Skills, Not Prompts: Why to Build Portable AI Skills for Construction, Management and VET
Published 30 September 2026
A skill is not a personality. It is a versioned procedure an AI agent can find, load only when the job matches, and run with the same guardrails on Claude, ChatGPT, Copilot, Cursor, Gemini and more. For construction, project management and VET work, which already runs on checklists, templates and handoffs, that is the real prize.
This briefing covers what a skill is, why it is worth building, why building it once for any AI matters, and where to start in writing, people management, construction management and education.
What a skill actually is
A skill is a folder, not a persona. It holds a SKILL.md file (a name, a description and the instructions), plus optional references, assets and scripts folders. The agent sees only the name and description at start-up, loads the body when a task matches, and opens the deeper files only when a step needs them. This is called progressive disclosure, and it works the same way everywhere the standard is used.
Metadata
Only the name and description load, and they are always on. Think of it as the position description on the org chart.
Body
The full SKILL.md loads when the task matches, so you only pay for it when it is used. It is the ITP, SOP or work instruction.
Resources
References, templates and scripts load on demand, when a step needs them. Think spec book, form template and calculator.
Anthropic frames it well: building a skill is like putting together an onboarding guide for a new hire. Rather than building a separate, custom agent for every use case, you specialise one agent with composable capability.
Why build them
- Prompts are one-off; skills are create-once. Skills load on demand, so you do not repeat the same guidance in every conversation.
- The context cost drops. Dumping the whole company manual into every chat is expensive and noisy. Metadata stays cheap and the body is paid for only when the job starts.
- It gives an agent a unit of work. A repeatable workflow gets a name, inputs, steps, decision points and a defined output. Use an MCP server for live data and controlled actions, and a skill for the sequence, examples and templates.
- It is institutional memory. Skills are version-controlled folders you can review, compare and roll back. A prompt dies in a thread; a skill survives staff turnover.
- It carries the practitioner's judgement. The refuse-list lives in the skill: no invented quantities, no sending, no legal advice. It is written once instead of re-negotiated every conversation.
- It raises the quality floor. In a randomised trial on mid-level professional writing, AI assistance cut time by 40% and lifted quality by 18%. A skill is how you keep that gain while matching your own house standard.
Why one folder for every AI matters
Anthropic released the format as an open standard for cross-platform portability in December 2025. It is now read by Claude, ChatGPT and Codex, GitHub Copilot and Copilot Studio, Cursor, Gemini CLI, Hermes Agent and many others. You write the procedure once and copy the folder, instead of re-writing it for each vendor. Five rules make that portability hold:
- Write job-first, not tool-first. Say "fill the report template from these inputs", not "use Claude's PDF tool".
- One skill, one job. If the body needs a table of contents, split it.
- The description is the trigger, not a summary. It decides when the model considers the skill at all.
- Push deterministic steps into scripts. Maths, form fills and pattern matching should not be things the model can creatively redo.
- Keep always-on rules out of skills. Every installed skill's name and description is loaded on every turn. Standing rules belong in AGENTS.md or the system prompt, and unused skills should be deleted.
One caution. Format portability is not behavioural equivalence. Expect consistent structure and guardrails across tools, and check the judgement on each one. No public cross-tool benchmark was found for this briefing.
Four places to use them
Each area follows the same pattern: the skill holds the evidence pack, the structure, and the boundaries.
Writing and knowledge work
Define your voice as observable behaviours, not adjectives, and pair every rule with an annotated sample. Build first: a house-style editor, tender and capability responses, board and PCG reports, learner-facing communications. Guardrail: no invented facts, figures or citations, and a named human reviewer.
People management
Skills can prepare the evidence pack for a conversation, structure it, and set the boundaries: 1:1 prep, review drafting from recorded evidence, delegation briefs, onboarding packs. Guardrail: never send, never decide. No adverse-action content without HR review, and the manager owns the outcome.
Construction management
This is the ITP pattern for a thinking job: the description is the hold point, the body is the inspection steps, the scripts are the checks. Build first: variation and EOT packs, clause memos, tender comparison, delay chronologies, monthly reports, NCR close-out. Guardrail: cite the clause, separate fact from claim from opinion, no invented dates or quantities.
