diff --git a/ci.sh b/ci.sh
index 760d6fbfb..3fe1b1ce8 100755
--- a/ci.sh
+++ b/ci.sh
@@ -133,16 +133,16 @@ cargo batch  \
     --- build --release --manifest-path examples/stm32u5/Cargo.toml --target thumbv8m.main-none-eabihf --out-dir out/examples/stm32u5 \
     --- build --release --manifest-path examples/stm32wb/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32wb \
     --- build --release --manifest-path examples/stm32wl/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32wl \
-    --- build --release --manifest-path examples/boot/application/nrf/Cargo.toml --target thumbv7em-none-eabi --features embassy-nrf/nrf52840 --out-dir out/examples/boot/nrf --bin b \
-    --- build --release --manifest-path examples/boot/application/nrf/Cargo.toml --target thumbv8m.main-none-eabihf --features embassy-nrf/nrf9160-ns --out-dir out/examples/boot/nrf --bin b \
-    --- build --release --manifest-path examples/boot/application/rp/Cargo.toml --target thumbv6m-none-eabi --out-dir out/examples/boot/rp --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32f3/Cargo.toml --target thumbv7em-none-eabi --out-dir out/examples/boot/stm32f3 --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32f7/Cargo.toml --target thumbv7em-none-eabi --out-dir out/examples/boot/stm32f7 --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32h7/Cargo.toml --target thumbv7em-none-eabi --out-dir out/examples/boot/stm32h7 --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32l0/Cargo.toml --target thumbv6m-none-eabi --out-dir out/examples/boot/stm32l0 --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32l1/Cargo.toml --target thumbv7m-none-eabi --out-dir out/examples/boot/stm32l1 --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32l4/Cargo.toml --target thumbv7em-none-eabi --out-dir out/examples/boot/stm32l4 --bin b \
-    --- build --release --manifest-path examples/boot/application/stm32wl/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/boot/stm32wl --bin b \
+    --- build --release --manifest-path examples/boot/application/nrf/Cargo.toml --target thumbv7em-none-eabi --features embassy-nrf/nrf52840,skip-include --out-dir out/examples/boot/nrf  \
+    --- build --release --manifest-path examples/boot/application/nrf/Cargo.toml --target thumbv8m.main-none-eabihf --features embassy-nrf/nrf9160-ns,skip-include --out-dir out/examples/boot/nrf \
+    --- build --release --manifest-path examples/boot/application/rp/Cargo.toml --target thumbv6m-none-eabi --features skip-include --out-dir out/examples/boot/rp \
+    --- build --release --manifest-path examples/boot/application/stm32f3/Cargo.toml --target thumbv7em-none-eabi --features skip-include --out-dir out/examples/boot/stm32f3 \
+    --- build --release --manifest-path examples/boot/application/stm32f7/Cargo.toml --target thumbv7em-none-eabi --features skip-include --out-dir out/examples/boot/stm32f7 \
+    --- build --release --manifest-path examples/boot/application/stm32h7/Cargo.toml --target thumbv7em-none-eabi --features skip-include --out-dir out/examples/boot/stm32h7 \
+    --- build --release --manifest-path examples/boot/application/stm32l0/Cargo.toml --target thumbv6m-none-eabi --features skip-include --out-dir out/examples/boot/stm32l0 \
+    --- build --release --manifest-path examples/boot/application/stm32l1/Cargo.toml --target thumbv7m-none-eabi --features skip-include --out-dir out/examples/boot/stm32l1 \
+    --- build --release --manifest-path examples/boot/application/stm32l4/Cargo.toml --target thumbv7em-none-eabi --features skip-include --out-dir out/examples/boot/stm32l4 \
+    --- build --release --manifest-path examples/boot/application/stm32wl/Cargo.toml --target thumbv7em-none-eabihf --features skip-include --out-dir out/examples/boot/stm32wl \
     --- build --release --manifest-path examples/boot/bootloader/nrf/Cargo.toml --target thumbv7em-none-eabi --features embassy-nrf/nrf52840 \
     --- build --release --manifest-path examples/boot/bootloader/nrf/Cargo.toml --target thumbv8m.main-none-eabihf --features embassy-nrf/nrf9160-ns \
     --- build --release --manifest-path examples/boot/bootloader/rp/Cargo.toml --target thumbv6m-none-eabi \
diff --git a/embassy-boot/boot/src/firmware_updater/asynch.rs b/embassy-boot/boot/src/firmware_updater/asynch.rs
index 0b3f88313..20731ee0a 100644
--- a/embassy-boot/boot/src/firmware_updater/asynch.rs
+++ b/embassy-boot/boot/src/firmware_updater/asynch.rs
@@ -56,6 +56,16 @@ impl<DFU: NorFlash, STATE: NorFlash> FirmwareUpdater<DFU, STATE> {
         }
     }
 
