Multi-Tool Compositions

A single-tool skill does one thing — reads a file, runs a command, generates text. A multi-tool composition chains multiple tools in sequence: read the config file, analyze its structure, generate a migration script, run the migration, verify the result, commit the changes. Each step uses a different tool capability. The skill orchestrates the entire pipeline.

The design challenge is handoffs between steps. Step 1 produces output that Step 2 consumes. Step 2 produces output that Step 3 consumes. If any step fails, the pipeline must handle the failure gracefully — not crash, not continue with garbage input, but stop, report what failed, and suggest how to fix it. This means your core instructions need both the happy path and the failure paths for every step.

The most common multi-tool pattern is the read-analyze-write-verify loop. Claude reads existing code or data (tool: Read), analyzes it against a set of criteria (tool: extended thinking), writes the updated version (tool: Write/Edit), and verifies the result passes quality checks (tool: Bash). This four-step loop is the backbone of refactoring skills, migration skills, and code generation skills.

Design multi-tool skills as pipelines with checkpoints. After each step, the skill verifies the output before passing it to the next step. "After generating the migration script, run it in dry-run mode and verify no errors. Only proceed to execution if the dry run succeeds." Checkpoints prevent cascading failures — a bug in step 2 does not corrupt steps 3 through 6. They also provide natural retry points. If step 4 fails, retry from step 4 with the verified output from step 3, rather than restarting the entire pipeline.