Turn 37 Prebuilt Sales Skills Into an Auditable CRM Playbook

Salesforce and Anthropic's Claudeforce puts 37 prebuilt sales skills in a Claude plugin for select pilot customers; open beta starts September 2026. Salesforce and Anthropic announced Claudeforce on August 26, 2026, in San Francisco. The first shipping component is a Salesforce plugin inside Claude. The plugin names meeting preparation, deal health review, and pipeline review as examples. Additional prebuilt skills are scheduled to begin launching in late 2026. Salesforce and Anthropic say the skills were jointly built to draw on Claude's reasoning, agentic tool use, and generative interface generation.

Treat the 37 skills as a pattern library and audit each for stage, owner, trigger, data, rule, and KPI before enabling.

Actions taken through the plugin are routed back through Salesforce so that business rules are enforced at the point of execution. Administrative setup is described as a single connection performed once by an administrator, with authentication and permissions managed centrally. If a skill can create, update, or summarize CRM records, it needs the same governance as any other workflow: scope, owner, trigger, data, rule, and metric.

Audit the 37 skills as a pattern library

Build a one-row inventory for each of the 37 skills. Do not start by enabling everything. Start by classifying each skill as adopt, customize, or retire. A skill earns adoption only when it fits an existing stage, has a named owner, and changes a measurable outcome.

1. Map the stage. Assign the skill to one primary stage: prospecting, qualification, opportunity, negotiation, close, post-sale, or pipeline hygiene. If a skill spans multiple stages, split it into separate use cases or mark it for customization.
2. Name the owner. The owner is the person accountable for the output, not the model. Use SDR, account executive, sales engineer, sales manager, RevOps, or a support role. If no one owns the output, the skill is a demo.
3. Define the trigger. Record what starts the skill: a record update, a meeting scheduled, a stage change, an aging opportunity, a forecast date, or a manual request. Vague triggers create unpredictable agent behavior.
4. List the data dependency. Identify the CRM objects and fields the skill reads or writes. Include account, contact, opportunity, activity, notes, tasks, forecast, and any pricing or approval data. A skill that depends on incomplete data should be blocked, not forced.
5. State the business rule. Write the rule the skill must respect: territory, approval, discount, data privacy, stage criteria, forecast confidence, or customer segmentation. Because actions are routed back through Salesforce, the audit should confirm that the rule is enforced at execution, not only in the prompt.
6. Choose the KPI. Pick one primary metric and one guardrail. Examples: meeting preparation time, deal review cycle, pipeline coverage, forecast variance, activity completion, stale opportunity count, or time to next action. The KPI should show whether the skill improved the workflow or merely added output.
7. Classify the skill. Mark it adopt if it matches the current process. Mark it customize if it needs local rules, data checks, or a different owner. Mark it retire if it duplicates another skill, has no owner, or cannot be measured.

Use the same inventory for the additional prebuilt skills scheduled to begin launching in late 2026. Do not let new skills enter production before they pass the same audit. The pattern library grows only when each new skill proves its role.

Deploy the audited skills as a controlled agent workflow

After classification, deploy the adopted skills in a controlled sequence. Start with the workflow that has the clearest owner and the least risk. A pipeline review or deal health review is often a better first target than a skill that writes directly to records, because the output is easier to review and the business rule is easier to state.

• Run read-only first where possible. Let the skill summarize, score, or flag records before it creates or updates them. Once the output is trusted, allow execution through the Salesforce plugin.
• Keep permissions central. The administrative setup is described as a single connection performed once by an administrator, with authentication and permissions managed centrally. Use that central control to limit which users, roles, and records can invoke each skill.
• Test the enforcement point. For each adopted skill, run a test that should fail under a business rule. Confirm that the action is blocked or corrected at the Salesforce layer, not only in the Claude interface.
• Audit the generated interface. The skills are said to draw on generative interface generation. If a skill produces a view, form, or dashboard, record what it displays, what actions it exposes, and who can see it. A generated interface is a control surface, not a cosmetic layer.
• Log the execution path. For every action, capture the trigger, the user or role, the skill, the CRM object, the rule applied, and the result. This is the evidence that the agent workflow is auditable.

A skill that cannot be mapped to a stage, owner, trigger, data dependency, business rule, and KPI is not ready for production. It is a candidate for customization or retirement.

Enter each adopted skill into the CRM inventory with owner, trigger, rule, and KPI before production use.