diff --git a/internal/captcha/grid.go b/internal/captcha/grid.go
index 79695d2..089d5e1 100644
--- a/internal/captcha/grid.go
+++ b/internal/captcha/grid.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"sort"
 
 	"github.com/lukaszraczylo/captcha-solver-mcp/internal/imageutil"
 	"github.com/lukaszraczylo/captcha-solver-mcp/internal/llm"
@@ -17,6 +18,9 @@ type GridOptions struct {
 	Cols        int
 	ImageWidth  int
 	ImageHeight int
+	Samples     int
+	Upscale     int
+	Annotate    bool
 }
 
 // GridResult lists tiles to click (row-major 0-based) plus optional pixel
@@ -25,17 +29,20 @@ type GridResult struct {
 	Tiles              []int             `json:"tiles"`
 	Centroids          []imageutil.Point `json:"centroids,omitempty"`
 	Confidence         float64           `json:"confidence"`
+	NoTargets          bool              `json:"no_targets"`
 	RecheckRecommended bool              `json:"recheck_recommended"`
 }
 
 func gridPrompt(o GridOptions, strict bool) string {
 	p := fmt.Sprintf(
 		"You are solving an image-grid captcha. The image is a %dx%d grid of tiles, "+
-			"numbered row-major starting at 0 (top-left). Instruction: %q. "+
-			"Return the 0-based indices of every tile that matches the instruction. "+
-			"If after selection more matching tiles would likely appear (dynamic grid), set recheck_recommended true. "+
-			`Respond ONLY with compact JSON: {"tiles":[..],"confidence":<0..1>,"recheck_recommended":<bool>}.`,
-		o.Rows, o.Cols, o.Instruction)
+			"numbered row-major starting at 0 (top-left) through %d (numbers may be overlaid on the image). "+
+			"Instruction: %q. "+
+			"Select EVERY cell that contains ANY part of the target, including small or partial portions. "+
+			"If no cell contains the target, return an empty tiles array and set no_targets true. "+
+			"If this is a dynamic grid where new matching tiles may appear after selection, set recheck_recommended true. ",
+		o.Rows, o.Cols, o.Rows*o.Cols-1, o.Instruction)
+	p += `Respond ONLY with compact JSON: {"tiles":[..],"confidence":<0..1>,"no_targets":<bool>,"recheck_recommended":<bool>}.`
 	if strict {
 		p += " You MUST output valid JSON only. No prose."
 	}
@@ -43,23 +50,118 @@ func gridPrompt(o GridOptions, strict bool) string {
 }
 
 // SolveGrid asks the vision model which tiles to click and (when grid geometry
-// is supplied) maps them to pixel centroids for humanized clicking.
+// is supplied) maps them to pixel centroids for humanized clicking. When
+// o.Samples > 1 it self-consistency votes across N solves, keeping tiles that
+// appear in a strict majority of successful samples.
 func SolveGrid(ctx context.Context, c llm.Client, img []byte, mediaType string, o GridOptions) (GridResult, error) {
+	img, mediaType = preprocessGrid(img, mediaType, o)
 	images := []llm.Image{{MediaType: mediaType, Data: img}}
+
+	if o.Samples <= 1 {
+		res, err := solveGridOnce(ctx, c, images, o, llm.Options{})
+		if err != nil {
+			return GridResult{}, err
+		}
+		res.Centroids = mapCentroids(res.Tiles, o)
+		return res, nil
+	}
+
+	opts := llm.Options{Temperature: 0.5}
+	results := make([]GridResult, 0, o.Samples)
+	for i := 0; i < o.Samples; i++ {
+		res, err := solveGridOnce(ctx, c, images, o, opts)
+		if err != nil {
+			continue
+		}
+		results = append(results, res)
+	}
+	if len(results) == 0 {
+		return GridResult{}, fmt.Errorf("model did not return parseable tile selection")
+	}
+	voted := voteGrid(results)
+	voted.Centroids = mapCentroids(voted.Tiles, o)
+	return voted, nil
+}
+
+// preprocessGrid optionally upscales then annotates the screenshot. On any
+// preprocessing error it falls back to the original bytes so a solve is never
+// failed by an optional enhancement.
+func preprocessGrid(img []byte, mediaType string, o GridOptions) ([]byte, string) {
+	if o.Upscale > 1 {
+		if up, mt, err := imageutil.Upscale(img, o.Upscale); err == nil {
+			img, mediaType = up, mt
+		}
+	}
+	if o.Annotate {
+		if ann, err := imageutil.AnnotateGrid(img, o.Rows, o.Cols); err == nil {
+			img, mediaType = ann, "image/png"
+		}
+	}
+	return img, mediaType
+}
+
+// solveGridOnce performs a single solve with one stricter retry on parse fail.
+func solveGridOnce(ctx context.Context, c llm.Client, images []llm.Image, o GridOptions, opts llm.Options) (GridResult, error) {
 	for attempt := 0; attempt < 2; attempt++ {
-		raw, err := c.Vision(ctx, gridPrompt(o, attempt > 0), images, llm.Options{})
+		raw, err := c.Vision(ctx, gridPrompt(o, attempt > 0), images, opts)
 		if err != nil {
 			return GridResult{}, err
 		}
 		var res GridResult
 		if err := json.Unmarshal([]byte(extractJSON(raw)), &res); err == nil && res.Tiles != nil {
-			res.Centroids = mapCentroids(res.Tiles, o)
 			return res, nil
 		}
 	}
 	return GridResult{}, fmt.Errorf("model did not return parseable tile selection")
 }
 