+    // Make sure we are running a booted firmware to avoid reverting to a bad state.
+    async fn verify_booted(&mut self, aligned: &mut [u8]) -> Result<(), FirmwareUpdaterError> {
+        assert_eq!(aligned.len(), STATE::WRITE_SIZE);
+        if self.get_state(aligned).await? == State::Boot {
+            Ok(())
+        } else {
+            Err(FirmwareUpdaterError::BadState)
+        }
+    }
+
     /// Obtain the current state.
     ///
     /// This is useful to check if the bootloader has just done a swap, in order
@@ -98,6 +108,8 @@ impl<DFU: NorFlash, STATE: NorFlash> FirmwareUpdater<DFU, STATE> {
         assert_eq!(_aligned.len(), STATE::WRITE_SIZE);
         assert!(_update_len <= self.dfu.capacity() as u32);
 
+        self.verify_booted(_aligned).await?;
+
         #[cfg(feature = "ed25519-dalek")]
         {
             use ed25519_dalek::{PublicKey, Signature, SignatureError, Verifier};
@@ -217,8 +229,16 @@ impl<DFU: NorFlash, STATE: NorFlash> FirmwareUpdater<DFU, STATE> {
     /// # Safety
     ///
     /// Failing to meet alignment and size requirements may result in a panic.
-    pub async fn write_firmware(&mut self, offset: usize, data: &[u8]) -> Result<(), FirmwareUpdaterError> {
+    pub async fn write_firmware(
+        &mut self,
+        aligned: &mut [u8],
+        offset: usize,
+        data: &[u8],
+    ) -> Result<(), FirmwareUpdaterError> {
         assert!(data.len() >= DFU::ERASE_SIZE);
+        assert_eq!(aligned.len(), STATE::WRITE_SIZE);
+
+        self.verify_booted(aligned).await?;
 
         self.dfu.erase(offset as u32, (offset + data.len()) as u32).await?;
 
@@ -232,7 +252,14 @@ impl<DFU: NorFlash, STATE: NorFlash> FirmwareUpdater<DFU, STATE> {
     ///
     /// Using this instead of `write_firmware` allows for an optimized API in
     /// exchange for added complexity.
-    pub async fn prepare_update(&mut self) -> Result<&mut DFU, FirmwareUpdaterError> {
+    ///
+    /// # Safety
+    ///
+    /// The `aligned` buffer must have a size of STATE::WRITE_SIZE, and follow the alignment rules for the flash being written to.
+    pub async fn prepare_update(&mut self, aligned: &mut [u8]) -> Result<&mut DFU, FirmwareUpdaterError> {
+        assert_eq!(aligned.len(), STATE::WRITE_SIZE);
+        self.verify_booted(aligned).await?;
+
         self.dfu.erase(0, self.dfu.capacity() as u32).await?;
 
         Ok(&mut self.dfu)
@@ -255,13 +282,14 @@ mod tests {
         let flash = Mutex::<NoopRawMutex, _>::new(MemFlash::<131072, 4096, 8>::default());
         let state = Partition::new(&flash, 0, 4096);
         let dfu = Partition::new(&flash, 65536, 65536);
+        let mut aligned = [0; 8];
 
         let update = [0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66];
         let mut to_write = [0; 4096];
         to_write[..7].copy_from_slice(update.as_slice());
 
         let mut updater = FirmwareUpdater::new(FirmwareUpdaterConfig { dfu, state });
-        block_on(updater.write_firmware(0, to_write.as_slice())).unwrap();
+        block_on(updater.write_firmware(&mut aligned, 0, to_write.as_slice())).unwrap();
         let mut chunk_buf = [0; 2];
         let mut hash = [0; 20];
         block_on(updater.hash::<Sha1>(update.len() as u32, &mut chunk_buf, &mut hash)).unwrap();
diff --git a/embassy-boot/boot/src/firmware_updater/blocking.rs b/embassy-boot/boot/src/firmware_updater/blocking.rs
index 551150c4f..f03f53e4d 100644
--- a/embassy-boot/boot/src/firmware_updater/blocking.rs
+++ b/embassy-boot/boot/src/firmware_updater/blocking.rs
@@ -58,6 +58,16 @@ impl<DFU: NorFlash, STATE: NorFlash> BlockingFirmwareUpdater<DFU, STATE> {
         }
     }
 
