diff --git a/docs/README.md b/docs/README.md
index 0ce8210f..21917873 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -79,6 +79,14 @@
     * [Game Boy–only modules](stdlib/gb.md)
     
     * [X16–only modules](stdlib/x16.md)
+    
+## Guides
+
+* [Differences from C](various/cdiff.md) 
+
+* [Differences from other assemblers](various/asmdiff.md) 
+
+* [Optimization guide](various/optimization.md) 
 
 ## Implementation details
 
diff --git a/docs/doc_index.md b/docs/doc_index.md
index 7450d0cd..42b40102 100644
--- a/docs/doc_index.md
+++ b/docs/doc_index.md
@@ -81,6 +81,14 @@
     * [Game Boy–only modules](stdlib/gb.md)
     
     * [X16–only modules](stdlib/x16.md)
+    
+## Guides
+
+* [Differences from C](various/cdiff.md) 
+
+* [Differences from other assemblers](various/asmdiff.md) 
+
+* [Optimization guide](various/optimization.md) 
 
 ## Implementation details
 
diff --git a/docs/various/asmdiff.md b/docs/various/asmdiff.md
new file mode 100644
index 00000000..e3efc369
--- /dev/null
+++ b/docs/various/asmdiff.md
@@ -0,0 +1,34 @@
+[< back to index](../doc_index.md)
+
+# Differences in assembly
+
+## Syntax
+
+* High and low bytes of an 16-bit value are acquired using the `hi` and `lo` functions, not the `>` and `<` operators.
+
+* Anonymous labels and local labels are not supported.  
+All labels defined in assembly are global.   
+Colons are required in label declarations.
+
+* Macros are inserted using the `+` operator.
+
+* Raw bytes are inserted using the Millfork array syntax, not with any pseudoopcode (like `!byte`, `db` or `fcb`)
+
+* Assembly blocks cannot contain definitions of constants or variables.
+
+* 6502: To enforce zero-page addressing, wrap the argument in the `lo` function: `lo(arg)`
+
+* 6502: To enforce absolute addressing, add a 16-bit zero to the argument: `0000 + arg`
+
+* GameBoy: The $FF page loads/stores are written `LDH (C),A`, not `LD ($FF00+C),A`.
+
+* GameBoy: The loads/stores that postincrement/postdecrement HL write the HL register as `HLI` or `HLD`, not `HL+` or `HL-`
+
+* Z80: Indexing using the index register uses the `IX(1)` syntax, not `(IX+1)` or `1(IX)`.
+
+* Z80: Most undocumented instructions are not supported. The only one supported is `SLL`.
+
+* 6809: `0,X` is assembled the same as `,X`.
+
+  
+
diff --git a/docs/various/cdiff.md b/docs/various/cdiff.md
new file mode 100644
index 00000000..0dfee5a7
--- /dev/null
+++ b/docs/various/cdiff.md
@@ -0,0 +1,78 @@
+[< back to index](../doc_index.md)
+
+# Differences from C
+
+## Syntax
+
+* Block comments `/* */` are not supported, use line comments `//`.
+
+* You cannot put multiple statements on one line.  
+Semicolons are allowed at the end of the line, but no code can follow them.
+
+* There are no `++` or `--` operators. Use `+= 1` and `-= 1`.
+
+* Pointer types are declared using the `pointer.` prefix, not `*` suffix. To dereference them, use `p[0]` instead of `*p`.
+
+* There is no unary `&` operator.
+Pointers to an object are acquired using the `.pointer` suffix.
+Raw addresses are acquired using the `.addr` suffix.
+The numeric values of the pointer and of the raw address may differ.
+
+* Operator precedence works differently.
+Bitwise and bitshift operators have the same precedence as arithmetic operators,
+and mixing different operators with the same precedence is usually forbidden.
+This prevents most ambiguities in bit-twiddling code, but requires care when porting code from or to C. 
+
+* There is no `!` operator. The negation is expressed as the `not` function.
+
+* The modulo operator is written `%%`. As with `/`, it's only defined for unsigned division. 
+
+* The `for` loops are range-based. Arbitrary, C-like `for` loops are not supported.
+
+* Variable declaration and initialization have to be separate.
+
+* Integer literals starting with zero and containing just digits are decimal, not octal.
+For octal literals, use the `0o` prefix.
+
+* String literals are not null-terminated by default. Use the `z` suffix for null-terminated strings.
+
+* In `if`, `do/while`, `while` and `for` statements, parentheses are not required, but braces are.
+The `else` branch also requires braces, unless the only statement in the `else` block is an `if` statement.
+
+* There has to be a space between many operators and character literals,
+as the parser may treat the apostrophe as a part of the operator.
+
+## Preprocessor
+
+* The preprocessor cannot expand symbols to something more complex than an identifier or a literal.
+
+* The preprocessor cannot define symbols for use in other files.
+
+* The closest to C's `#define` is `#use`.
+Unlike `#define`, such definitions are not visible within the preprocessor.
+
+* Directives like `#use` and `#pragma` affect the entire file, not just the code after the directive.
+
+* The preprocessor cannot include files.
+
+## Semantics
+
+* There is no automatic integer promotion. An operation of two bytes will yield a byte. For example:
+
+        byte a
+        byte b
+        word w
+        w = a * b
+
+    may yield unexpected results. To prevent this, cast at least one argument:
+    
+        w = word(a) * b
+
+    This issue applies mostly to the `*` and `<<` operators.
+
+* There is no padding in structs.
+
+
+    
+
+  
diff --git a/docs/various/faq.md b/docs/various/faq.md
index c6bfb9f6..27e38565 100644
--- a/docs/various/faq.md
+++ b/docs/various/faq.md
@@ -83,6 +83,10 @@ or generate a static binary and link it manually using the `file` directive.
 
