Retromulticsshellcomputing-history
Louis Pouzin Named the Shell, Then Colleagues Built It From His Flowchart After He Left
Overnight maintenance jobs pushed commands toward a reusable language, while the surviving 1965 design notes reveal how much human error its author expected.

The shell’s designer left the country. The flowchart stayed behind, which turned out to be enough for somebody else to get the software running.
In his first-person history, Louis Pouzin traces the idea through his work on CTSS in 1963 and 1964. He wanted to leave long jobs running overnight. RUNCOM let saved command sequences accept arguments, so a person could reuse a sequence with different inputs rather than sit there typing the evening away.
The ambition was to combine commands “just like subroutine libraries,” Pouzin wrote. His work in the MAD language had also exposed duplicated effort between commands and the routines programs could call. He wanted them to share more machinery.
Pouzin dates the name shell to late 1964 or early 1965. While colleagues debated syntax, he drew the execution flow before returning to France in mid-1965. His recollection credits Glenda Schroeder at MIT and an unnamed GE colleague with building the first Multics shell from it after he left.
The April 2, 1965 design notebook gives the diagram some substance. The shell would sit between a typed console message and a callable procedure. The first word named the command; the remaining arguments supplied its inputs. The paper also proposed letting users replace the shell with a private version. Even the thing that interpreted your requests could be another thing you requested.
Pouzin expected people to mistype, omit arguments and interrupt commands. His requirements said output should make sense without a decoding table and errors should explain what the user needed to do next. A command line in 1965 was already expected to offer more help than a mysterious number and a shrug.
His April 7 RUNCOM paper explained how a named, stored sequence could act like a single procedure. Fixed parameters stayed in the sequence; substitutable arguments supplied the details that changed each run. One recipe, different quantities, without rewriting dinner from scratch.
The first implementation was hardly the end. In Tom Van Vleck’s firsthand account, early EPL versions ran so slowly under the 645 simulator that the team created a tiny mini-shell for simple commands without shell metacharacters. A leading period sent a request through that faster route. Users effectively had an express lane, provided their command qualified.
Bill Southworth then demonstrated a processor based on Calvin Mooers’ TRAC language. Van Vleck says the team ultimately produced a third EPL implementation with simpler syntax and adequate performance, and Bob Daley had a major influence on it. Multics replaced the word shell with command processor; the original name later returned in Unix.
A future automation tool could pass the same practical test: can another person run the saved work, change its inputs and understand a failure without having the designer beside them? Pouzin left an implementation path that other people could follow. His account still leaves one attribution unfinished: the name of Schroeder’s GE colleague.