+    // Make sure we are running a booted firmware to avoid reverting to a bad state.
+    fn verify_booted(&mut self, aligned: &mut [u8]) -> Result<(), FirmwareUpdaterError> {
+        assert_eq!(aligned.len(), STATE::WRITE_SIZE);
+        if self.get_state(aligned)? == State::Boot {
+            Ok(())
+        } else {
+            Err(FirmwareUpdaterError::BadState)
+        }
+    }
+
     /// Obtain the current state.
     ///
     /// This is useful to check if the bootloader has just done a swap, in order
@@ -100,6 +110,8 @@ impl<DFU: NorFlash, STATE: NorFlash> BlockingFirmwareUpdater<DFU, STATE> {
         assert_eq!(_aligned.len(), STATE::WRITE_SIZE);
         assert!(_update_len <= self.dfu.capacity() as u32);
 
+        self.verify_booted(_aligned)?;
+
         #[cfg(feature = "ed25519-dalek")]
         {
             use ed25519_dalek::{PublicKey, Signature, SignatureError, Verifier};
@@ -219,8 +231,15 @@ impl<DFU: NorFlash, STATE: NorFlash> BlockingFirmwareUpdater<DFU, STATE> {
     /// # Safety
     ///
     /// Failing to meet alignment and size requirements may result in a panic.
-    pub fn write_firmware(&mut self, offset: usize, data: &[u8]) -> Result<(), FirmwareUpdaterError> {
+    pub fn write_firmware(
+        &mut self,
+        aligned: &mut [u8],
+        offset: usize,
+        data: &[u8],
+    ) -> Result<(), FirmwareUpdaterError> {
         assert!(data.len() >= DFU::ERASE_SIZE);
+        assert_eq!(aligned.len(), STATE::WRITE_SIZE);
+        self.verify_booted(aligned)?;
 
         self.dfu.erase(offset as u32, (offset + data.len()) as u32)?;
 
@@ -234,7 +253,13 @@ impl<DFU: NorFlash, STATE: NorFlash> BlockingFirmwareUpdater<DFU, STATE> {
     ///
     /// Using this instead of `write_firmware` allows for an optimized API in
     /// exchange for added complexity.
-    pub fn prepare_update(&mut self) -> Result<&mut DFU, FirmwareUpdaterError> {
+    ///
+    /// # Safety
+    ///
+    /// The `aligned` buffer must have a size of STATE::WRITE_SIZE, and follow the alignment rules for the flash being written to.
+    pub fn prepare_update(&mut self, aligned: &mut [u8]) -> Result<&mut DFU, FirmwareUpdaterError> {
+        assert_eq!(aligned.len(), STATE::WRITE_SIZE);
+        self.verify_booted(aligned)?;
         self.dfu.erase(0, self.dfu.capacity() as u32)?;
 
         Ok(&mut self.dfu)
@@ -264,7 +289,8 @@ mod tests {
         to_write[..7].copy_from_slice(update.as_slice());
 
         let mut updater = BlockingFirmwareUpdater::new(FirmwareUpdaterConfig { dfu, state });
-        updater.write_firmware(0, to_write.as_slice()).unwrap();
+        let mut aligned = [0; 8];
+        updater.write_firmware(&mut aligned, 0, to_write.as_slice()).unwrap();
         let mut chunk_buf = [0; 2];
         let mut hash = [0; 20];
         updater
diff --git a/embassy-boot/boot/src/firmware_updater/mod.rs b/embassy-boot/boot/src/firmware_updater/mod.rs
index a37984a3a..55ce8f363 100644
--- a/embassy-boot/boot/src/firmware_updater/mod.rs
+++ b/embassy-boot/boot/src/firmware_updater/mod.rs
@@ -26,6 +26,8 @@ pub enum FirmwareUpdaterError {
     Flash(NorFlashErrorKind),
     /// Signature errors.
     Signature(signature::Error),
+    /// Bad state.
+    BadState,
 }
 
