Get your hours back: 10 AI skills for the back office
sop-writer
write the process down once so somebody else can do it
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: sop-writer
description: Writes a process up as a numbered, owned SOP somebody else can run. Use for documenting a process, a handover, or a standard operating procedure.
---
# The process written down once, so somebody else can do it
An SOP a stranger to the job can follow and still get it right.
## Claude does
1. Ask for four inputs, guess none: {PROCESS_NAME}; how the job is done today ({WALKTHROUGH}); the owner and every role ({OWNER_AND_ROLES}); the real approval limits, SLAs and targets.
2. Apply the three-sentence scope test: purpose in one or two sentences, scope in three. If it does not fit, split into two SOPs.
3. Pick one of four formats and say why: simple checklist, complex linear checklist, watertight process (regulated or money-touching), or flow diagram.
4. Write the governance header: version, last-updated date, review schedule (monthly compliance, quarterly stakeholder, annual audit), named owner or DRI.
5. Number the steps. One action each, opening with an imperative verb in the present tense. A step joining two actions with "and" gets split.
6. Name each decision point as a question with branches, and each exception with what to do and who to tell.
7. Write `sop-spec.md`: header, purpose, scope, roles, ordered steps with owners, decision points, approval levels with SLAs and thresholds, exceptions, metrics.
## Then Codex does
Claude decides the shape and wording; Codex builds the document: the template, the ASCII map, a RACI grid of every step against every role, the approval and exception tables, a one-page checklist, and a lint of every step. Bulk work.
```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```
Fill in first: {PROCESS_NAME}, {SPEC_FILE} (path to `sop-spec.md`), {OWNER_AND_ROLES}, {FORMAT}.
## Claude checks
1. Document step count equals the spec.
2. Every step opens with an imperative verb and joins no two actions with "and". Most SOPs fail here, so one breach rejects the file.
3. Purpose two sentences or fewer, scope three, header carrying version, date, review schedule and named owner.
4. Every RACI row has exactly one Accountable, and every owner is a named role, never "the team".
5. No SLA, threshold or target that was not supplied. An invented number is a rejection.
6. Fresh-eyes test: any step needing knowledge not on the page is rewritten.
7. Reject and rerun: fix `sop-spec.md` and rerun, never hand-patch Codex's output.
## Rules
- Fails the three-sentence test, split the SOP.
- One action per step. "And" means two steps.
- Never ship an SOP nobody has piloted: one non-owner runs it, then a superior reads it fresh.
- Never invent an SLA, approval limit, tolerance or target. Ask, or mark it missing.
- 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 builds the document, tables and map itself and says so.
## Built from
- w95/awesome-claude-corporate-skills, https://github.com/w95/awesome-claude-corporate-skills - 189 stars (api.github.com, 5 Sep 2026); the six components, RACI, SLA approval chains.
- Process Street, https://www.process.st/writing-standard-operating-procedures/ - four formats, verb-first rule, pilot and fresh-eyes gates.
- Anthropic Agent Skills best practices, https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices - authoring standard.
- obra/superpowers, https://github.com/obra/superpowers - 281,934 stars (api.github.com, 5 Sep 2026); verification before completion.
Prompt for Codex
# The process written down once, so somebody else can do it - Codex task
## You are given
- {PROCESS_NAME} - the process being documented.
- {SPEC_FILE} - path to `sop-spec.md`. Holds header fields, purpose, scope, roles, ordered steps with owners, decision points, approval levels with SLAs and escalation thresholds, exceptions, and metric targets. It is the only source of truth.
- {OWNER_AND_ROLES} - the named owner or DRI, plus every role in the process.
- {FORMAT} - one of `simple-checklist`, `complex-linear-checklist`, `watertight-process`, `flow-diagram`.
## Produce
In the working folder:
1. `SOP-<process-name>.md` - the full document in this order: governance header (version, last-updated date, review schedule, named owner), purpose, scope, roles, numbered steps, decision points, approval chain, exceptions, metrics. Lay the steps out in the style {FORMAT} calls for.
