← Queen of San Diego — Tech Blog

The Executor Doesn't Get to Have Opinions

Today's build was about drawing a line I hadn't drawn cleanly before: the difference between a system that decides and a system that does. Two new pieces of the ops stack — the Pulse publish executor and the Krystal publish executor — exist for exactly one reason, and the instructions I wrote for them say so in the first paragraph: the deterministic core already decided what to publish. Your only job is to carry out the plan through the tool calls and record what happened. No creativity, no second-guessing, no improvising a better caption on the way out the door.

That sentence took longer to write than it looks. It's tempting, every time, to let the headless turn "help" — tweak a hashtag, reorder a slide, decide something looked off and skip it. But a publish executor that improvises is a publish executor you can't audit. The whole point of separating the deciding step from the doing step is that the doing step becomes boring, and boring is what you want at 6am when nobody's watching the run happen. So the executor gets a plan, a fixed set of Composio MCP tool calls, and a mandate to log the result — success, failure, or partial — with no interpretation layer in between.

Where creativity is allowed — and how tightly it's fenced

The interesting part is that not everything in today's queue got the boring treatment. Some of it needed voice, and voice needs constraints of its own kind.

• JADA's persona file got another pass — she's a 1938 wooden yawl, not a person playing a boat, and the instructions are explicit that she's been rigged as a yawl for decades: never a ketch, never a schooner. Getting the rig type wrong isn't a style nitpick, it's a factual error a real boat owner would catch instantly.
• The naval history piece for The Pulse — this one on the Strait of Magellan, the passage nobody knew existed until thirty-some ships' worth of dead reckoning found it in October 1520 — needed a second pass into 140-170 words of Ken Burns-style narration. The rule there was as strict as anything the publish executor follows: use only facts that appear in the source article, never add or embellish. Same discipline as the deterministic core, just applied to prose instead of API calls.
• Google review replies for JADA got tightened too: one to two sentences, 45 words max, and — the rule that actually matters — mirror exactly one specific phrase the guest used. Not marketing language back at them. Their own words, handed back warmer.

Laid side by side, the pattern is the same shape wearing different clothes. The publish executor has zero creative latitude because a wrong post is expensive and unrecoverable once it's live. The narration and the review replies have creative latitude, but it's fenced by source material — the article's facts, the guest's own phrase, the boat's actual rigging history. Neither system gets to invent. One just has a much smaller cage than the other.

The lesson

Every time I've gotten burned by an LLM step in this stack, it wasn't because the model was dumb — it was because I hadn't told it which kind of step it was in. Deciding steps get judgment. Doing steps get a plan and a log line. Voice steps get a fence made of real facts. Write the instructions like the job depends on knowing the difference, because it does.