 #[cfg(feature = "defmt")]
@@ -34,6 +36,7 @@ impl defmt::Format for FirmwareUpdaterError {
         match self {
             FirmwareUpdaterError::Flash(_) => defmt::write!(fmt, "FirmwareUpdaterError::Flash(_)"),
             FirmwareUpdaterError::Signature(_) => defmt::write!(fmt, "FirmwareUpdaterError::Signature(_)"),
+            FirmwareUpdaterError::BadState => defmt::write!(fmt, "FirmwareUpdaterError::BadState"),
         }
     }
 }
diff --git a/embassy-boot/boot/src/lib.rs b/embassy-boot/boot/src/lib.rs
index 45a87bd0e..016362b86 100644
--- a/embassy-boot/boot/src/lib.rs
+++ b/embassy-boot/boot/src/lib.rs
@@ -51,6 +51,8 @@ impl<const N: usize> AsMut<[u8]> for AlignedBuffer<N> {
 
 #[cfg(test)]
 mod tests {
+    #![allow(unused_imports)]
+
     use embedded_storage::nor_flash::{NorFlash, ReadNorFlash};
     #[cfg(feature = "nightly")]
     use embedded_storage_async::nor_flash::NorFlash as AsyncNorFlash;
@@ -120,9 +122,13 @@ mod tests {
             dfu: flash.dfu(),
             state: flash.state(),
         });
-        block_on(updater.write_firmware(0, &UPDATE)).unwrap();
+        block_on(updater.write_firmware(&mut aligned, 0, &UPDATE)).unwrap();
         block_on(updater.mark_updated(&mut aligned)).unwrap();
 
+        // Writing after marking updated is not allowed until marked as booted.
+        let res: Result<(), FirmwareUpdaterError> = block_on(updater.write_firmware(&mut aligned, 0, &UPDATE));
+        assert!(matches!(res, Err::<(), _>(FirmwareUpdaterError::BadState)));
+
         let flash = flash.into_blocking();
         let mut bootloader = BootLoader::new(BootLoaderConfig {
             active: flash.active(),
@@ -188,7 +194,7 @@ mod tests {
             dfu: flash.dfu(),
             state: flash.state(),
         });
-        block_on(updater.write_firmware(0, &UPDATE)).unwrap();
+        block_on(updater.write_firmware(&mut aligned, 0, &UPDATE)).unwrap();
         block_on(updater.mark_updated(&mut aligned)).unwrap();
 
         let flash = flash.into_blocking();
@@ -230,7 +236,7 @@ mod tests {
             dfu: flash.dfu(),
             state: flash.state(),
         });
-        block_on(updater.write_firmware(0, &UPDATE)).unwrap();
+        block_on(updater.write_firmware(&mut aligned, 0, &UPDATE)).unwrap();
         block_on(updater.mark_updated(&mut aligned)).unwrap();
 
         let flash = flash.into_blocking();
diff --git a/examples/boot/application/nrf/Cargo.toml b/examples/boot/application/nrf/Cargo.toml
index 5939a43b1..b98f73f39 100644
--- a/examples/boot/application/nrf/Cargo.toml
+++ b/examples/boot/application/nrf/Cargo.toml
@@ -24,3 +24,4 @@ cortex-m-rt = "0.7.0"
 [features]
 ed25519-dalek = ["embassy-boot/ed25519-dalek"]
 ed25519-salty = ["embassy-boot/ed25519-salty"]
+skip-include = []
diff --git a/examples/boot/application/nrf/src/bin/a.rs b/examples/boot/application/nrf/src/bin/a.rs
index 06c237781..021d77f3b 100644
--- a/examples/boot/application/nrf/src/bin/a.rs
+++ b/examples/boot/application/nrf/src/bin/a.rs
@@ -12,6 +12,9 @@ use embassy_nrf::wdt::{self, Watchdog};
 use embassy_sync::mutex::Mutex;
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
@@ -55,13 +58,13 @@ async fn main(_spawner: Spawner) {
         button.wait_for_any_edge().await;
         if button.is_low() {
             let mut offset = 0;
+            let mut magic = [0; 4];
             for chunk in APP_B.chunks(4096) {
                 let mut buf: [u8; 4096] = [0; 4096];
                 buf[..chunk.len()].copy_from_slice(chunk);
-                updater.write_firmware(offset, &buf).await.unwrap();
+                updater.write_firmware(&mut magic, offset, &buf).await.unwrap();
                 offset += chunk.len();
             }
-            let mut magic = [0; 4];
             updater.mark_updated(&mut magic).await.unwrap();
             led.set_high();
             cortex_m::peripheral::SCB::sys_reset();
diff --git a/examples/boot/application/rp/Cargo.toml b/examples/boot/application/rp/Cargo.toml
index 4a2c5dd8f..007b6839c 100644
--- a/examples/boot/application/rp/Cargo.toml
+++ b/examples/boot/application/rp/Cargo.toml
@@ -29,6 +29,7 @@ debug = [
     "embassy-boot-rp/defmt",
     "panic-probe"
 ]
+skip-include = []
 
