Get your hours back: 10 AI skills for the back office
policy-drafter
the boring documents a real business needs
How the two work together
Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.
Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.
No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.
Prompt for Claude
---
name: policy-drafter
description: Writes the boring policies a real business needs - governed, versioned, owned, enforceable - and refuses to invent legal wording it cannot source.
---
# The boring documents a real business needs, written properly
A governed policy: version, owner, review date, scope, numbered rules, escalation thresholds, sign-off block.
## Claude does
1. Demand four inputs first: {POLICY_TOPIC}, {COMPANY_NAME}, a named {POLICY_OWNER} (a person, not a department), and the sites or countries covered. An unowned policy is not a policy.
2. Apply the three-sentence scope test: purpose in 1-2 sentences, scope in at most 3. If it will not fit, split it into two policies and say which.
3. Pick one of four Process Street archetypes: simple checklist, complex linear checklist, watertight process, flow diagram. Name the choice and why: a conduct policy is not a checklist.
4. Write the governance header: version, last-updated date, review schedule, named owner. Then purpose, scope, applicability, definitions.
5. Write rules as numbered steps, imperative verb first, present tense, active voice, one action each. If a step contains "and", split it.
6. Add the approval and escalation table: each level with its own SLA and monetary or complexity threshold. Then exception handling: who grants one, how it is logged.
7. Set the review cadence: monthly compliance check, quarterly stakeholder review, annual audit.
8. Refuse to guess. Statutory notice periods, legal minimums, insurance limits and anything jurisdiction-specific get marked `[CONFIRM WITH ADVISER]`, never a number.
## Then Codex does
Codex takes the approved outline and does the bulk work: expanding each rule into prose, regenerating it per site or country, building the glossary, and writing a one-page plain-English summary and a staff acknowledgement form. Repetitive formatting at volume. Claude fills {POLICY_TOPIC}, {COMPANY_NAME}, {POLICY_OWNER}, {JURISDICTIONS} and {APPROVED_OUTLINE} into CODEX.md first.
```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```
## Claude checks
1. Every step starts with an imperative verb and holds one action. A step containing "and" is a reject.
2. Scope still three sentences or fewer after expansion. Codex pads it.
3. Every `[CONFIRM WITH ADVISER]` marker survived. An unsourced legal figure is the worst failure here: authoritative-looking and invented.
4. A named owner on every version; version number and date changed.
5. Hand it to someone who has never done the job. If they cannot follow it, it fails.
6. Any check fails: quote the failing line, send it back, rerun. Never patch it yourself.
## Rules
- One policy per document. Scope beyond three sentences means split it.
- No step without a verb. No rule without an owner. No document without a review date.
- Never fill in a legal figure, notice period or statutory right from memory.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
## If Codex is not installed
Claude writes the expansion, per-site versions and summary itself, and says so.
## Built from
- w95/awesome-claude-corporate-skills - https://github.com/w95/awesome-claude-corporate-skills - 189 stars via api.github.com; its sop-builder gives the governance header, SLAs, thresholds, review cadence.
- Process Street, 16 Steps to Writing SOPs - https://www.process.st/writing-standard-operating-procedures/ - the four archetypes, verb-first steps, fresh-eyes gate.
- Anthropic, Skill authoring best practices - https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices - the authoring spec.
Prompt for Codex
# The boring documents a real business needs - Codex task
You are expanding an approved policy outline into finished documents. You have no memory of the conversation that produced it. Everything you need is below.
## You are given
- `{POLICY_TOPIC}` - the subject of the policy, for example "expenses", "remote working", "data retention".
- `{COMPANY_NAME}` - the business the policy belongs to. Use it verbatim.
- `{POLICY_OWNER}` - one named person accountable for the policy. Not a department.
- `{JURISDICTIONS}` - the countries or sites this policy covers, comma separated. One output document per entry.
- `{APPROVED_OUTLINE}` - the agreed structure: governance header, purpose, scope, definitions, numbered rules, approval and escalation table, exception handling, review cadence. Markers reading `[CONFIRM WITH ADVISER]` are deliberate.
## Produce
Write into the working folder:
1. `policy-{POLICY_TOPIC}-<jurisdiction>.md`, one file per entry in `{JURISDICTIONS}`, with `<jurisdiction>` replaced by that entry lowercased and hyphenated. Each carries, in this order: a governance header table (version, last updated, review schedule, owner), purpose (1-2 sentences), scope (max 3 sentences), who it applies to, definitions, numbered rules, approval and escalation table with a stated SLA and threshold per level, exception handling, review cadence, sign-off block.
2. `glossary.md` - every defined term used across the documents, alphabetical, one line each.
3. `summary-one-page.md` - plain English, under 400 words, no jargon, aimed at a new starter.
4. `acknowledgement-form.md` - name, role, date, signature line, and one sentence confirming what is being acknowledged.
## Rules
- Every numbered rule starts with an imperative verb, present tense, active voice.
- One action per rule. If a rule would contain "and", split it into two numbered rules.
- Scope stays at or under three sentences in every output file. Do not pad it.
- Copy every `[CONFIRM WITH ADVISER]` marker through unchanged. Never replace one with a number, a notice period, a statutory right or a legal citation. Inventing one is a failure of the task.
- Do not add rules that are not in `{APPROVED_OUTLINE}`. Expand wording only.
- Do not add sections, appendices or extra files beyond the four listed.
- British English. No em dash characters.
- Every document names `{POLICY_OWNER}` and carries a version number and date.
## Return
Print, and nothing else:
- The list of files written, with the byte size of each.
- The rule count per file.
- The exact scope sentence count per file.
- Every line still containing `[CONFIRM WITH ADVISER]`, quoted with its file and line number.
- Any rule containing the word "and", quoted with its file and line number, under the heading `SPLIT CANDIDATES`. If there are none, print `SPLIT CANDIDATES: none`.
Built from the best public work on this
Sources
- **w95/awesome-claude-corporate-skills** - https://github.com/w95/awesome-claude-corporate-skills - 189 stars read from api.github.com (created 26 Feb 2026, updated 2 Sep 2026). Its `07-operations/sop-builder/SKILL.md` was opened and read: it supplies the governance header carrying version, update date, review schedule and a named owner, plus DRI assignment, per-level SLAs, escalation thresholds and the monthly/quarterly/annual review cadence used above.
- **Process Street, 16 Essential Steps to Writing Standard Operating Procedures** - https://www.process.st/writing-standard-operating-procedures/ - opened in full. Supplies the four format archetypes (simple checklist, complex linear checklist, watertight process, flow diagram), the rule that every step starts with a verb, the make-it-scannable rule, and the two validation gates: pilot with a subset, then fresh-eyes review by a superior.
- **Anthropic, Skill authoring best practices** - https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices - Anthropic's own normative spec for SKILL.md: name and description limits, the body budget, the degrees-of-freedom model and the plan-validate-execute pattern this skill follows.
Best public prompt we found for this job
https://www.process.st/writing-standard-operating-procedures/
The most useful transferable instruction from it is the validation gate: pilot the document with a subset of the people who will use it, then have a fresh pair of eyes review it. Its stated logic is that a good process produces a good result regardless of the operator's experience - which is why the acceptance test in this skill is handing the finished policy to someone who has never done the job, rather than asking the author whether it reads well.
Want this running in your business?
I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.
Foxera