In the Hangman I know, there is a limit to the number of incorrect guesses :-), and a little diagram to show how close to failing you are...
I've looked at your program, and can't break it (except that the program's method of stopping is to "fall off the end", which I've not tried but I assume must be OK in GnuCOBOL 2), so that's good, but it is only a short program. There's not much particularly "COBOL" about it, which is a criticism of the task chosen and its implementation, rather than the code.
There are no "rules" about how to write your own COBOL programs, but "in action" you would have some local standards about how things are done. These standards are rarely perfect, but at least other members of the team can pick up someone's program and not have to understand the individual style.
Data-names/identifiers and labels: Make these descriptive; group definitions together which are related
Data-types: PIC A(...) does not get used except by beginners; don't use PIC 9(...) just because something is "a number", unless it is used where a numeric definition is required, use PIC X(...)
Data sizes, descriptions and validation: use what is needed, no more, and no less. That PIC A(100) will confuse a reader. What is that for? Is it really only a WORD? Or slackly used for something else? Limit it, and describe it to your user ("You have to guess a word which is a maximum of N letters long"). You have 256 entries for
guesses when only 52 are needed (upper- and lower-case letters). Why process user-input which is not valid?
Formatting/indenting: use it well and consistently. I arrange things more like @Kwebble mentioned in their comment. Use your editor effectively to aid you to write code for the human reader.
Full-stops/periods in the PROCEDURE DIVISION: as was mentioned, keep these to a minimum, and don't attach any to actual code. This will ease copy/pasting of code, as you will never have to decide whether the new location requires the full-stop/period or not.
88-level/condition-names: use them. Don't use literals in the PROCEDURE DIVISION for tests, except for the trivial. When using 88-levels/condition-names in tests, you also have the option to use SET with that 88-level to change the value stored.
COBOL is a language of fixed-length fields (unless using variable-length fields). COBOL does not have strings. If you want to know the length of the data, you have to calculate it yourself by counting the trailing spaces. FUNCTION LENGTH will tell you the length of a data-item, and you work backwards from that.
01 word pic X(100).
01 word-length pic 9(3).
01 guess pic X.
88 clear-to-no-guesses VALUE ZERO.
05 FILLER occurs 256 times.
10 FILLER PIC X.
88 letter-guessed VALUE "1".
01 FILLER pic X.
88 done VALUE "Q".
88 done-not VALUE "7".
01 FILLER pic X.
88 no-missing-letters VALUE ":".
88 missing-letter VALUE "3".
01 i pic 9(3).
* TODO: pick random word from word list
move "hello" to word
move 5 to word-length
* TODO: show this in debug mode only
display "word: " word
set clear-to-no-guesses TO TRUE
set done-not TO TRUE
perform until done
( function ord ( guess ) )
SET no-missing-letters TO TRUE
MOVE ZERO TO i
add 1 TO i
( function ord ( word ( i : 1 ) ) )
display word ( i : 1 ) with no advancing
display "_" with no advancing
display " "
SET done TO TRUE
Note I've replaced the VARYING by TIMES. There is nothing "varying" about the number of iterations for any given execution of the PERFORM, so TIMES better describes the loop. And it is faster. You have to code the initialisation and use of the subscripting item, so many people wouldn't do that. They'd go for less typing that better description.
Note the VALUE clauses on the 88s for the flags. The value does not matter (as long as they are different). The flags are only referenced in conditions and given a value by SET. In the real world, you'd use Y and N (or similar) on the VALUEs, but those values have no intrinsic meaning to the code. They are just values.
No literals with "meaning" scattered through the procedure code.
I'm not a fan of reference-modification, because it obscures. Here, you could REDEFINES word with OCCURS and use subscripting, giving the elements of word, and the data-name or index used for subscripting, a meaningful name. Not leave someone thinking "what does
word ( i : 1 ) actually indicate?" (not so important here, but often is).
Note also that you use word ( i : 1 ) twice. Instead, you could MOVE word ( i : 1 ) to a new data-item, and use that subsequently.
You can write your COBOL programs as you have done, for yourself, and have fun. If you got a job writing COBOL programs, you'd be in for a bit of a change in the way you'd have to do things.
You can already start to "read" the program. A few more data-name changes, a very simple thing to do, and you'll see further readability.
Making a program readable means a programmer can understand it better. We don't expect a non-programmer to be able to make genuine sense of COBOL code. The more you simplify, the more more you describe the data, and the structure of the program, with the code, the less likely you are to confuse the programmer looking for an obscure error at 2am or wanting to implement a change to a program to timescales.
MOVE VAR1 ( i : j ) TO VAR3
The latter is COBOL, the former is... make up you own mind whether you'd like to encounter that, meaning the same thing...