Role
Who the model is. One line narrows its vocabulary and the standard it holds itself to.
"You are a senior comms lead in financial services."
TANZO.AI / RESOURCES
Tanzo’s AI
Kitchen
Find every recipe for your AI journey.
Cook your path with us.
Free tools for anyone learning to work with AI properly. No account, nothing stored — everything on this page runs in your browser.
Prompt Structure
How a prompt is built.
Six parts, in this order. Omit one and the model fills the gap with an assumption — which is where generic output comes from. The builder assembles all six for you.
Objective
The job and the outcome you want from the reader — not just the task.
"...an email that secures written approval for £40k."
Context
Who the audience is, what they know, what happened before, what is at stake.
"CFO. Rejected a similar request in Q1 on ROI grounds."
Constraints
How it should sound and what it must never do. The most under-used lever.
"No em dashes, no buzzwords, no rhetorical questions."
Input
Your raw material, fenced clearly so content is not mistaken for instruction.
INPUT: """ ...your text here... """
Output Spec
Format, length, structure, and what not to return. Kills preamble and filler.
"Only the final email. 150–200 words. No preamble."
The Tool
Build your prompt.
Select a mode, tick a voice or framework, set the filters, add your text. The prompt on the right assembles as you go. Nothing is transmitted — this runs entirely in your browser.
Conversational For anything a person reads as a message — emails, LinkedIn posts, articles, replies, updates. It governs how the writing sounds: removing machine patterns and putting a real voice back in.
01 — Choose your voice
1st tick = Priority 1 · 2nd = Priority 2
Your first selection sets the dominant personality. Tick a second only where you genuinely need both. Where they conflict, Priority 1 governs and the clashing Priority 2 instruction is dropped.
02 — Set the context
03 — Things to avoid
Optional — removes machine tells
04 — Your text
Draft, notes or bullets
0 characters
01 — Output mode
Single = one · Power = chained
One framework, one focused deliverable. Use this when you know exactly what you need.
02 — Choose your framework
Tick one
03 — Set the business context
04 — Known constraints
Optional
05 — Your data, notes or draft
Findings, numbers or an outline
0 characters
Your prompt
Prompt strength
0%
Paste this into any model as your first message. If the output is close but not right, reply with one correction at a time rather than rewriting the whole prompt.
The difference
Prompt or skill?
Same capability. The difference is whether it survives the conversation.
Option A
Prompt
Written now, used once
- LifespanOne conversation
- InvokedYou paste it in, every time
- ConsistencyVaries with how you phrase it
- SharingCopy the text to a colleague
- ContentsText only
- SetupSeconds
Use for one-off work
Option B
Skill
Written once, used forever
- LifespanPermanent, until you remove it
- InvokedTriggers itself when the task matches
- ConsistencyIdentical every single run
- SharingSend a folder; they get your exact setup
- ContentsInstructions plus templates and scripts
- SetupMinutes to install
Use for work you repeat
The library
Skills you can actually download.
Every skill below lives in Anthropic's public repository, github.com/anthropics/skills. Nothing here is ours and nothing is rehosted — you copy the file, download it, or view the source on GitHub, and it stays current because it comes straight from the repository. The list checks itself against the repository each time this page loads.
Fastest route — no download at all
checking repository…
If you use Claude Code, register the repository as a marketplace and install a whole set in one go:
/plugin marketplace add anthropics/skills/plugin install document-skills@anthropic-agent-skills/plugin install example-skills@anthropic-agent-skillsOn a paid Claude.ai plan these example skills are already available — nothing to install. For everything else, use the copy or download buttons on each card below.
Vendor marketplaces
Software you already pay for increasingly ships its own skills. Maintained, tested, updated. Check your tools first.
Community directories
Aggregators index thousands of public skills by function. Broad coverage, uneven quality — a search index, not an endorsement.
Write your own
For anything specific to how you work, no directory will have it. Skill Creator above exists for exactly this.
Two things before you install anything. Licences differ — the four document skills are
source-available rather than open source, most others are Apache 2.0, and each folder's LICENSE.txt
governs. And a skill is instructions your assistant will follow, so read the file first, especially
anything that fetches from unfamiliar URLs or runs scripts.
Make them yours
One file that personalises all of them.
Rather than editing every skill you install — which breaks the moment the author updates it — write a single organisation profile skill. It carries your terminology, your metrics, your formats and your tone, and every other skill inherits it. Fill this in, download the file, install it once.
01 — Who you are
02 — How you talk
One per line
03 — What matters
04 — Hard rules
Optional
SKILL.md
Profile completeness
0%
Save this as SKILL.md inside a folder named organisation-profile, then install that folder the same way as any other skill. It will apply automatically from then on.
Getting started
Five steps, two minutes.
This is where almost everyone gets stuck. It is simpler than it looks.
1
Get it
Download a SKILL.md below, or install a whole bundle with one command.
2
Place it
One folder per skill in your skills directory. Bundles handle this for you.
3
Check it
Ask your assistant which skills it has. Missing means wrong folder or bad description.
4
Trigger it
Describe your task in plain words. Never call the skill by its name.
5
Improve it
Every correction you make in chat goes back into the file. That is the compounding part.
Step 04
The one that catches people. Say "I need to explain why marketing came in over budget last month" — not "run variance-analysis". The assistant matches your words against the skill's description and loads it on its own. If nothing happens, the description is wrong, not you.
Anatomy
What is actually in the file.
Worth a look even if you never write one — it tells you what to check when judging someone else's.
---
name: variance-analysis1
description: Decompose budget vs2
actual variances into drivers.
Use when explaining a monthly
P&L movement to leadership.
---
# Variance Analysis
1. Identify the drivers3
2. Quantify each in currency
3. Write one-line commentary
## Rules4
- Never estimate. Mark gaps [TBC].
- Always state the data source.
## Output5
- Table plus three bullets.
- No preamble.
1
Name
Lowercase, hyphenated. Identifies it — nothing more.
2
Description
The trigger, and the only part that decides whether the skill ever fires. Needs the words a person would actually use.
3
Instructions
The procedure, as ordered steps.
4
Rules
Edge cases and constraints. Usually the thinnest section and always the most valuable.
5
Output spec
Exact shape of the result, and what not to return.
On platforms. This format is used by Claude across its apps and developer tools. Other
assistants have close equivalents under different names, and the same file usually transfers with
light editing. Install paths change as these tools evolve — check current documentation.
Pitfalls
Five ways to get it wrong.
These account for most of the disappointment with skills.
Description reads "Helps with emails"
→
Use the words someone types when they need it
One skill covering all of finance
→
Three narrow skills that each do one job
Instructions with no rules or edge cases
→
Write down what to do when data is missing
Install it once and never touch it again
→
Every correction in chat goes back in the file
Generic skills with no context about you
→
Install the organisation profile above
Go further
Skills fix a task.
A system fixes the workflow.
A well-built library gets an individual a long way. Getting a whole organisation to the same standard is a different problem — it needs shared definitions, guardrails and a way to keep everything current as the work changes. That is what Tanzo AI designs.
Reply within one working day
Copied to clipboard