 [profile.release]
 debug = true
diff --git a/examples/boot/application/rp/src/bin/a.rs b/examples/boot/application/rp/src/bin/a.rs
index 69850069b..c8497494c 100644
--- a/examples/boot/application/rp/src/bin/a.rs
+++ b/examples/boot/application/rp/src/bin/a.rs
@@ -18,7 +18,11 @@ use panic_probe as _;
 #[cfg(feature = "panic-reset")]
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
+
 const FLASH_SIZE: usize = 2 * 1024 * 1024;
 
 #[embassy_executor::main]
@@ -43,7 +47,7 @@ async fn main(_s: Spawner) {
     let mut buf: AlignedBuffer<4096> = AlignedBuffer([0; 4096]);
     defmt::info!("preparing update");
     let writer = updater
-        .prepare_update()
+        .prepare_update(&mut buf.0[..1])
         .map_err(|e| defmt::warn!("E: {:?}", defmt::Debug2Format(&e)))
         .unwrap();
     defmt::info!("writer created, starting write");
diff --git a/examples/boot/application/stm32f3/Cargo.toml b/examples/boot/application/stm32f3/Cargo.toml
index 24abd90d4..5b3faf8f8 100644
--- a/examples/boot/application/stm32f3/Cargo.toml
+++ b/examples/boot/application/stm32f3/Cargo.toml
@@ -26,3 +26,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32f3/src/bin/a.rs b/examples/boot/application/stm32f3/src/bin/a.rs
index c94676f09..c0a11d699 100644
--- a/examples/boot/application/stm32f3/src/bin/a.rs
+++ b/examples/boot/application/stm32f3/src/bin/a.rs
@@ -13,6 +13,9 @@ use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
 use embassy_sync::mutex::Mutex;
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
@@ -31,13 +34,13 @@ async fn main(_spawner: Spawner) {
     let mut updater = FirmwareUpdater::new(config);
     button.wait_for_falling_edge().await;
     let mut offset = 0;
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     for chunk in APP_B.chunks(2048) {
         let mut buf: [u8; 2048] = [0; 2048];
         buf[..chunk.len()].copy_from_slice(chunk);
-        updater.write_firmware(offset, &buf).await.unwrap();
+        updater.write_firmware(magic.as_mut(), offset, &buf).await.unwrap();
         offset += chunk.len();
     }
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     updater.mark_updated(magic.as_mut()).await.unwrap();
     led.set_low();
     cortex_m::peripheral::SCB::sys_reset();
diff --git a/examples/boot/application/stm32f7/Cargo.toml b/examples/boot/application/stm32f7/Cargo.toml
index 529a01aad..b6a6f9cd8 100644
--- a/examples/boot/application/stm32f7/Cargo.toml
+++ b/examples/boot/application/stm32f7/Cargo.toml
@@ -16,6 +16,7 @@ defmt = { version = "0.3", optional = true }
 defmt-rtt = { version = "0.4", optional = true }
 panic-reset = { version = "0.1.1" }
 embedded-hal = { version = "0.2.6" }
+embedded-storage = "0.3.0"
 
 cortex-m = { version = "0.7.6", features = ["inline-asm", "critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
@@ -26,3 +27,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32f7/src/bin/a.rs b/examples/boot/application/stm32f7/src/bin/a.rs
index fc2702c91..dea682a96 100644
--- a/examples/boot/application/stm32f7/src/bin/a.rs
+++ b/examples/boot/application/stm32f7/src/bin/a.rs
@@ -2,6 +2,8 @@
 #![no_main]
 #![feature(type_alias_impl_trait)]
 
