Run the business: 10 Claude skills for the jobs only you can judge

house-rules

turn the way you do things into a skill Claude follows every time

How to use it

Claude only. Nothing else to install. Paste the prompt into Claude, or drop the folder into your skills folder and Claude picks it up on its own.

Every step, every judgement and every check happens in one place. No second tool, no hand-off, no command line if you do not want one.

The prompt

---
name: house-rules
description: Turns your own documents, notes or a described process into a reusable skill file Claude loads and follows, so the same job comes out the same way every time.
---

# Turn the way you do things into a skill Claude follows every time

You hand over a document, a folder of notes, or just talk through how the job goes. You get back a finished skill file that Claude loads whenever that job comes up.

## What it does

1. **Take the source before taking the description:** ask for the actual material first, the handbook, the checklist, the old emails, the folder of notes, the pricing sheet. book-to-skill, the public converter this borrows from, points out that the input does not have to be a book: internal documentation, runbooks, onboarding guides, brand and tone of voice documents, specs and standards all work, and its rule of thumb is that if you reopen a document often enough to wish you had memorised it, it is a candidate. Only if there is no document do you interview the owner instead.
2. **Ask when it should fire, not what it does, and write that answer down first:** Anthropic's own skill-creator asks four questions before anything is written: what should this enable, when should it trigger and on what phrases, what is the expected output format, and does it need test cases. Get those four answers before drafting. The trigger answer becomes the description line and nothing else in the file does that job.
3. **Pull out the structure, not a summary:** book-to-skill is explicit that a skill is not a book report. Go through the source and separate four things: named methods with their real names kept exactly as the owner says them, decision rules of the form when X, do Y, the numbers and thresholds, and the anti patterns, meaning the things that went wrong before and must not happen again. A paragraph that is not one of those four is background and does not go in the file.
4. **Find the thresholds and make the owner state them out loud:** every real process has numbers that live in somebody's head, the deposit percentage, the callback window, the discount nobody may go below, the day of the month it must be done by. Ask for each one by name. Where a number is missing, write the question into the draft in place of the number rather than inventing a plausible figure, and tell the owner which questions are open.
5. **Write the description as triggering conditions only, never as a summary of the steps:** obra/superpowers, in its writing-skills skill, reports that when a description summarised the workflow, the agent followed the description and skipped the file, doing one review where the skill clearly showed two. Its rule is that the description says when to use, and the body says what to do. Keep it under about 500 characters and start from the situation the owner is actually in.
6. **Match the shape of each rule to the way the job goes wrong:** superpowers found in head to head testing that the same guidance works or backfires depending on the failure. If people know the rule and skip it under pressure, write a flat prohibition. If they comply but the output comes out the wrong shape, write a positive recipe stating what the output is and in what order, because a prohibition just invites negotiation. If a required item keeps getting left out, add it as a named slot in a template. If the behaviour depends on a condition, write it as a conditional on something observable. Classify first, then write.
7. **Keep the file short and push the bulk into companion files:** Anthropic's skill-creator describes three levels of loading, the name and description always in context, the body loaded whenever the skill fires, and bundled files read only when needed. It puts the body under about 500 lines and says that if you are near the limit, add a layer with clear pointers to where to go next. superpowers is stricter, under 500 words for an ordinary skill. Keep the body to the decisions, and put the long price lists, form templates and reference tables in separate files the skill names.
8. **Test it against a real job before calling it finished:** take a job that has already been done, run the new skill on it cold, and compare the output with what the business actually produced. superpowers puts this at the centre: if you did not watch what happens without the skill, you do not know whether the skill teaches the right thing. Where the skill and reality disagree, ask the owner which is right, then fix the file. Where the skill hedged instead of deciding, that is a missing threshold from step 4.
9. **Hand back the file plus the list of open questions:** the finished skill, in a folder named after it, with a plain English note saying what it now does automatically, what it will ask the owner every time, and which questions from step 4 are still unanswered and are therefore holding a gap in the file.

## Then it checks

1. The file opens with YAML frontmatter containing a name and a description, and nothing else is required to be there. The name matches the folder name, is 64 characters or fewer, uses only lowercase letters, numbers and hyphens, and does not start or end with a hyphen. The description is 1,024 characters or fewer.
2. The description describes when to use the skill and does not restate the steps. If a reader could follow the description instead of opening the file, rewrite it.
3. Every number, threshold, rate and deadline in the file traces back to something the owner said or to a line in their own document. Point at the source for each one. Nothing is filled in from what is typical.
4. Every step is specific enough that a person could disagree with it. Any step that would be true of any business at all is deleted or replaced with what this business actually does.
5. The body is under 500 lines, and anything longer than that lives in a named companion file the body points to.
6. The skill has been run once against a real past job and the differences have been resolved with the owner, not silently smoothed over.

Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.

