routine-builder

---
name: routine-builder
description: Turns a job you repeat every week into a written, versioned routine you can run on demand, by interviewing you once, deciding whether it needs a single prompt or a workflow, and handing the file writing to Codex.
---

# Turn a job you do every week into something that runs itself

You describe the job once. You get back a named, versioned routine you paste in and run, plus a log line every time it runs.

## Claude does

1. Ask for the job in the user's own words, then ask **5 to 10 clarifying questions** and stop (the doc-coauthoring pattern in anthropics/skills). Required: what triggers it, what goes in, what comes out, who sees it, and one real past example of a good result.
2. Climb the ladder from Anthropic's *Building effective agents*: **single prompt first**, then a workflow, then an agent. Add a step only when it demonstrably improves the outcome. Say which rung and why in one line.
3. If it is a workflow, name one pattern: prompt chaining, routing, parallelisation, orchestrator-workers, or evaluator-optimizer. Never blend two.
4. Write it in the **10 fixed slots** (prompt-eng tutorial, ch.09): role, task context, tone, rules, examples, input data, immediate task, think step by step, output format, prefill. Long input at the top, the request at the bottom.
5. Supply **3 to 5 examples**, relevant, diverse and structured, each in `<example>` tags. Six for classification, with the order shuffled.
6. Write rules as **instructions, not prohibitions**, each led by an action verb: Extract, Rank, Rewrite, Summarise, Sort. Keep constraints for safety and output format, with the reason attached.
7. Fill the **log template** (Google whitepaper, Table 21): name and version, goal, model and version, temperature, token limit, top-K, top-P, full prompt, output. Start at temperature 0.2, top-P 0.95, top-K 30; use 0 where there is one correct answer.
8. Refuse to guess the trigger, the recipient, or any number the user did not give. Leave `[]` where a fact is missing.

## Then Codex does

Codex writes the folder out: the routine file, the filled log, a run checklist and a README, all from Claude's decisions. That is mechanical file production, identical for every routine, so it belongs to the bulk worker, not the interview.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills these first: `{ROUTINE_NAME}`, `{JOB_DESCRIPTION}`, `{TRIGGER}`, `{PATTERN}`, `{TEN_SLOTS}`, `{EXAMPLES}`, `{LOG_SETTINGS}`.

## Claude checks

1. Count the slots. Ten, in order, question last, or reject.
2. Count the examples. Under three, or all one shape, means rerun with diverse ones.
3. Find prohibitions with no reason attached and rewrite them as instructions.
4. Confirm no step was added that the ladder did not justify. Anthropic's own failure case was a vague brief producing duplicated, off-target work.
5. Confirm the log names a model version and a temperature. Without a version, next month's result cannot be compared.
6. Reject once, quoting the failing check. If it fails twice, Claude writes that part itself.

## Rules

- Public information only.
- Never invent a fact, a number or a quote. Missing facts become `[]`.
- Anything sent in someone's name says whose name it is.
- Do not build an agent where a prompt works.
- Version on every edit; never overwrite silently.

## If Codex is not installed

Claude writes the four files itself and says it did that half by hand.

## Built from

- anthropics/skills, https://github.com/anthropics/skills, 174,340 stars via api.github.com; skill-creator and doc-coauthoring give the 500-line cap and the 5 to 10 questions.
- Anthropic, Building effective agents, https://www.anthropic.com/engineering/building-effective-agents, the escalation ladder and the five workflow patterns.
- anthropics/prompt-eng-interactive-tutorial, https://github.com/anthropics/prompt-eng-interactive-tutorial, 38,070 stars via api.github.com; the 10-element order, ch.09.
- Lee Boonstra, Prompt Engineering whitepaper, https://www.kaggle.com/whitepaper-prompt-engineering, Table 21 logging template and the temperature settings.

# Turn a job you do every week into something that runs itself - Codex task

You are writing the files for one reusable routine. Every decision is already made and supplied below. Do not redesign it, add steps, or invent content.

## You are given

- `{ROUTINE_NAME}` - short kebab-case name, e.g. `weekly-supplier-chase`.
- `{JOB_DESCRIPTION}` - the repeating job in the user's own words.
- `{TRIGGER}` - what starts it: a day, an event, or a manual run.
- `{PATTERN}` - one of: single prompt, prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer.
- `{TEN_SLOTS}` - the routine text, already ordered as role, task context, tone, rules, examples, input data, immediate task, think step by step, output format, prefill.
- `{EXAMPLES}` - 3 to 5 worked examples.
- `{LOG_SETTINGS}` - model and version, temperature, token limit, top-K, top-P.

## Produce

Four files in the working folder.

1. `{ROUTINE_NAME}.md` - the routine, with all ten slots as headings `## 1 Role` through `## 10 Prefill`, in that order, question last. Each example in its own `<example>` tag inside one `<examples>` block.
2. `{ROUTINE_NAME}-log.md` - a table with these nine columns and one blank row: Name and version, Goal, Model and version, Temperature, Token limit, Top-K, Top-P, Full prompt, Output. Carry `{LOG_SETTINGS}` into the header. Version starts at `v1.0`.
3. `RUN.md` - a numbered checklist, 5 to 8 steps: when to run it, what to paste, where to paste it, what to check in the output, where to record the run.
4. `README.md` - under 200 words: the job, the trigger, the pattern in one sentence, and a link line to the other three files.

## Rules

- Copy `{TEN_SLOTS}` verbatim. Reorder nothing, reword nothing, add no slot.
- If a placeholder arrives empty, write `[]` in its place and list it under `## Missing` in `README.md`. Never guess.
- Write rules as instructions led by an action verb. Where a prohibition is needed for safety or format, give the reason on the same line.
- No em dash characters. British English.
- Plain Markdown only. No scripts, no network calls, no installs.
- Create, move or delete nothing outside the working folder.
- Overwrite nothing. If `{ROUTINE_NAME}.md` exists, write `{ROUTINE_NAME}-v2.md`.

## Return

Print this and nothing else:

1. `FILES:` each filename with its byte size.
2. `SLOTS:` the ten heading lines exactly as written, in file order.
3. `EXAMPLES:` the count of `<example>` tags found.
4. `LOG:` the nine column headers, comma separated.
5. `MISSING:` each empty placeholder, or `none`.
6. `VERSION:` the version string written into the log.