public kit’s fictional library equipment-loan task. The sample output and review decisions illustrate how to inspect a response; they are not a transcript of a model evaluation.

What to inspect in an output

Check evidence: can a claim be traced to supplied material, or has an assumption been presented as a finding? Check context: does the recommendation match users, authority, environment, and actual work? Check feasibility: does it depend on unavailable data or an unconfirmed capability?

Check behavior: are failure paths, recovery, accessibility, permissions, and human review addressed where relevant? Check usefulness: does the artifact help someone make a decision, or merely restate a framework? Record unsupported recommendations and correct the instructions or inputs as appropriate.

Connecting skills across a product workflow

Use skills together when the next decision needs a different kind of analysis. Research synthesis can organize observations into needs and opportunities; workflow analysis can clarify tasks, decisions, and handoffs. A Product Experience Blueprint then connects the supported findings to capabilities, scope, dependencies, and validation.

With that blueprint reviewed, the backlog skill can help define delivery work while preserving the user need and rationale. Carry source references, open questions, and review decisions forward so an assumption does not become a confirmed requirement simply because it appears in the next artifact.

Choose the skills needed for the task rather than running the whole library. At each handoff, review the output with the relevant practitioners and specialists. As implementation or research changes the evidence, update the blueprint and linked backlog together.

Human responsibility and known limits

I treat generated artifacts as drafts to inspect against the source material. AI can omit context or invent a connection even when the instructions ask for evidence. Review recommendations with practitioners and specialists who understand the work and its consequences.

I tested repository-scoped skill discovery and four bounded tasks in Codex, plus custom-agent discovery and blueprint/backlog tasks in VS Code Copilot. These smoke tests checked source use, uncertainty, authority, recovery, and task scope. They do not establish general reliability or time savings. Evaluate representative tasks in your own environment, retain failures as well as successes, and record the human decisions that shaped the final artifact.

Using the guidance across tools

I developed this guidance for Codex, the Snowflake-based Prospector CoWork environment used at Freeport-McMoRan, and GitHub Copilot. The public kit provides portable Markdown and separate Codex and Copilot adapter guides. Choose the guide for your environment and verify how it discovers or receives the instructions.

Once installed and discovered in Codex, skills can be selected automatically when a request matches their descriptions. You can describe your task naturally or name a skill when you want a particular method. Verify both approaches during setup; automatic selection depends on the request and available skills. Other environments may require choosing an agent or supplying the instructions explicitly.

Models, permissions, connected sources, and tool access differ across environments. The toolkit supplies instructions; your environment supplies those capabilities. Run the included fictional exercise in the environment you intend to use and review the result before applying it to project work.

Try a bounded task

Begin with the fictional task included in the download. Review its claims and handoffs against the source notes, then repeat the exercise with a small set of notes from your own work.

A useful first result is an artifact you can explain and check: what the sources support, what remains uncertain, and what decision comes next. Carry that reviewed artifact into a blueprint or backlog when the task calls for it.

What the public edition contains

The download contains 19 general-practice skills and two replaceable context skills, along with a task router, adapter guides, a fictional exercise, and review criteria. The router helps you choose the smallest relevant set of skills for the task.

The library below describes each capability. Read the selected instructions, add your context and sources, and use the review criteria to inspect the resulting artifact. The public edition is a separate adaptation of my workplace toolkit.