2. `raci.md` - a grid of every step (rows) against every role in {OWNER_AND_ROLES} (columns), each cell R, A, C, I or blank. Exactly one A per row.
3. `process-map.txt` - an ASCII map: every numbered step a box, every decision point a branch with named outcomes, every exception an exit.
4. `approval-chain.md` - a table of level, approver role, triggering threshold, SLA, escalation path if the SLA is missed. Values copied from {SPEC_FILE} only.
5. `checklist.md` - the one-page version, one tick box per step, no prose.
6. `lint.md` - one row per step: number, first word, whether that first word is an imperative verb, whether the step contains " and " joining two actions, PASS or FAIL.
## Rules
- Copy every step, owner, threshold, SLA and metric verbatim from {SPEC_FILE}. Never add a step, merge two, reorder them or improve the wording.
- Never invent a number. A missing threshold, SLA or target becomes `[MISSING: what is needed]`, recorded in `lint.md`.
- Do not rewrite or split a failing step. Flag it FAIL in `lint.md` and leave it exactly as written.
- Every owner is a named role from {OWNER_AND_ROLES}. Never "the team", "someone" or "TBC".
- Do not edit {SPEC_FILE}.
- Deterministic: the same inputs produce identical output.
## Return
Print, in this order:
1. Step count in {SPEC_FILE} and in `SOP-<process-name>.md`. State plainly whether they match.
2. Sentence count of purpose and of scope.
3. Steps failing the verb-first check, with their numbers.
4. Steps containing " and " joining two actions, with their numbers.
5. Number of RACI rows, and any row without exactly one A.
6. Every `[MISSING: ...]` marker and the file it sits in.
7. Approval levels written, and any lacking a threshold or an SLA.
8. The first ten lines of `SOP-<process-name>.md`.
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, 5 Sep 2026; created 26 Feb 2026, updated 2 Sep 2026). Its `07-operations/sop-builder/SKILL.md` was opened and read: six output components (process definition, step-by-step workflow, RACI, approval chains, ASCII process map, decision points), a template header carrying version, update date, review schedule and named owner, plus DRI assignment, per-level SLAs, escalation thresholds and a monthly/quarterly/annual review cadence.
- Process Street, 16 Essential Steps to Writing Standard Operating Procedures, https://www.process.st/writing-standard-operating-procedures/ - opened in full. Supplies the numbered method, the four format archetypes (simple checklist, complex linear checklist, watertight process, flow diagram) with a stated fit for each, the "always start a task with a verb" rule, the scannability rule, and two validation gates: step 11 pilot with a subset, step 12 fresh-eyes review by a superior.
- Anthropic, Agent Skills best practices, https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices - the official authoring standard. Hard limits (name 64 characters, description 1024 characters), the under-500-line body budget, one-level-deep references, the degrees-of-freedom model, and the plan, validate, execute pattern this skill is built on.
- obra/superpowers, https://github.com/obra/superpowers - 281,934 stars (read from api.github.com, 5 Sep 2026). Largest public skills framework and accepted into Anthropic's marketplace; its `skills/writing-skills/anthropic-best-practices.md` supplies the verification-before-completion gate and the checklist-in-response discipline used in the Claude checks section.
Best public prompt we found for this job
https://www.process.st/writing-standard-operating-procedures/
The most useful single rule it carries, and the one this skill enforces hardest, is that every task in a procedure "always start a task with a verb" - present tense, active voice, one action. It pairs that with the validation logic behind the fresh-eyes test: an SOP is proven by handing it to somebody who has never done the job, because a good process should produce a good result regardless of the operator's experience.
The closest working prompt to copy is `07-operations/sop-builder/SKILL.md` in https://github.com/w95/awesome-claude-corporate-skills, which is where the six-component output contract and the governance header in this skill come from.
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