Run a skill

Every skill in this Playbook follows the same four-step pattern. Once you've run two or three, the rest are muscle memory.

Before you start

You need:

• Claude Desktop, OpenAI Codex, or Microsoft 365 Copilot signed in. Any one of the three runs every skill in this Playbook. (See the install pages if you haven't done this yet — pick the runtime your company already uses.)
• The skill page open in a separate window — you'll be copying its prompt body and looking at its Inputs to attach list.
• The input files the skill names. Don't start without them; the AI will fabricate values if it's asked to run a skill with missing inputs.

Step 1 — Load the skill

Two ways. Use the first.

The reusable way (recommended)

Save the skill's prompt body once. You'll only do this the first time you run that skill.

In Claude Desktop:

1. Sidebar → Projects → + New Project.
2. Name it after the skill (e.g., Monthly distributor scorecard).
3. Click Set project instructions, paste the skill's prompt body in, save.

In OpenAI Codex:

1. Sidebar → Skills → New Skill.
2. Name it after the Playbook skill.
3. Paste the prompt body into Instructions, save.

In Microsoft 365 Copilot (Agent Builder):

1. Open Copilot Chat → Agents in the right rail → + New agent.
2. Name it after the Playbook skill.
3. Paste the prompt body into Instructions. Save.

To run later: open the Project (Claude), Skill (Codex), or agent (Copilot) and start a new chat inside it. The prompt body is automatically in scope.

The one-off way

Open a fresh chat. Paste the skill's prompt body into the message box. Move to Step 2 below in the same message.

Fine for testing a skill once. Don't use it for anything you'll run more than twice.

Step 2 — Attach the inputs

Each skill page has a section called Inputs to attach with a named list of files. Drag every file in that list into the conversation. (Codex / Claude: use the paperclip icon. Copilot: use the Add content icon → Upload images and files for local files or Attach cloud files for OneDrive / SharePoint files.)

All three runtimes read PDFs, CSVs, XLSX, DOCX, plain text, code, and images directly. If the skill calls for a CSV and your data lives in an Excel file, save a copy as CSV first (File → Save As → CSV UTF-8) — every runtime handles CSV more reliably than XLSX with merged cells, pivot tables, or macros. Copilot has a hard cap of ten files or SharePoint pages per prompt; if you bump into it, package inputs into a single SharePoint folder and reference the folder.

If a file is missing, don't fake it. Either get the missing file before running, or ask the AI: One of the inputs is missing — which one, and what should I do? It will tell you which input it's missing and whether the skill can still produce useful output without it.

Step 3 — Run it

Type one of these as your message:

• Run the skill.
• Generate the report.
• Run.

Press Enter. Most skills return in 20–90 seconds. Agent tasks that fetch web content take 2–4 minutes and you'll see progress text ("reading attachment 1 of 3," "fetching pricing page") while it works.

You don't need to babysit. Switch tabs, get coffee. Both runtimes notify you when the run completes if you've granted notifications.

Step 4 — Review and edit

The output is a starting point, not a finished deliverable. Every skill in this Playbook produces output a human owner reviews before it leaves the company. The review checklist:

1. Spot-check the headline numbers. AI math is usually right but verify the top-line figures against the source files.
2. Read for context the AI doesn't have. It doesn't know about Distributor X's late payments, Supplier Y's tooling-cost dispute, the off-the-record reason your VP doesn't want to highlight a particular product line. Edit those out or in as appropriate.
3. Edit for voice. The default output is neutral business prose. Add the strong opinion, the specific name, the inside reference that makes it sound like you wrote it.
4. Check the citations or references. If the skill cites a source, click through. Hallucinated citations are rare on Claude Sonnet 4.6 and GPT-5.5 but not zero.

If something is off, don't start from scratch. Tell the skill what's wrong:

The recommendation in section 2 contradicts what I know about Distributor X's contract terms — they have a 90-day cure window. Redo just that section.

The skill will revise without re-running everything.

When a skill stops working

A skill that worked last month and produces garbage today usually means one of three things:

1. The input format changed. Your ERP export now has different column names, or someone reformatted the rebate spreadsheet. Add a note to the top of your message: Heads-up: column 'Net Sell-Through' is now called 'NST_Net'. The skill will adapt.
2. The model was updated. Both Anthropic and OpenAI ship model updates regularly. A new model may handle ambiguity differently. Re-run the skill with the same inputs; if the output is reliably worse, tighten the prompt or email Jason.
3. You're attaching more than the skill expects. Extra files dilute the model's attention. Attach only what the Inputs to attach list names — no "just in case" extras.

If the skill stops working and you can't get it back, email headstart@nunncurtislabs.com. The Playbook comes with a lifetime guarantee for a reason.

Tips that aren't on the skill pages

A few things that apply to every skill:

• Start a fresh chat for each run. Long threads accumulate context that can leak into the next run's output. One run, one chat. The Project/Skill instructions persist; the chat history doesn't need to.
• Don't over-prompt. The skill's prompt body is tuned. Adding "Make it really good" or "Be thorough" usually makes the output worse, not better. Trust the prompt, run it, edit the output.
• Save good outputs as Project knowledge (Claude) or attach them to next month's run (Codex). Skills get better when they can see what last month's good version looked like.

Where to go next

• Scheduled tasks — how to make a skill run on a cadence so you don't have to remember to.
• Skills section in the sidebar — pick one and run it.
• Guarantee and support — what to do if you get stuck.