+use core::cell::RefCell;
+
 #[cfg(feature = "defmt-rtt")]
 use defmt_rtt::*;
 use embassy_boot_stm32::{AlignedBuffer, BlockingFirmwareUpdater, FirmwareUpdaterConfig};
@@ -9,8 +11,13 @@ use embassy_executor::Spawner;
 use embassy_stm32::exti::ExtiInput;
 use embassy_stm32::flash::{Flash, WRITE_SIZE};
 use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_sync::blocking_mutex::Mutex;
+use embedded_storage::nor_flash::NorFlash;
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
@@ -27,16 +34,16 @@ async fn main(_spawner: Spawner) {
 
     let config = FirmwareUpdaterConfig::from_linkerfile_blocking(&flash);
     let mut updater = BlockingFirmwareUpdater::new(config);
-    let mut writer = updater.prepare_update().unwrap();
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
+    let writer = updater.prepare_update(magic.as_mut()).unwrap();
     button.wait_for_rising_edge().await;
     let mut offset = 0;
     let mut buf = AlignedBuffer([0; 4096]);
     for chunk in APP_B.chunks(4096) {
         buf.as_mut()[..chunk.len()].copy_from_slice(chunk);
         writer.write(offset, buf.as_ref()).unwrap();
-        offset += chunk.len();
+        offset += chunk.len() as u32;
     }
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     updater.mark_updated(magic.as_mut()).unwrap();
     led.set_low();
     cortex_m::peripheral::SCB::sys_reset();
diff --git a/examples/boot/application/stm32h7/Cargo.toml b/examples/boot/application/stm32h7/Cargo.toml
index d7539a53f..0a7e19b1d 100644
--- a/examples/boot/application/stm32h7/Cargo.toml
+++ b/examples/boot/application/stm32h7/Cargo.toml
@@ -16,6 +16,7 @@ defmt = { version = "0.3", optional = true }
 defmt-rtt = { version = "0.4", optional = true }
 panic-reset = { version = "0.1.1" }
 embedded-hal = { version = "0.2.6" }
+embedded-storage = "0.3.0"
 
 cortex-m = { version = "0.7.6", features = ["inline-asm", "critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
@@ -26,3 +27,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32h7/src/bin/a.rs b/examples/boot/application/stm32h7/src/bin/a.rs
index 1a54464d0..719176692 100644
--- a/examples/boot/application/stm32h7/src/bin/a.rs
+++ b/examples/boot/application/stm32h7/src/bin/a.rs
@@ -2,6 +2,8 @@
 #![no_main]
 #![feature(type_alias_impl_trait)]
 
+use core::cell::RefCell;
+
 #[cfg(feature = "defmt-rtt")]
 use defmt_rtt::*;
 use embassy_boot_stm32::{AlignedBuffer, BlockingFirmwareUpdater, FirmwareUpdaterConfig};
@@ -9,8 +11,13 @@ use embassy_executor::Spawner;
 use embassy_stm32::exti::ExtiInput;
 use embassy_stm32::flash::{Flash, WRITE_SIZE};
 use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_sync::blocking_mutex::Mutex;
+use embedded_storage::nor_flash::NorFlash;
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
@@ -26,17 +33,17 @@ async fn main(_spawner: Spawner) {
     led.set_high();
 
     let config = FirmwareUpdaterConfig::from_linkerfile_blocking(&flash);
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     let mut updater = BlockingFirmwareUpdater::new(config);
-    let mut writer = updater.prepare_update().unwrap();
+    let writer = updater.prepare_update(magic.as_mut()).unwrap();
     button.wait_for_rising_edge().await;
     let mut offset = 0;
     let mut buf = AlignedBuffer([0; 4096]);
     for chunk in APP_B.chunks(4096) {
         buf.as_mut()[..chunk.len()].copy_from_slice(chunk);
         writer.write(offset, buf.as_ref()).unwrap();
-        offset += chunk.len();
+        offset += chunk.len() as u32;
     }
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     updater.mark_updated(magic.as_mut()).unwrap();
     led.set_low();
     cortex_m::peripheral::SCB::sys_reset();
diff --git a/examples/boot/application/stm32l0/Cargo.toml b/examples/boot/application/stm32l0/Cargo.toml
index e90da259b..998df4dc0 100644
--- a/examples/boot/application/stm32l0/Cargo.toml
+++ b/examples/boot/application/stm32l0/Cargo.toml
@@ -26,3 +26,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32l0/src/bin/a.rs b/examples/boot/application/stm32l0/src/bin/a.rs
index 4033ac590..ce80056e6 100644
--- a/examples/boot/application/stm32l0/src/bin/a.rs
+++ b/examples/boot/application/stm32l0/src/bin/a.rs
@@ -4,22 +4,26 @@
 