## Rules

- Public information only.
- Never invent a fact, a number or a quote.
- Never invent a step the owner did not describe. A gap in the process is written into the file as a question, not filled with a sensible default.
- Keep the owner's own words for anything that has a name in the business, the stages, the job titles, the product names. Renaming things to sound tidier is how a skill stops matching the business.
- Never put a password, an API key, a bank detail, a customer's personal data or a supplier price that is under a confidentiality agreement into a skill file. Skill files get copied and shared. Reference where the value lives instead of writing it down.
- If the process touches tax, employment, health and safety or anything with a legal duty attached, the skill it produces prepares the work for a qualified professional to check, and must say so in its own Rules section. It does not become the advice.

## Built from

- virgiliojr94/book-to-skill, https://github.com/virgiliojr94/book-to-skill, 28,944 stars: its SKILL.md is a numbered conversion pipeline, Steps 0 to 11, that validates the input, asks the content type, asks what the skill is for, then extracts frameworks, principles, techniques and anti patterns rather than summarising. Steps 1 and 3 here take that extraction list and its rule that a skill is structure and not a summary, and step 1 takes its "Beyond books" section, which names internal documentation, runbooks, onboarding guides and brand documents as valid inputs.
- anthropics/skills, skill-creator, https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md, 175,004 stars: step 2 uses its four capture intent questions verbatim in substance, and step 7 uses its progressive disclosure model, metadata always in context, body when the skill fires, bundled resources on demand, with the body kept under about 500 lines and a pointer added when it grows past that.
- obra/superpowers, writing-skills, https://github.com/obra/superpowers/blob/main/skills/writing-skills/SKILL.md, 282,768 stars: step 5 takes its finding that a description summarising the workflow causes the agent to follow the description and skip the file. Step 6 takes its "Match the Form to the Failure" table, which pairs each type of failure with the form of guidance that fixes it and warns that prohibitions backfire on shaping problems. Step 8 takes its core principle that a skill written without watching the failure first is not known to teach the right thing.
- Agent Skills specification, https://agentskills.io/specification, page last modified 4 August 2026, referenced by both obra/superpowers and coreyhaines31/marketingskills as the open format their skills target: the frontmatter limits in check 1, name required and at most 64 characters of lowercase letters, numbers and hyphens with no hyphen at either end, description required and at most 1,024 characters.

Built from the best public work on this

Sources for house-rules

Every source below was opened and read on 7 September 2026. Every star count is the exact

`stargazers_count` read from api.github.com at 17:46 UTC that day, not a rounded badge and not an

estimate. Where a source does not support what the skill says it

supports, that is stated plainly rather than smoothed over.

virgiliojr94/book-to-skill

The public converter that turns a written document into an agent skill, read today at

https://github.com/virgiliojr94/book-to-skill. The repository page shows **28,944 stars**, which

matches the figure already in the skill. Its default branch is `master` rather than `main`, and the

file that matters, `SKILL.md`, sits at the top level of the repository. It is a numbered pipeline

running from Step 0, an out of scope check, through Step 1 validate input, Step 1.5 identify content

type, Step 2 extract text, Step 3 analyse structure, Step 4 ask purpose, and on to Step 11, publishing

the finished skill, so the description of it in the skill is accurate. Its Philosophy section opens

with the line this skill leans on hardest: "Extract structure, not summaries. A skill isn't a book

report." It then lists what should be pulled out instead, which is named frameworks, actionable

principles, techniques, anti-patterns and voice calibration. Step 3 of this skill takes four of those

five and translates them for a business owner rather than an author: named methods, decision rules of

the form when X do Y, numbers and thresholds, and the things that went wrong before. The same section

also says "Preserve the author's precision. Frameworks often have specific names for reasons", and

that is the origin of the rule in this skill about keeping the owner's own words for anything that has

a name in the business. **One correction worth recording:** the "Beyond books" material quoted in step

1 is in the repository's README.md, not in its SKILL.md. The cited URL is the repository root, so it

loads the README and the claim stands, but anybody checking the SKILL.md alone will not find it. That

section names internal documentation, architecture decision records, runbooks, onboarding guides,

brand and design systems including tone of voice documents, research clusters, and specs and

standards, all as valid inputs, and it ends with the sentence step 1 paraphrases: "If you re-open a

document often enough to wish you'd memorized it, it's a candidate."

anthropics/skills, skill-creator

Anthropic's own public skill for writing skills, read today at

https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md. The repository page

shows **175,004 stars**, matching the figure in the skill. Its "Capture Intent" section holds the four

questions step 2 is built on, and they appear there in this order and almost this wording: what should

this skill enable Claude to do, when should this skill trigger and on what user phrases or contexts,

what is the expected output format, and should test cases be set up to verify the skill works. The

skill's instruction to answer all four before drafting matches the file, which tells the writer to

extract those answers from the conversation first and have the user confirm before proceeding. Its