 Since the compiler is a work-in-progress, some of the mentioned issues might be improved upon in the future.
 
+### I have experience with C and/or assembly. What should I keep in mind?
+
+See the [differences from C](./cdiff.md) and the [differences from assembly](./asmdiff.md).
+
 ### Why is it called Millfork?
 
 It stands for **MI**ddle **L**evel **L**anguage **FOR** **K**ommodore computers.
diff --git a/docs/various/optimization.md b/docs/various/optimization.md
new file mode 100644
index 00000000..15e8fa94
--- /dev/null
+++ b/docs/various/optimization.md
@@ -0,0 +1,103 @@
+[< back to index](../doc_index.md)
+
+# Optimization guidelines
+
+## Command-line options
+
+* The default options provide literally no optimizations.
+Consider using at least `-O1` for quick compilation and `-O4` for release builds.
+
+* Inlining can drastically improve performance. Add `-finline` to the command line.
+
+* If you're not using self-modifying code or code generation,
+enabling interprocedural optimizations (`-fipo`) and stdlib optimizations (`-foptimize-stdlib`) can also help.
+
+* For convenience, all options useful for debug builds can be enabled with `-Xd`,
+and for release builds with `-Xr`.
+
+* 6502 only: If you are sure the target will have a CPU that supports so-called illegal/undocumented 6502 instructions,
+consider adding the `-fillegals` option. Good examples of such targets are NES and C64.
+
+## Alignment
+
+* Consider adding `align(fast)` or even `align(256)` to arrays which you want to access quickly.
+
+* 6502 only: Consider adding `align(fast)` to the hottest functions.
+
+* If you have an array of structs, consider adding `align(X)` to the definition of the struct,
+where `X` is a power of two. Even if this makes the struct 12 bytes instead of 11, it can still improve performance.
+
+## Variables
+
+* Use the smallest type you need. Note that Millfork supports integers of any size from 1 to 16 bytes.
+
+* Consider using multiple arrays instead of arrays of structs.
+
+* Avoid reusing temporary variables.
+It makes it easier for the optimizer to eliminate the variable entirely.   
+
+## Functions
+
+* Write many functions with no parameters and use `-finline`.
+This will simplify the job for the optimizer and increase the chances of certain powerful optimizations to apply.
+
+* Avoid passing many parameters to functions.
+Try to minimize the number of bytes passed as parameters and returned as return values.
+
+## Loops
+
+* For `for` loops that use a byte-sized variable and whose body does not involve function calls or further loops,
+use a unique iteration variable. Such variable will have a bigger chance of being stored in a CPU register.  
+For example:
+
+        byte i
+        byte j
+        for i,0,until,30 { .... }
+        for j,0,until,40 { .... }
+
+    is usually better than:
+    
+        byte i
+        for i,0,until,30 { .... }
+        for i,0,until,40 { .... }
+
+* 8080/Z80 only: The previous tip applies also for loops using word-sized variables.
+
+* When the iteration order is not important, use `paralleluntil` or `parallelto`.
+The compiler will try to choose the optimal iteration order.
+
+* Since 0.3.18: When the iteration order is not important,
+use `for ix,ptr:array` to iterate over arrays of structs.
+
+* 6502 only: When iterating over an array larger than 256 bytes, whose element count is a composite number,
+consider splitting it into less-than-256-byte sized slices and use them within the same iteration.
+For example, instead of:
+
+        word i
+        for i,0,paralleluntil,1000 {
+           screen[i] = ' 'scr
+        }
+
+    consider:
+            
+        byte i
+        for i,0,paralleluntil,250 { 
+            screen[i+000] = ' 'scr
+            screen[i+250] = ' 'scr
+            screen[i+500] = ' 'scr
+            screen[i+750] = ' 'scr
+        }
+        
+    Note that the compiler might do this optimization automatically
+    for simpler loops with certain iteration ranges, but it is not guaranteed.
+
+# Arithmetic
+
+* Avoid 16-bit arithmetic. Try to keep calculations 8-bit for as long as you can.
+If you can calculate the upper and lower byte of a 16-bit value separately, it's usually better to do so.
+
+* Avoid arithmetic larger than 16-bit.
+
+* Use `nonet` if you are sure that the result of shifting will fit into 9 bits.
+Use `nonet` when doing byte addition that you want to promote to a word.
+