 #[cfg(feature = "defmt-rtt")]
 use defmt_rtt::*;
-use embassy_boot_stm32::{AlignedBuffer, FirmwareUpdater};
+use embassy_boot_stm32::{AlignedBuffer, FirmwareUpdater, FirmwareUpdaterConfig};
 use embassy_embedded_hal::adapter::BlockingAsync;
 use embassy_executor::Spawner;
 use embassy_stm32::exti::ExtiInput;
 use embassy_stm32::flash::{Flash, WRITE_SIZE};
 use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_sync::mutex::Mutex;
 use embassy_time::{Duration, Timer};
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     let flash = Flash::new_blocking(p.FLASH);
-    let mut flash = BlockingAsync::new(flash);
+    let flash = Mutex::new(BlockingAsync::new(flash));
 
     let button = Input::new(p.PB2, Pull::Up);
     let mut button = ExtiInput::new(button, p.EXTI2);
@@ -28,18 +32,19 @@ async fn main(_spawner: Spawner) {
 
     led.set_high();
 
-    let mut updater = FirmwareUpdater::default();
+    let config = FirmwareUpdaterConfig::from_linkerfile(&flash);
+    let mut updater = FirmwareUpdater::new(config);
     button.wait_for_falling_edge().await;
     let mut offset = 0;
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     for chunk in APP_B.chunks(128) {
         let mut buf: [u8; 128] = [0; 128];
         buf[..chunk.len()].copy_from_slice(chunk);
-        updater.write_firmware(offset, &buf, &mut flash, 128).await.unwrap();
+        updater.write_firmware(magic.as_mut(), offset, &buf).await.unwrap();
         offset += chunk.len();
     }
 
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
-    updater.mark_updated(&mut flash, magic.as_mut()).await.unwrap();
+    updater.mark_updated(magic.as_mut()).await.unwrap();
     led.set_low();
     Timer::after(Duration::from_secs(1)).await;
     cortex_m::peripheral::SCB::sys_reset();
diff --git a/examples/boot/application/stm32l1/Cargo.toml b/examples/boot/application/stm32l1/Cargo.toml
index 8ac0fac85..10b58c172 100644
--- a/examples/boot/application/stm32l1/Cargo.toml
+++ b/examples/boot/application/stm32l1/Cargo.toml
@@ -26,3 +26,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32l1/src/bin/a.rs b/examples/boot/application/stm32l1/src/bin/a.rs
index 00ddda636..1e9bf3cb9 100644
--- a/examples/boot/application/stm32l1/src/bin/a.rs
+++ b/examples/boot/application/stm32l1/src/bin/a.rs
@@ -10,9 +10,13 @@ use embassy_executor::Spawner;
 use embassy_stm32::exti::ExtiInput;
 use embassy_stm32::flash::{Flash, WRITE_SIZE};
 use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_sync::mutex::Mutex;
 use embassy_time::{Duration, Timer};
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
@@ -31,15 +35,15 @@ async fn main(_spawner: Spawner) {
     let config = FirmwareUpdaterConfig::from_linkerfile(&flash);
     let mut updater = FirmwareUpdater::new(config);
     button.wait_for_falling_edge().await;
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     let mut offset = 0;
     for chunk in APP_B.chunks(128) {
         let mut buf: [u8; 128] = [0; 128];
         buf[..chunk.len()].copy_from_slice(chunk);
-        updater.write_firmware(offset, &buf).await.unwrap();
+        updater.write_firmware(magic.as_mut(), offset, &buf).await.unwrap();
         offset += chunk.len();
     }
 
