Automation

ChatGPT, Claude, Gemini, and Copilot all ship a way to re-run a saved prompt. A model is in the loop every time. The run draws on the person's plan. The output can vary. The model underneath will change. The run acts with that person's credentials.

Use it when the run needs judgment. Do not use it when the numbers have to match a recipe you already settled.

Plexara does not schedule these prompts. The assistant does. What Plexara holds is the procedure the prompt names, when you point the scheduler at a prompt in the library.

Every prompt page in Plexara has a copyable invocation: "Run the `monday-market-briefing` prompt with ..." Paste that line into ChatGPT Tasks, Claude scheduled tasks, or Gemini scheduled actions. When the schedule fires, the assistant connects to Plexara the way it always does, resolves the prompt from the library, and runs it. Arguments, version, approval status, and the audit of every tool call stay in Plexara. The clock stays in the assistant.

On the ACME demo, My Monday Market Briefing is a personal prompt. Last week's revenue by region, top and bottom stores, weather alerts that touch store states, an inventory attention list, pending items in the review queue. It answers in conversation. A person who wants that every Monday pastes the invocation into the assistant they already use.

A private prompt in the assistant works the same way if it calls Plexara tools directly. A governed Plexara prompt adds review, sharing, version history, usage columns, and attached reference material. The library is When prompts become shared infrastructure. Running one by hand or on a schedule is Lesson 404.

A managed script is a small program the assistant writes during the conversation where the logic was settled. Plexara stores it, versions it, and executes it. No model at run time. The authoring loop, the language, and the regional flash worked example are in the report that runs without the agent. The 600 series is the how-to. What this page adds is the job that looks like ETL, done without a data-team ticket.

Ask the assistant that has Plexara loaded: pull county demographics from the Census API, produce a CSV, register it as a table, feed a dashboard, run it monthly. On ACME that conversation produced three scripts. `acme-county-demographics` makes fifty-one Census ACS 5-year calls and registers two CSVs as tables. `acme-county-competition` does the same for County Business Patterns. `acme-expansion-screen` joins those tables to the store master and writes a CSV plus a JSON feed the Expansion Screen dashboard reads. Nobody on a data team opened a repository.

Hourly is the same shape. `acme-dc-weather-watch` calls the National Weather Service six times and patches the data island of an HTML dashboard, on the hour, UTC. After a run, the Scripts page shows the schedule in words, the next fire, and how the last run went. A Failing tile is the list of what needs a person. A failed scheduled run mails its owner with the reason and the log. Failures are not retried. The same script on the same inputs fails the same way.

A delivery destination Deasil sets up for the deployment is how a CSV lands in a bucket. The script names the destination. It never holds a credential.

Claude Code, Codex, and Cursor are how a lot of people write the first draft of a script. Their own schedulers still run a model when the schedule fires. That is a scheduled prompt. Use it when the run needs judgment. Do not use it when the logic is already settled.

Authoring and executing are separate decisions. The agent that wrote the script can hand it to Plexara. Plexara runs it under the author's access, with a run record and no model in the loop. Replit Scheduled Deployments also run with no model: a shell command on a schedule. Plexara adds checks those schedulers do not: a script cannot be saved unless it parses, a dry run executes under the author with nothing persisted, and a run can never do what its author could not.

A model per run is a cost and a variance. Processing one million rows costs roughly $800 to $2,000 through an LLM, against under $1 with SQL or rules. The ACME regional flash, the worked example in the earlier insight, is three queries and 1.85 seconds, and it charges no tokens against anyone's subscription because no model ran.

Reliability falls when you ask a model to do the same tool-using job twice. tau-bench measures pass^k, all k trials succeed, and found GPT-4o function-calling agents under 50 percent of tasks, with pass^8 under 25 percent in retail. A scheduled prompt that has to invent Tuesday's ranking will invent a different Tuesday. A managed script will not.

The model under a scheduled prompt will change. OpenAI gives at least six months' notice for generally available models. Anthropic gives at least 60 days. GPT-5.5 leaves ChatGPT on October 14, 2026. A prompt written against one model silently runs on another. A managed script does not care which model the author used last quarter.

Fivetran's 2026 benchmark put 53 percent of engineering capacity on pipeline maintenance, 4.7 failures a month, about 13 hours each. The Plexara answer is not a data team in front of every extract. A script that does not parse cannot be saved. Validate reports what it would reach before it runs. A dry run executes for real under the author, with nothing persisted. The language has no network, filesystem, clock, or randomness, so the class of string-built SQL and unbounded I/O has nowhere to live. Every version records its author. A failed run mails once and waits for an edit.

A lot of standing work needs both: numbers that stay the same, and a decision that can change.

On ACME, three managed scripts compute category velocity, supplier margin, and regional weather, with no model at run time. The Weekly Merchandising Review is a scheduled prompt. It runs those scripts for the numbers, reads them against five knowledge pages (seasonality calendar, returns policy, store formats, stock health bands, weather playbook), and writes a promote, discount, discontinue, or watch list. The assistant is on a weekly clock. The scripts are on theirs. Both are audited.

That is the pattern: managed scripts produce the numbers; a scheduled prompt reads them. Do not pay a model to re-derive a figure a script already computed. Do not ask a script to decide a discount.

Judgment each run: schedule a prompt in the assistant you already use. Settled logic: have that assistant write a managed script, and let Plexara run it. Both: scripts for the numbers, a scheduled prompt for the call.

The assistant schedules prompts. Plexara schedules scripts. The model is rented. The prompts and scripts are owned.