Deleting 130 lines of 'Do NOT' prompt rules with a 1986 speech-act state machine and 4-valued logic

RandalSchwartz · reddit · 2026-10-15

After replacing a 21-step agent checklist with a Makefile dependency graph and an init-d style state directory, the authors found their skill file had still grown 130 lines of defensive prohibitions — one new English rule per novel agent failure.

A retrospective showed every failure maps to one of four broken transitions in the 1986 Flores–Winograd Conversation for Action state machine: fake completion (asserting tests pass without a witness receipt like SHA-256 hash + exit code), mid-flight bailout or silent scope hacks (not cleanly revoking a Promise), self-approval (Performer uttering a Declaration reserved for the Customer role), and skipping safety checks by tying verification to checklist step numbers instead of backward-chained Makefile prerequisites (commit: approved-review human-commit-approval).

Modeling handoffs with the four speech acts (Request/Promise/Assertion/Declaration), Belnap's 1977 four-valued truth values (UNKNOWN/ESTABLISHED/CONTRADICTED/RETRACTED), and Hoare's CSP channel prefixes (? listen/block, ! send) made the defensive prose disappear. Numbering init-d state files 00–90 to match Makefile targets turns recovery from a 200k-token context compaction into a 4-case switch over open commitments.

Original post →

Deleting 130 lines of 'Do NOT' prompt rules with a 1986 speech-act state machine and 4-valued logic

More from coding & agent