Agent Skills Need A Job Spec Before They Need A Model.

The useful idea inside Agent Skills is not that a company can buy a smarter button. It is that repeatable work can be packaged as instructions, resources, scripts, and checks that an AI agent loads when the task calls for them. For a building-industry business, that is only valuable if the skill begins with a job spec.

OpenAI's current Codex skill documentation frames a skill as a task-specific package that helps an agent follow a workflow reliably. The open Agent Skills standard describes the same pattern as a folder with instructions plus optional scripts, references, templates, and assets. OpenAI's recent agent work also points to longer, more cross-functional delegated tasks. That combination is powerful, but it makes sloppy setup more expensive.

The job spec comes first

Before a remodeler, builder, designer, supplier, distributor, showroom, or trade contractor creates an agent skill, the team should name the real job. Not "help with estimating." Name the operational task: compare three vendor quotes against the signed scope, draft a missing-selections follow-up, flag contract allowances that changed, turn a site note into a client update, or prepare a weekly purchasing exception list.

A skill without that job boundary becomes generic assistant behavior. A skill with the boundary can be tested. It has approved sources, a trigger, a format, a reviewer, known failure modes, and a log of what happened.

A practical skill card

Datum's version of an agent skill starts as a plain operational card before it becomes automation. The card should be short enough for an owner or manager to inspect, but specific enough for an AI worker to run the same way twice.

• Job: the exact recurring task the skill owns.
• Trigger: the user action or system event that starts the workflow.
• Sources: the approved files, emails, CRM notes, drawings, quotes, catalogs, job-cost exports, or web pages the skill may use.
• Inputs and outputs: typed fields in, reviewable artifact out.
• Permissions: what the skill may read, what it may change, and what requires human approval.
• Failure path: what happens when sources are missing, conflicting, stale, or low confidence.
• Receipt: prompt, source list, tool calls, output, edits, approval status, errors, and final version.

This is the difference between operational AI and automation theater. The first creates a reusable work surface. The second creates a demo that looks impressive until it meets a messy project folder.

Search guidance points the same direction

Google's current guidance for AI Overviews and AI Mode says the fundamentals still matter: helpful original content, crawlability, technical clarity, and structured data that supports normal Search features. Google also says there is no special schema.org markup required for generative AI Search, and its documentation updates confirm FAQ rich results are being removed rather than becoming a new optimization lever.

That matters because the same discipline should govern public content and internal AI workflows. A building business should not create hidden AI-only pages, fake source signals, or unsupported schema. It should publish useful proof and build internal agent skills that can show their sources, limits, and receipts.

Where building teams should start

Start with one painful, repeatable job where the source material already exists. A quote comparison skill is a good example. The sources are the signed scope, vendor quotes, product specs, exclusions, allowance notes, and project constraints. The output is not a final decision. It is a comparison table, exception list, missing-information list, and recommended review questions.

The human still owns the decision. The AI skill owns the structured first pass, the trace, and the reminder that no source means no confident claim.

The Datum rule

Do not ask whether an agent skill is impressive. Ask whether it can survive a normal week inside the business. Can the office manager see the status? Can the project manager inspect the sources? Can the owner tell which output was approved? Can the team rerun it next Friday? Can a mistake be traced to a bad source, bad instruction, bad tool call, or bad review habit?

For the building industry, agent skills should be narrow workers, not vague coworkers. Give them a job spec, a source set, a permission boundary, a review path, and a receipt. Then decide whether the skill deserves more automation.

Next step, if this note maps to a problem on your desk: Private Training — a private working session for your team ($1,500+).