diff --git a/README.md b/README.md
index 48b90f1..90aa740 100644
--- a/README.md
+++ b/README.md
@@ -157,8 +157,18 @@ caseforge lint --spec openapi.yaml
 --exclude-tag string    Comma-separated OpenAPI tags to exclude (e.g. 'deprecated')
 --auth-bootstrap      Wrap all secured-endpoint cases with an auth setup step
 --with-oracles        Mine response body constraints via LLM and inject as assertions (requires LLM)
+--force               Regenerate even when spec hash matches existing output
+--annotation-batch N  Number of operations to annotate per LLM call (0 = one call per op; recommended: 8–20)
 ```
 
+**Smart regeneration:** `gen` hashes the spec file on each run. If the hash matches the previously generated output, it exits early with a `✓ Spec unchanged` message. Use `--force` to bypass this.
+
+**Batch annotation:** By default each operation is annotated in a separate LLM call. On large specs (20+ operations) this can take several minutes. `--annotation-batch 10` groups 10 operations per call, reducing round-trips dramatically.
+
+**Rate-limit backoff:** 429 responses from LLM providers trigger automatic exponential backoff (5 s → 15 s → 30 s → 60 s) before retrying.
+
+**Hurl output headers:** Every generated `.hurl` file includes `# case_id=` and `# case_name=` comment headers for traceability.
+
 ### `caseforge run`
 
 ```
diff --git a/cmd/gen.go b/cmd/gen.go
index ad06dc1..c297e64 100644
--- a/cmd/gen.go
+++ b/cmd/gen.go
@@ -403,7 +403,7 @@ func runGen(cmd *cobra.Command, args []string) error {
 	if genMaxCasesPerOp > 0 {
 		engine.SetMaxCasesPerOp(genMaxCasesPerOp)
 	}
-	if genAnnotationBatch > 0 {
+	if genAnnotationBatch >= 1 {
 		engine.SetAnnotationBatch(genAnnotationBatch)
 	}
 	newCases, err := engine.Generate(parsedSpec)
diff --git a/internal/mcp/server_test.go b/internal/mcp/server_test.go
index b43b15a..677c163 100644
--- a/internal/mcp/server_test.go
+++ b/internal/mcp/server_test.go
@@ -27,7 +27,18 @@ func TestNewServerReturnsNonNil(t *testing.T) {
 }
 
 func TestServerHasGenerateTestCasesTool(t *testing.T) {
-	assert.NotNil(t, generateTestCasesTool())
+	tool := generateTestCasesTool()
+	require.NotNil(t, tool)
+	assert.Equal(t, "generate_test_cases", tool.Name)
+
+	rawSchema, _ := tool.InputSchema.(json.RawMessage)
+	var schema map[string]any
+	require.NoError(t, json.Unmarshal(rawSchema, &schema))
+	required, _ := schema["required"].([]any)
+	assert.Contains(t, required, "spec")
+
+	props, _ := schema["properties"].(map[string]any)
+	assert.Contains(t, props, "annotation_batch", "annotation_batch must be in schema for MCP clients to discover it")
 }
 
 func TestServerHasLintSpecTool(t *testing.T) {
diff --git a/internal/mcp/tools.go b/internal/mcp/tools.go
index 34afd88..54844a6 100644
--- a/internal/mcp/tools.go
+++ b/internal/mcp/tools.go
@@ -39,6 +39,10 @@ func generateTestCasesTool() *mcpsdk.Tool {
                     "type": "string",
                     "description": "Output format: hurl|markdown|csv|postman|k6 (default: hurl)",
                     "enum": ["hurl", "markdown", "csv", "postman", "k6"]
+                },
+                "annotation_batch": {
+                    "type": "integer",
+                    "description": "Number of operations to annotate per LLM call; reduces round-trips on large specs (0 = one call per op, recommended: 8–20)"
                 }
             },
             "required": ["spec"]
@@ -48,9 +52,10 @@ func generateTestCasesTool() *mcpsdk.Tool {
 
 func makeGenerateHandler(ctx context.Context, req *mcpsdk.CallToolRequest) (*mcpsdk.CallToolResult, error) {
 	var args struct {
-		Spec   string `json:"spec"`
-		Output string `json:"output"`
-		Format string `json:"format"`
+		Spec            string `json:"spec"`
+		Output          string `json:"output"`
+		Format          string `json:"format"`
+		AnnotationBatch int    `json:"annotation_batch"`
 	}
 	if err := json.Unmarshal(req.Params.Arguments, &args); err != nil {
 		r := &mcpsdk.CallToolResult{}
@@ -88,6 +93,9 @@ func makeGenerateHandler(ctx context.Context, req *mcpsdk.CallToolRequest) (*mcp
 	)
 	engine.AddSpecTechnique(methodology.NewChainTechnique())
 	engine.AddSpecTechnique(methodology.NewSecuritySpecTechnique())
+	if args.AnnotationBatch >= 1 {
+		engine.SetAnnotationBatch(args.AnnotationBatch)
+	}
 
 	cases, err := engine.Generate(parsedSpec)
 	if err != nil {
diff --git a/internal/methodology/engine.go b/internal/methodology/engine.go
index 2bf976d..0219ee6 100644
--- a/internal/methodology/engine.go
+++ b/internal/methodology/engine.go
@@ -87,8 +87,8 @@ func (e *Engine) SetMaxCasesPerOp(n int) {
 
 // SetAnnotationBatch sets the number of operations to annotate per LLM call.
 // 0 (default) uses sequential mode: one call per operation.
-// Values > 0 batch that many operations into a single call, reducing round-trips
-// at the cost of larger prompts. Recommended range: 5–20.
+// Values >= 1 batch that many operations into a single call, reducing round-trips
+// at the cost of larger prompts. Recommended range: 8–20.
 func (e *Engine) SetAnnotationBatch(n int) {
 	e.annotationBatch = n
 }