"Progressive Disclosure" section is the source of step 7 and is a three level model: metadata, meaning

the name and description, always in context at around 100 words, the SKILL.md body loaded whenever the

skill triggers with under 500 lines given as the ideal, and bundled resources under scripts,

references and assets loaded only as needed. Its key pattern reads "Keep SKILL.md under 500 lines; if

you're approaching this limit, add an additional layer of hierarchy along with clear pointers about

where the model using the skill should go next to follow up", which is exactly what step 7 tells the

owner to do with long price lists and form templates. One extra point the file supports and the skill

does not claim: skill-creator independently agrees with step 5, because its guidance on the

description field says all the "when to use" information belongs in the description and not in the

body. So the two strongest sources here reach the same conclusion by different routes.

obra/superpowers, writing-skills

The most heavily starred public treatment of how to write a skill that actually changes behaviour,

read today at https://github.com/obra/superpowers/blob/main/skills/writing-skills/SKILL.md. The

GitHub API record shows **282,768 stars**, read from https://api.github.com/repos/obra/superpowers at

17:46 UTC, and that is the figure carried in the skill. Three things in this file carry three

different steps. First, step 5. Its section on the description field says "The description should ONLY

describe triggering conditions. Do NOT summarize the skill's process or workflow in the description",

and then gives the evidence: "Testing revealed that when a description summarizes the skill's

workflow, an agent may follow the description instead of reading the full skill content. A description

saying 'code review between tasks' caused an agent to do ONE review, even though the skill's flowchart

clearly showed TWO reviews". Changing the description to a plain trigger fixed it. The same section

also says "Keep under 500 characters if possible", which is where step 5's length guidance comes from,

and it is guidance rather than a hard limit, since the format itself allows 1,024. Second, step 6. The

file has a table headed "Match the Form to the Failure" with four rows, and step 6 follows it row for

row: a rule skipped under pressure gets a prohibition, output of the wrong shape gets a positive

recipe stating what the output is and in what order, an omitted element gets a required slot in the

template, and behaviour that depends on a condition gets a conditional keyed to something observable.

The file's warning is stronger than the skill's and worth knowing: in head to head wording tests the

prohibition version of shaping guidance produced more of the unwanted content than the recipe version,

and trended worse than giving no guidance at all. Third, step 8. The file's core principle line reads

"If you didn't watch an agent fail without the skill, you don't know if the skill teaches the right

thing", and its testing section insists on a no-guidance control. Step 8 translates that from agent

testing into something a business owner can actually do, which is to run the new skill cold on a job

already done and compare it with what the business really produced. That is a fair translation rather

than a quotation, and the skill presents it as such. For completeness, the file's word budget is

tighter than Anthropic's line budget, at under 500 words for an ordinary skill, which is what step 7

reports.

Agent Skills specification

The open format specification both of the repositories above target, read today at

https://agentskills.io/specification. The page loads, and its own metadata records a last modified

timestamp of 2026-08-04, which is the 4 August 2026 date given in the skill. Its frontmatter table is

the source of check 1 and every constraint in that check is correct: `name` is required, at most 64

characters, lowercase letters, numbers and hyphens only, and must not start or end with a hyphen;

`description` is required and at most 1,024 characters. The specification adds two rules check 1 does

not mention and which are worth knowing when a skill is rejected for no obvious reason: the name must

not contain two hyphens in a row, and it must match the parent directory name, which is why this skill

insists the name and the folder agree. The page also confirms the 500 line guidance quoted from

Anthropic under "Progressive disclosure", where it says to keep the main SKILL.md under 500 lines and

move detailed reference material to separate files, so check 5 is standing on the format itself and

not only on one repository's house style. On the claim that both repositories reference it:

obra/superpowers links to this exact page from its frontmatter section and again from its deployment

checklist, so that half is exact. coreyhaines31/marketingskills links to the site root,

https://agentskills.io, describing itself as working with any agent that supports the Agent Skills

spec, rather than to the specification page itself. The substance is right, the link is one level

shallower than the skill implies.

Best public prompt we found for this job

**obra/superpowers, the writing-skills skill,

https://github.com/obra/superpowers/blob/main/skills/writing-skills/SKILL.md**. Anthropic's

skill-creator is the better guide to the mechanics, and book-to-skill is the better guide to turning a

document into structure, but writing-skills is the only one of the three that tested its advice and

reported what happened, including the cases where the obvious advice made things worse. The one line

worth copying verbatim is its core principle:

If you didn't watch an agent fail without the skill, you don't know if the skill teaches the right
thing.

For a business owner that reads as: until you have run the new skill on a job you have already done

and compared the two, you do not have a skill, you have a wish. It is why step 8 exists and why check

6 will not pass without it.

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.