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     updater.mark_updated(magic.as_mut()).await.unwrap();
     led.set_low();
     Timer::after(Duration::from_secs(1)).await;
diff --git a/examples/boot/application/stm32l4/Cargo.toml b/examples/boot/application/stm32l4/Cargo.toml
index ec79acdeb..713a6527e 100644
--- a/examples/boot/application/stm32l4/Cargo.toml
+++ b/examples/boot/application/stm32l4/Cargo.toml
@@ -26,3 +26,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32l4/src/bin/a.rs b/examples/boot/application/stm32l4/src/bin/a.rs
index 54579e4ac..a514ab5be 100644
--- a/examples/boot/application/stm32l4/src/bin/a.rs
+++ b/examples/boot/application/stm32l4/src/bin/a.rs
@@ -10,8 +10,12 @@ use embassy_executor::Spawner;
 use embassy_stm32::exti::ExtiInput;
 use embassy_stm32::flash::{Flash, WRITE_SIZE};
 use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_sync::mutex::Mutex;
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
@@ -29,15 +33,15 @@ async fn main(_spawner: Spawner) {
     let config = FirmwareUpdaterConfig::from_linkerfile(&flash);
     let mut updater = FirmwareUpdater::new(config);
     button.wait_for_falling_edge().await;
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     let mut offset = 0;
     for chunk in APP_B.chunks(2048) {
         let mut buf: [u8; 2048] = [0; 2048];
         buf[..chunk.len()].copy_from_slice(chunk);
-        updater.write_firmware(offset, &buf).await.unwrap();
+        updater.write_firmware(magic.as_mut(), offset, &buf).await.unwrap();
         offset += chunk.len();
     }
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
-    updater.mark_updated(&mut flash, magic.as_mut()).await.unwrap();
+    updater.mark_updated(magic.as_mut()).await.unwrap();
     led.set_low();
     cortex_m::peripheral::SCB::sys_reset();
 }
diff --git a/examples/boot/application/stm32wl/Cargo.toml b/examples/boot/application/stm32wl/Cargo.toml
index dfaece6cf..4c8bbd73f 100644
--- a/examples/boot/application/stm32wl/Cargo.toml
+++ b/examples/boot/application/stm32wl/Cargo.toml
@@ -26,3 +26,4 @@ defmt = [
       "embassy-stm32/defmt",
       "embassy-boot-stm32/defmt",
 ]
+skip-include = []
diff --git a/examples/boot/application/stm32wl/src/bin/a.rs b/examples/boot/application/stm32wl/src/bin/a.rs
index 0c6fa05f9..52a197a5c 100644
--- a/examples/boot/application/stm32wl/src/bin/a.rs
+++ b/examples/boot/application/stm32wl/src/bin/a.rs
@@ -4,21 +4,25 @@
 
 #[cfg(feature = "defmt-rtt")]
 use defmt_rtt::*;
-use embassy_boot_stm32::{AlignedBuffer, FirmwareUpdater};
+use embassy_boot_stm32::{AlignedBuffer, FirmwareUpdater, FirmwareUpdaterConfig};
 use embassy_embedded_hal::adapter::BlockingAsync;
 use embassy_executor::Spawner;
 use embassy_stm32::exti::ExtiInput;
 use embassy_stm32::flash::{Flash, WRITE_SIZE};
 use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_sync::mutex::Mutex;
 use panic_reset as _;
 
+#[cfg(feature = "skip-include")]
+static APP_B: &[u8] = &[0, 1, 2, 3];
+#[cfg(not(feature = "skip-include"))]
 static APP_B: &[u8] = include_bytes!("../../b.bin");
 
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     let flash = Flash::new_blocking(p.FLASH);
-    let mut flash = Mutex::new(BlockingAsync::new(flash));
+    let flash = Mutex::new(BlockingAsync::new(flash));
 
     let button = Input::new(p.PA0, Pull::Up);
     let mut button = ExtiInput::new(button, p.EXTI0);
@@ -30,15 +34,15 @@ async fn main(_spawner: Spawner) {
     let mut updater = FirmwareUpdater::new(config);
     button.wait_for_falling_edge().await;
     //defmt::info!("Starting update");
+    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     let mut offset = 0;
     for chunk in APP_B.chunks(2048) {
         let mut buf: [u8; 2048] = [0; 2048];
         buf[..chunk.len()].copy_from_slice(chunk);
         //        defmt::info!("Writing chunk at 0x{:x}", offset);
-        updater.write_firmware(offset, &buf).await.unwrap();
+        updater.write_firmware(magic.as_mut(), offset, &buf).await.unwrap();
         offset += chunk.len();
     }
-    let mut magic = AlignedBuffer([0; WRITE_SIZE]);
     updater.mark_updated(magic.as_mut()).await.unwrap();
     //defmt::info!("Marked as updated");
     led.set_low();