Trainers, assessors and lecturers
ASQA is clear that AI must not replace human decision-making in VET. A written, versioned procedure is the simplest evidence that oversight exists. Build first: evidence mapping, assessment-tool validation, marking-guide checks, reasonable adjustments, feedback banks. Guardrail: no competency judgements, a named human validator, and no auto-issuing.
What to build first
Rank jobs by how often they happen multiplied by the cost of getting them wrong. Start where a wrong output costs money, a student outcome or a safety record, not where the demo looks best. Do not start with a persona such as "you are an expert construction manager". Personas leak; jobs compound.
- Variation and EOT entitlement pack (construction)
- Evidence mapping to unit requirements (education)
- PCG or board report (construction)
- Assessment-tool validation (education)
- House-style editor (writing)
- Performance-review drafting (people management)
- Contract clause memo (construction)
- Tender evaluation matrix (construction)
- Trainer PD capability plan (education)
- 1:1 prep pack (people management)
Where to find real skills
Before writing your own, read how the published ones are built. These are the best places to start.
Official Agent Skills repo (~179k stars)
Spec, template, skill-creator and real docx, pdf, pptx and xlsx skills. This is the format everyone copies.
OpenAI curated skills (~28k stars)
The same SKILL.md idea on the Codex side. Useful if you run both.
Construction office skills (ACS): 8 portable skills
Eight jobs for construction managers, PMs and site and office staff: minutes and actions, RFIs, site diary, inbox triage, ITP hold points, variation and EOT drafts, NCR and punch lists, and tender comparison. One skill, one artefact, with a human hold point on every send and commercial commit.
Hygiene (do not skip)
Skills can ship scripts. Treat a random GitHub skill like unsigned code.Every installed description rides every turn - install 8 jobs, not 800.Prefer repos that show a real SKILL.md, a references folder and a refuse-list over "prompt pack" READMEs.Vet with something like a skill-auditor, or the royalpinto list, before it touches a project folder.
Where the evidence is thin
- Adoption is broad but shallow. More than 40% of construction firms use AI daily, but fewer than 1.5% use it across multiple processes. Skills are a credible fix for that gap, but no source found here shows skill reuse working at organisational scale in construction or VET.
- Most of the "why" comes from vendors (Anthropic, OpenAI, Microsoft, Google), who have a commercial interest in the format winning. The independent evidence is one randomised writing trial and regulator documents that set obligations rather than benefits.
- Context has a cost. Every installed skill's description is loaded each turn. More skills is not better.
- The profession is optimistic and anxious at once. RICS found 69% of project managers see AI adding value, while 44% are concerned about its impact on their own role.
- Gap: no Australian study of skill reuse in construction VET was located. That gap is itself a content opportunity.
Sources
- Agent Skills open standard (agentskills.io)
- Agent Skills client showcase
- Anthropic Engineering: Equipping agents for the real world with Agent Skills
- Anthropic: Introducing Agent Skills
- Claude Docs: Agent Skills overview
- Gemini CLI: Agent Skills
- OpenAI Developers: Build skills
- Microsoft Learn: Skills overview for agents
- AGENTS.md open format
- ASQA: Responsible use of AI in VET
- TEQSA: Enacting assessment reform in a time of AI
- TEQSA: Gen AI knowledge hub
- Microsoft: 2026 Work Trend Index
- MIT News: ChatGPT boosts productivity on professional writing tasks
- Science: Experimental evidence on the productivity effects of generative AI
- UNESCO: AI competency framework for teachers
- RICS: AI in construction report 2025
- Future Skills Organisation: National AI Skills Forum 2026 insights paper
- InnovationAus: National push to embed AI skills across the VET sector
- Dan Cumberland Labs: AI in the construction industry 2026
- Glean: How to create a brand voice guide for AI tools
Keep reading
More on AI, construction, education and project management.
Use AI as a guide, not the final word
AI is powerful, but it can be wrong or out of date, and it does not produce compliant or engineered work on its own. Always review the outputs yourself. Make sure any codes, standards or regulations involved are current and match your local specifications before you rely on them.