+// voteGrid keeps tiles appearing in a strict majority of successful samples.
+// Confidence is the average agreement of the chosen tiles; NoTargets is true
+// when a majority of samples reported no targets.
+func voteGrid(results []GridResult) GridResult {
+	n := len(results)
+	tileCounts := make(map[int]int)
+	noTargets := 0
+	recheck := 0
+	for _, r := range results {
+		seen := make(map[int]struct{}, len(r.Tiles))
+		for _, t := range r.Tiles {
+			if _, dup := seen[t]; dup {
+				continue
+			}
+			seen[t] = struct{}{}
+			tileCounts[t]++
+		}
+		if r.NoTargets {
+			noTargets++
+		}
+		if r.RecheckRecommended {
+			recheck++
+		}
+	}
+
+	tiles := make([]int, 0, len(tileCounts))
+	var agreeSum float64
+	for t, cnt := range tileCounts {
+		if cnt*2 > n { // strict majority: more than half
+			tiles = append(tiles, t)
+			agreeSum += float64(cnt) / float64(n)
+		}
+	}
+	sort.Ints(tiles)
+
+	var conf float64
+	if len(tiles) > 0 {
+		conf = agreeSum / float64(len(tiles))
+	}
+	return GridResult{
+		Tiles:              tiles,
+		Confidence:         conf,
+		NoTargets:          noTargets*2 > n,
+		RecheckRecommended: recheck*2 > n,
+	}
+}
+
 func mapCentroids(tiles []int, o GridOptions) []imageutil.Point {
 	if o.Rows <= 0 || o.Cols <= 0 || o.ImageWidth <= 0 || o.ImageHeight <= 0 {
 		return nil
diff --git a/internal/captcha/grid_test.go b/internal/captcha/grid_test.go
index d8693a5..b4fddce 100644
--- a/internal/captcha/grid_test.go
+++ b/internal/captcha/grid_test.go
@@ -1,12 +1,32 @@
 package captcha
 
 import (
+	"bytes"
 	"context"
+	"image"
+	"image/color"
+	"image/png"
 	"testing"
 
 	"github.com/lukaszraczylo/captcha-solver-mcp/internal/llm"
 )
 
