Toolkit
Five helper skills.
A skill is a saved set of instructions your AI assistant can follow every time. Each step in your path tells you which one to use, so you never need to work through this page in order.
Download all five (ZIP)How to add a skill to your AI tool
- Download the skill’s ZIP and unzip it. You’ll get a folder with a file called SKILL.md.
- Put the folder where your tool looks for skills:
- Claude Code:
.claude/skills/in your project - Codex:
.agents/skills/in your project - VS Code with Copilot:
.github/skills/ - Claude Cowork: upload the ZIP in Customize → Skills (if your workspace allows it)
- Claude Code:
- Ask for it by name, e.g. “Use the scope-work skill on this request.”
Scope work
Turn a vague request into a clear, testable brief.
- You give it
- “Add an enquiry form for the workshop. Make it quick. We need bookings.” The business can confirm availability only after checking its diary.
- Good output looks like
- The brief calls the form an enquiry, names confirmation as a later human step, specifies fields and validation, and leaves a real booking integration out of scope.
Going further
Trigger test
Ask the assistant to use scope-work on the input. Then open a fresh task with this unrelated request: A vague but harmless formatting task that already has explicit acceptance criteria. Check whether automatic discovery selected the appropriate workflow. Explicit invocation is a separate test.
Compare two tools
Run the same input through two hosts. Record which skill loaded, where the evidence came from, the output structure and any unauthorized actions. Compare against the expected result and revise one instruction before retesting.
For your current step
Review evidence
Check every claim against the facts you supplied.
- You give it
- Draft: “Same-day repair for R250.” Source A: “Staff confirm timing and price after inspection.” Source B: an undated sample flyer with R250.
- Good output looks like
- Timing is unsupported and the flyer cannot establish a current price. Rewrite as: “Request an inspection. We’ll confirm price and timing after reviewing your enquiry.”
Going further
Trigger test
Ask the assistant to use review-evidence on the input. Then open a fresh task with this unrelated request: A request to invent a clearly fictional sample slogan. Check whether automatic discovery selected the appropriate workflow. Explicit invocation is a separate test.
Compare two tools
Run the same input through two hosts. Record which skill loaded, where the evidence came from, the output structure and any unauthorized actions. Compare against the expected result and revise one instruction before retesting.
Review project changes
Review a change before you hand it over.
- You give it
- The form brief requires an email address and a review queue. A diff changes the success copy to “Booking confirmed” and disables email validation.
- Good output looks like
- Flag both defects, point to changed lines or files, reproduce the invalid submission, and show why confirmation wording fails the brief.
Going further
Trigger test
Ask the assistant to use review-project-changes on the input. Then open a fresh task with this unrelated request: A request to summarize a finished document with no change or review intent. Check whether automatic discovery selected the appropriate workflow. Explicit invocation is a separate test.
Compare two tools
Run the same input through two hosts. Record which skill loaded, where the evidence came from, the output structure and any unauthorized actions. Compare against the expected result and revise one instruction before retesting.
Test automation
Test an automation before you switch it on.
- You give it
- Input A-102 is a complete enquiry. Replay A-102. Input A-103 has no email. An AI step returns the unsupported label ‘urgent’.
- Good output looks like
- A-102 produces one review record across two attempts. A-103 and the unsupported label go to human review. The schedule stays paused until an owner reviews the log.
Going further
Trigger test
Ask the assistant to use test-automation on the input. Then open a fresh task with this unrelated request: A request to explain what a trigger is without a workflow to test. Check whether automatic discovery selected the appropriate workflow. Explicit invocation is a separate test.
Compare two tools
Run the same input through two hosts. Record which skill loaded, where the evidence came from, the output structure and any unauthorized actions. Compare against the expected result and revise one instruction before retesting.
Prepare handover
Prepare a clear handover with proof it works.
- You give it
- The demo page passes mobile and keyboard checks. The form was tested with sample data. Production delivery and password reset were never tested.
- Good output looks like
- State the demo evidence, list production checks as open, identify the owner and record rollback instructions without implying go-live approval.
Going further
Trigger test
Ask the assistant to use prepare-handover on the input. Then open a fresh task with this unrelated request: A request to publish an approved release immediately without preparing a handover. Check whether automatic discovery selected the appropriate workflow. Explicit invocation is a separate test.
Compare two tools
Run the same input through two hosts. Record which skill loaded, where the evidence came from, the output structure and any unauthorized actions. Compare against the expected result and revise one instruction before retesting.