High-value uses: from specs to checks

The best early wins for Claude Code in SAP testing aren't flashy end-to-end scripts. They're the slow, text-heavy jobs that sit around automation: turning a functional spec into test cases, planning data, reading a 4,000-line log, or listing what a transport might break. In each of them the agent produces a draft quickly, and a person who knows the process makes it right. This lesson covers six such uses, the prompt pattern that makes them work, and the repository setup that keeps them safe.

A prompt pattern that holds up

Vague prompts produce plausible nonsense: invented field names, transactions that don't exist in your release, data that doesn't match your configuration. A short, structured prompt fixes most of that, and it makes the output easy to review because you asked for the format in advance.

Notice what the prompt forbids: invention. The ASSUMED marker turns a hidden error into a visible question. When Maria reviews the draft, she spends her time on the business rules rather than hunting for fabricated details.

A CLAUDE.md for an SAP test repository

Put the rules every session needs in the repository's CLAUDE.md, which Claude Code loads at the start of each session. Anthropic recommends keeping each file under about 200 lines; /init generates a starting file. Here is the core of Contoso's:

# Contoso SAP test repository
## Context
- System under test: CMQ client 100 (quality). Never CMP or any production host.
- Contoso data: company code 1710, plant 1710 Detroit, sales area 1710/10/00, customer Great Lakes Fleet Services, material CMC-ALT-100.
- Approval rules: purchase requisitions over $10,000 need release; purchase orders over $25,000 need approval.
## How we write tests
- Test cases: Given / When / Then, one business outcome each, plus at least one negative case.
- Playwright: getByRole, getByLabel or getByText first. Never IDs that start with __. No waitForTimeout.
- Each test starts from a fresh browser context and creates or looks up its own data.
## Data and secrets
- Read credentials only from environment variables. Never write them into files, prompts or logs.
- Use only the masked extracts in data/masked/. Never paste production data.
## Commands
- npx playwright test --project=chromium
- npx playwright show-trace test-results/<test>/trace.zip
## Review
- Every generated test carries a REVIEW: <owner> comment until a person signs it off.

Remember the limit you learned in lesson 1: CLAUDE.md is guidance. The hard limits go in .claude/settings.json as permission rules, such as "deny": ["Read(./.env)", "Bash(git push *)"], and in a PreToolUse hook that rejects any command or URL containing a production host name. Anthropic's permission docs add one more caution: Bash rules match the command text Claude writes, so they aren't a security boundary around a program. For enforcement that doesn't depend on wording, turn on sandboxing.

Notes on three of the uses

• API checks and SAP's API Policy. SAP published a new API Policy on April 27, 2026. SAP's own overview says integration with autonomous AI agents is allowed only through an endorsed SAP architecture, such as the MCP Gateway in SAP Integration Suite. Community analyses place this in section 2.2.2 of policy version 4.2026a. A script Claude Code writes and a person runs is ordinary automation; an agent that plans and executes its own sequence of SAP API calls is the case the policy targets. Treat that line as a compliance check, confirm it with your SAP contract owner, and use published APIs only.
• Change-impact review. The agent reads object lists quickly and suggests which processes they touch, but it doesn't know your usage data. Confirm its suggestions against your own change-impact tooling and usage statistics before you cut regression scope.
• ABAP Unit drafting. SAP says ABAP development in Visual Studio Code started in Q2 2026, first in the ABAP Cloud scope, and that its ABAP MCP server for Eclipse and VS Code is generally available to connect third-party tools. Generated test classes still need the design skills from sap-abp-205: test doubles, one behavior per method, and meaningful assertions.

SAP is building AI into its own test tooling too. SAP Learning lists a Joule Test Management Assistant among its migration and modernization assistants, described as scoping testing and creating test cases on SAP Cloud ALM. Its availability for your contract wasn't confirmed when this course was reviewed, so check with SAP before planning around it.