+// smallPNG returns a tiny valid PNG for exercising preprocessing paths.
+func smallPNG(t *testing.T) []byte {
+	t.Helper()
+	img := image.NewRGBA(image.Rect(0, 0, 6, 6))
+	for y := 0; y < 6; y++ {
+		for x := 0; x < 6; x++ {
+			img.Set(x, y, color.RGBA{R: uint8(x * 40), G: uint8(y * 40), B: 80, A: 255})
+		}
+	}
+	var buf bytes.Buffer
+	if err := png.Encode(&buf, img); err != nil {
+		t.Fatalf("encode png: %v", err)
+	}
+	return buf.Bytes()
+}
+
 func TestSolveGrid(t *testing.T) {
 	f := &llm.Fake{VisionResp: `{"tiles":[0,4,8],"confidence":0.8,"recheck_recommended":true}`}
 	res, err := SolveGrid(context.Background(), f, []byte{1}, "image/png", GridOptions{
@@ -25,3 +45,53 @@ func TestSolveGrid(t *testing.T) {
 		t.Fatal("recheck flag lost")
 	}
 }
+
+func TestSolveGridVoting(t *testing.T) {
+	f := &llm.Fake{VisionResponses: []string{
+		`{"tiles":[0,1],"confidence":0.9}`,
+		`{"tiles":[0,2],"confidence":0.9}`,
+		`{"tiles":[0,1],"confidence":0.9}`,
+	}}
+	res, err := SolveGrid(context.Background(), f, []byte{1}, "image/png", GridOptions{
+		Instruction: "select all buses", Rows: 3, Cols: 3, Samples: 3,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// 0 in 3/3, 1 in 2/3 (>half), 2 in 1/3 (dropped).
+	if len(res.Tiles) != 2 || res.Tiles[0] != 0 || res.Tiles[1] != 1 {
+		t.Fatalf("voted tiles %+v, want [0 1]", res.Tiles)
+	}
+	if len(f.VisionCalls) != 3 {
+		t.Fatalf("expected 3 vision calls, got %d", len(f.VisionCalls))
+	}
+}
+
+func TestSolveGridNoTargets(t *testing.T) {
+	f := &llm.Fake{VisionResp: `{"tiles":[],"confidence":0.9,"no_targets":true}`}
+	res, err := SolveGrid(context.Background(), f, []byte{1}, "image/png", GridOptions{
+		Instruction: "select all buses", Rows: 3, Cols: 3,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !res.NoTargets {
+		t.Fatal("no_targets flag lost")
+	}
+	if len(res.Tiles) != 0 {
+		t.Fatalf("expected empty tiles, got %+v", res.Tiles)
+	}
+}
+
+func TestSolveGridPreprocessing(t *testing.T) {
+	f := &llm.Fake{VisionResp: `{"tiles":[0,4],"confidence":0.8}`}
+	res, err := SolveGrid(context.Background(), f, smallPNG(t), "image/png", GridOptions{
+		Instruction: "select all buses", Rows: 3, Cols: 3, Annotate: true, Upscale: 2,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(res.Tiles) != 2 {
+		t.Fatalf("expected 2 tiles, got %+v", res.Tiles)
+	}
+}
diff --git a/internal/captcha/rotation.go b/internal/captcha/rotation.go
index 4db8b45..431fa2d 100644
--- a/internal/captcha/rotation.go
+++ b/internal/captcha/rotation.go
@@ -21,7 +21,8 @@ func SolveRotation(ctx context.Context, c llm.Client, img []byte, mediaType, ins
 		"You are solving a rotation/choice puzzle captcha. Instruction: %q. "+
 			"If asked to pick one image, return its 0-based index as choice. "+
 			"If asked to rotate, return the number of clockwise steps as rotations. "+
-			`Respond ONLY with compact JSON: {"choice":<int>,"rotations":<int>,"confidence":<0..1>}.`,
+			"First reason step-by-step about the correct orientation/choice, then give the final answer "+
+			`as the last thing you output: compact JSON {"choice":<int>,"rotations":<int>,"confidence":<0..1>}.`,
 		instruction)
 	raw, err := c.Vision(ctx, prompt, []llm.Image{{MediaType: mediaType, Data: img}}, llm.Options{})
 	if err != nil {
diff --git a/internal/captcha/text.go b/internal/captcha/text.go
index de70d0c..f674db6 100644
--- a/internal/captcha/text.go
+++ b/internal/captcha/text.go
@@ -4,7 +4,9 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"sort"
 
+	"github.com/lukaszraczylo/captcha-solver-mcp/internal/imageutil"
 	"github.com/lukaszraczylo/captcha-solver-mcp/internal/llm"
 )
 
@@ -13,6 +15,8 @@ type TextOptions struct {
 	Hint          string
 	Charset       string
 	Length        int
+	Samples       int
+	Upscale       int
 	CaseSensitive bool
 }
 
@@ -36,6 +40,7 @@ func textPrompt(o TextOptions, strict bool) string {
 	if o.Hint != "" {
 		p += "Hint: " + o.Hint + ". "
 	}
+	p += "Carefully distinguish visually similar glyphs: 0 vs O, 1 vs l vs I, 5 vs S, 2 vs Z, 8 vs B. "
 	p += `Respond ONLY with compact JSON: {"text":"<chars>","confidence":<0..1>}.`
 	if strict {
 		p += " You MUST output valid JSON and nothing else. Do not explain."
@@ -43,11 +48,40 @@ func textPrompt(o TextOptions, strict bool) string {
 	return p
 }
 
