Skip to main content
Your agent has four skills. In Part 1 you saw one of them, diagnostic-triage-questions, change a reply. The registry has many more. In this part you read that skill’s page, find another and check it before you trust it, add it, then add a whole pack and make one skill your own. Your code doesn’t change, on any tab from Part 2.
Made your copy on Part 1’s page? Then your prompt versions here are one higher, and you’ve already done step 6 if you changed the fee.

1. Read a skill you’ve seen work

Open diagnostic-triage-questions, the skill that made the agent ask about the broken kettle. Its page shows:
  • What it changes: how many of its test cases passed, and what it did to tokens, turns and response time. Not every skill has a clear result, and the page says so when it doesn’t. When we wrote this, this skill had passed 13 of 14 cases, and its lift was withheld because that run didn’t record which model graded it.
  • Safety checks: a code scan, a review of what its instructions try to do, and a content review.
  • What’s inside: the instructions your model reads when it opens the skill. Read them: this text goes into your agent’s conversation. The questions you saw in Part 1 come from its list here.
You’ve seen this skill change a reply, and its page shows no lift. A score tells you how a skill did on its own tests, not whether it helps your agent.

2. Search for another

Say you want the agent to handle a complaint it has no rules for: a customer saying Acme broke a service promise. Open the registry and search for sla. Each result shows what the skill says it does, its SkillScore, and whether it passed its safety screening. Read the description, not just the score. Some of these results are about reviewing contract terms, and a search for refund also finds skills that write a returns policy page for a website. They can score well and still be no use to this agent.

3. Check it, then add it

Open sla-breach-response. It tells a model how to answer a missed commitment: confirm the miss, concede it, and offer only the remedy the agreement defines. When we wrote this, its page read “Passed 20 of 22 test cases with the skill, and 3 of 22 without it”. That’s a much clearer result than the last skill’s, so read the rest of the page the same way. Check which version it measured: this result is for v1, and adding the skill gets you v2. Check what its test cases ask, too. A strong result on those cases still says nothing about your agent’s questions. Scores give you a shortlist, and your own cases decide. Part 5 shows you how. Run this from your project, where .env has your DECIMAL_API_KEY:
Run python main.py again. The menu now has five skills, and you changed no code: the agent’s skills live on DecimalAI, and its next run reads them. The model decides whether to open the new one, and the kettle’s refund doesn’t call for it. Part 4 shows you what it chose.

4. Add a whole pack

A pack is a starting point for one kind of agent: a starter prompt and a set of skills. Look at the Support pack, then apply it. Applying it to the agent you already have keeps its runs and history together.
It shows what will change and asks first (example output):
--dry-run shows the change and stops. The CLI reference has the other options.
The starter became your prompt’s next version, and your older versions stay in its history. Your agent now has nine skills: the pack’s eight, including the new support-reply-format, csat-followup-request, mohitagw15856/escalation-tree and incident-status-comms, and your acme-returns-policy. It’s all one change on the agent’s Versions page.

5. Fill in the blanks

The starter has blanks in angle brackets for the parts only you know. On the Prompt & skills page, edit the new version:
  • Replace <product> with Acme's online store.
  • Replace the Tools line’s blank with the four tools your code attaches, and tell it to use the first one:
Save. That makes v3. Run python main.py again: the agent now runs on the new prompt, and the kettle answer is still right. fallback= is now only used if DecimalAI can’t be reached at startup.

6. Make one skill your own

A registry skill is someone else’s text. To change it, make an editable copy for your organization. Here you’ll change Acme’s restocking fee.
  1. Open acme-returns-policy. In the Manage it in DecimalAI card, click Create editable copy. Pick order-support as the agent, then click Create editable copy. Your copy keeps the skill’s name and takes the original’s place on order-support.
  2. Click Open your copy. On its Settings tab, in Body (Markdown), change 15% to 10%.
  3. Click Save Changes, then Publish (affects 1 agent). This makes a new version of your copy for your agent. It stays private: nothing goes to the public registry.
Run python main.py again:
$41.65 became $44.10, with no redeploy. That’s the same change as Part 1’s step 5, now reaching your own code. The wording varies from run to run, and now and then the reply gives the percentage without the amount. If yours does, run it again.
decimalai init order-support writes a runnable LangChain file for an agent that already exists in your workspace. --framework openai-agents writes one for that SDK instead, and decimalai init --help lists the rest.decimalai starter support downloads a local Support project with the pack’s prompt and skills copied into it. That copy is frozen: it needs no account and sends no traces, and nothing you change in the dashboard reaches it. The Support starter project walks through it.

Next: 4. See what it did

Read order-support’s traces: the skills it was offered and opened, the tools it called, and the versions each run used.