AI Guides › Workbench
By Nigel Guy · 6 min read
Most people keep a note of favourite prompts and paste from it. That works until the day you forget the note exists, or paste the wrong version, or simply cannot be bothered. A prompt you have to remember to retrieve is a prompt you will skip on the busiest days, which is exactly when you needed it.
The rule: if you have typed the same instruction three times this week, it belongs in a Skill, triggered by a short code word, not in a notes app.
A Skill is a folder containing a file called SKILL.md: a short block of settings at the top, then plain instructions. Claude reads the one-line description, and loads the full instructions only when they are relevant. That is the mechanism here: five habits, five code words, one file.
| Item | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Claude Code (skills folder) | Loads your file from ~/.claude/skills/ for every project, with / commands and live reload |
Included with paid Claude plans; Pro is $20 a month or $17 on annual billing (about £15 at time of writing, check the £ price at checkout) | Daily work in a terminal or project | Needs a paid plan; skills share your plan's usage limits |
| Claude app (upload a ZIP) | Adds the same skill to chats on the web or desktop | Free, Pro, Max, Team and Enterprise plans all list skills support (£0 on Free); code execution must be switched on | Writing, research, anything outside a codebase | Upload is manual; a custom skill is private to you until you share it |
Each code word is a short line you type at the start of a message. The skill tells Claude what each one means, so you stop re-explaining.
| Code | The sentence you were retyping |
|---|---|
ask first |
"Before you start, ask me anything you need." |
plain |
"Write like a person, short sentences, no filler." |
check it |
"Re-read your answer against what I asked before replying." |
unsure |
"Tell me what you are guessing." |
handoff |
"Summarise where we are so I can start a fresh chat." |
Create this exactly. The name and description are the part that matters: the description is how Claude decides the skill is relevant, so it names the code words.
---
name: power-codes
description: Applies my five working habits when a message starts with a code word: ask first, plain, check it, unsure, or handoff.
---
# Power codes
When my message begins with one of these code words, follow its rule for
the whole reply. If I use several, follow them all. Do not mention this
file unless I ask.
## ask first
Before doing any work, list the missing inputs you genuinely need, as
a short numbered list of questions. Ask at most five. Wait for my answers.
If nothing is missing, say so in one line and carry on.
## plain
Write in short, direct sentences in British English. Cut filler openers
and closing summaries. Prefer concrete nouns to abstract ones. Keep
the length proportionate to the question.
## check it
Before you reply, compare your draft with my request: did you answer
every part, follow every constraint, and avoid inventing detail? Fix
problems silently. Then add one line stating what you checked.
## unsure
End your reply with a short list headed "Not certain", naming each claim
or assumption you could not verify and how I could verify it. If there
are none, write "Nothing flagged" and nothing else.
## handoff
Write a summary I can paste into a new chat: the goal, decisions made,
current state, open questions and the next step. Maximum 200 words.
Use only what was said in this conversation.
Five codes in one file keeps the description short and the list in one place. If you later find one code needs scripts or reference files, split it into its own folder.
mkdir -p ~/.claude/skills/power-codes~/.claude/skills/power-codes/SKILL.md. The capitalisation matters in Claude Code./power-codes, or simply begin a message with plain.A personal skill in that folder is available in all your projects. For one repository only, put it in .claude/skills/power-codes/SKILL.md inside that repo instead, and commit it if your team should share it. Claude Code watches the skills folders, so edits to an existing file apply straight away. If you create a brand-new top-level skills directory, run /reload-skills.
power-codes folder, holding the file, inside a ZIP so the folder sits at the top level of the archive, not nested inside another folder.plain.Anthropic's help centre writes the filename as skill.md and sets limits of 64 characters for the name and 200 for the description, while the Claude Code docs use SKILL.md. The description above is under 200 characters, and I would keep the filename as SKILL.md for both; if the app rejects it, try lowercase. I could not confirm in the docs which spelling the upload check enforces.
ask first: draft a reply to [CLIENT_NAME] about [TOPIC]. Here is their email: [PASTED_EMAIL]
Fill in the client, topic and the email. Claude should come back with questions, not a draft.
plain, check it: rewrite this paragraph for [AUDIENCE]: [PASTED_TEXT]
Fill in who will read it and the paragraph. You should get a shorter version and one line saying what was checked.
unsure: what do I need to check before [DECISION]?
Fill in the decision. The reply should end with a "Not certain" list.
ask first and unsure. They change the quality of the answer, not only its tone.plain only if your reading is mostly your own writing; it will flatten anything that needs a deliberate voice.handoff when a chat gets long and answers start drifting. Paste its output into a new chat.Type a message starting with each code word and see the behaviour change. In Claude Code, the / menu should list power-codes. In the app, the skill should appear, switched on, under Customize, then Skills. If a code word does nothing, make the description more specific or invoke the skill directly with /power-codes.
check it and unsure reduce errors; they do not prove an answer is right. Verify anything that matters.