-// SolveText runs OCR via the vision model with one stricter retry on parse failure.
+// SolveText runs OCR via the vision model with one stricter retry on parse
+// failure. When o.Samples > 1 it self-consistency votes across N solves,
+// returning the most frequent text (tie-break by highest confidence).
 func SolveText(ctx context.Context, c llm.Client, img []byte, mediaType string, o TextOptions) (TextResult, error) {
+	if o.Upscale > 1 {
+		if up, mt, err := imageutil.Upscale(img, o.Upscale); err == nil {
+			img, mediaType = up, mt
+		}
+	}
 	images := []llm.Image{{MediaType: mediaType, Data: img}}
+
+	if o.Samples <= 1 {
+		return solveTextOnce(ctx, c, images, o, llm.Options{})
+	}
+
+	opts := llm.Options{Temperature: 0.5}
+	results := make([]TextResult, 0, o.Samples)
+	for i := 0; i < o.Samples; i++ {
+		res, err := solveTextOnce(ctx, c, images, o, opts)
+		if err != nil {
+			continue
+		}
+		results = append(results, res)
+	}
+	if len(results) == 0 {
+		return TextResult{}, fmt.Errorf("model did not return parseable captcha text")
+	}
+	return voteText(results), nil
+}
+
+// solveTextOnce performs a single solve with one stricter retry on parse fail.
+func solveTextOnce(ctx context.Context, c llm.Client, images []llm.Image, o TextOptions, opts llm.Options) (TextResult, error) {
 	for attempt := 0; attempt < 2; attempt++ {
-		raw, err := c.Vision(ctx, textPrompt(o, attempt > 0), images, llm.Options{})
+		raw, err := c.Vision(ctx, textPrompt(o, attempt > 0), images, opts)
 		if err != nil {
 			return TextResult{}, err
 		}
@@ -58,3 +92,37 @@ func SolveText(ctx context.Context, c llm.Client, img []byte, mediaType string,
 	}
 	return TextResult{}, fmt.Errorf("model did not return parseable captcha text")
 }
+
+// voteText returns the most frequent text across results. Ties break by the
+// highest confidence seen for the candidate; remaining ties break
+// deterministically by text. Confidence is the fraction of results that agree
+// with the winner.
+func voteText(results []TextResult) TextResult {
+	counts := make(map[string]int)
+	bestConf := make(map[string]float64)
+	for _, r := range results {
+		counts[r.Text]++
+		if r.Confidence > bestConf[r.Text] {
+			bestConf[r.Text] = r.Confidence
+		}
+	}
+	texts := make([]string, 0, len(counts))
+	for t := range counts {
+		texts = append(texts, t)
+	}
+	sort.Slice(texts, func(i, j int) bool {
+		a, b := texts[i], texts[j]
+		if counts[a] != counts[b] {
+			return counts[a] > counts[b]
+		}
+		if bestConf[a] != bestConf[b] {
+			return bestConf[a] > bestConf[b]
+		}
+		return a < b
+	})
+	winner := texts[0]
+	return TextResult{
+		Text:       winner,
+		Confidence: float64(counts[winner]) / float64(len(results)),
+	}
+}
diff --git a/internal/captcha/text_test.go b/internal/captcha/text_test.go
index ad68658..1895737 100644
--- a/internal/captcha/text_test.go
+++ b/internal/captcha/text_test.go
@@ -31,3 +31,21 @@ func TestSolveTextRetriesOnGarbage(t *testing.T) {
 		t.Fatalf("expected 2 attempts (one retry), got %d", len(f.VisionCalls))
 	}
 }
+
+func TestSolveTextVoting(t *testing.T) {
+	f := &llm.Fake{VisionResponses: []string{
+		`{"text":"ab12","confidence":0.7}`,
+		`{"text":"ab12","confidence":0.7}`,
+		`{"text":"ab13","confidence":0.9}`,
+	}}
+	res, err := SolveText(context.Background(), f, []byte{1}, "image/png", TextOptions{Samples: 3})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Text != "ab12" {
+		t.Fatalf("voted text %q, want ab12", res.Text)
+	}
+	if len(f.VisionCalls) != 3 {
+		t.Fatalf("expected 3 vision calls, got %d", len(f.VisionCalls))
+	}
+}