- Home
- Skills
- Prompt & Skill Engineering
- Template Builder
Works with the AI Agents you already use
Template Builder
Turns a one-off prompt into a reusable template with typed placeholders, and audits it for false confidence and hallucination risks before you use it
$5
/moSubscription · cancel anytime
Secure checkout via Stripe
- Always the latest version
- Delivered live to your agent via MCP
- Cancel anytime, access ends at period end
Template Builder
Example session with this skill installed
Turn this into a reusable template for ChatGPT, add placeholders, and check it for anything that would make the model overconfident:
Write a cold email to the head of marketing at Brightline Coffee, a specialty coffee roaster in Leeds with about 40 staff. We're Northwind Analytics; we help retailers see which promotions actually make money. Mention that we helped a bakery chain cut wasted discounts by 18%. Keep it under 120 words, friendly, and end with a question about a 15-minute call. Double-check your facts before answering.
- Read your context and instructions
- Compiled the template builder
This is mode (c): your prompt has no variable slots yet, so I added them and audited it at the same time. You asked for all three things, so I went ahead without stopping to confirm first. The judgment calls are flagged below.
Variables (5 is the limit)
- RECIPIENT_ROLE: head of marketing
- PROSPECT: Brightline Coffee, a specialty coffee roaster in Leeds with about 40 staff
- SENDER_PITCH: Northwind Analytics; we help retailers see which promotions actually make money
- PROOF_POINT: we helped a bakery chain cut wasted discounts by 18% (optional)
- TONE: friendly (optional, chosen from a fixed list)
What would make the model overconfident
"Double-check your facts before answering." This is the main problem. ChatGPT has nothing to check Brightline Coffee against. The line won't make the email more accurate, but it makes the email sound checked, and that is false confidence. I removed it and put in two things that actually help:
- A rule the model follows while writing: use only the facts in the inputs.
A list you can check: a "Facts used" list under the email that says which input each claim came from. Anything you can't trace back is easy to spot.
Made-up first name. The prompt names a role but not a person, so models tend to write "Hi Sarah". The template now tells it to greet by role or use a neutral greeting.
Invented details about the prospect. From just "specialty roaster in Leeds", models like to add a café, a seasonal blend or recent growth. That is now forbidden. So is looking the company up or using what the model thinks it knows about a business with that name, which could be a different company or out of date.
Assuming the prospect has the problem. The email can't say Brightline is losing money on promotions. It has to present that as a possibility.
Stretching the 18%. Models often add a timeframe, a pound figure or a client name. The template says to quote the number exactly as you give it and add nothing.
Made-up sender details. No invented sender name, job title or contact details. The sign-off leaves YOUR NAME for you to fill in.
Word limit. "Under 120 words" now says what counts (greeting through sign-off), and the model adds its word count at the end. That count is still the model's own guess, so check it with a counter.
Other changes
- The output order is now set at the top, before the rules it applies to.
Removed a clash: the email ends on exactly one question (the call), so the "maybe you have this problem" line has to be a statement, not a second question.
- RECIPIENT_ROLE, PROSPECT and SENDER_PITCH are required. If one is missing, the model names it instead of guessing.
- If PROOF_POINT is "none", the email leaves out the proof sentence and uses no numbers at all.
Judgment call: I kept the 120-word limit and the 15-minute call fixed rather than making them variables, because the limit is five variables and TONE seemed more useful to vary. If you'd rather vary length, swap TONE for WORD_LIMIT.
- Judgment call: I didn't add a subject line, because your prompt didn't ask for one.
PROMPT TEMPLATE
INPUTS
RECIPIENT_ROLE : [RECIPIENT_ROLE]
PROSPECT : [PROSPECT]
SENDER_PITCH : [SENDER_PITCH]
PROOF_POINT : [PROOF_POINT]
TONE : [TONE]
DEFAULTS AND ACCEPTED VALUES
TONE accepts exactly one of: friendly, warm-professional, direct, playful. If TONE is empty or not on this list, use friendly.
PROOF_POINT is optional. If it is empty or reads "none", leave the proof sentence out entirely and use no numbers anywhere in the email.
RECIPIENT_ROLE, PROSPECT and SENDER_PITCH are required. If any of them is empty, do not write the email; reply with one line naming the missing input.
TASK
Write a short B2B cold email on behalf of the sender described in {{SENDER_PITCH}}, addressed to the {{RECIPIENT_ROLE}} at the company described in {{PROSPECT}}.
OUTPUT FORMAT - produce these parts in this order and nothing else
1. Greeting line
2. Email body
3. Sign-off
4. Word count line
5. Facts used list
Do not write a subject line, a preamble, a P.S. or any commentary before or after these five parts.
FACT RULES - apply these while writing
- Use only facts stated in the inputs. Every claim about the prospect, the sender or a result must come from {{PROSPECT}}, {{SENDER_PITCH}} or {{PROOF_POINT}}.
- Add nothing about the prospect that {{PROSPECT}} does not state: no products, venues, awards, news, growth, customers, campaigns or problems.
- Do not state that the prospect has the problem the sender solves. Present it as a possibility, not as a fact about them.
- Do not browse, search or rely on anything you believe you know about a company with this name; it may be a different company or out of date.
- The recipient's name is unknown. Do not invent one. Greet them by role, or with a neutral greeting such as "Hi there,".
- Quote the result in {{PROOF_POINT}} exactly as given. Do not add a timeframe, a money amount, a client name, a method or a second figure, and do not round or reword the number.
- Do not invent the sender's personal name, job title, phone number, website or any other contact detail.
WRITING RULES
- Tone: {{TONE}}.
- Length: under 120 words, counting the greeting, the body and the sign-off.
- First sentence: about the prospect, using only what {{PROSPECT}} states, linked to the kind of problem described in {{SENDER_PITCH}}.
- Then one sentence on what the sender does, taken from {{SENDER_PITCH}}.
- Then one sentence giving the result in {{PROOF_POINT}}, unless PROOF_POINT is empty or "none".
- The body ends with one question asking whether the recipient would be open to a 15-minute call. This is the only question in the email.
- Sign-off: the word "Best," on its own line, then the words YOUR NAME in capitals on the next line, then the sender's company name from {{SENDER_PITCH}} on the line after.
WORD COUNT LINE
After the sign-off, leave one blank line, then write the number of words from greeting to sign-off in parentheses, in the form: (Word count: 97)
FACTS USED LIST
After the word count line, write the heading "Facts used" and list every factual claim the email makes, one per line, each followed by a dash and the name of the input it came from, for example: specialty coffee roaster - PROSPECT. If a claim cannot be matched to an input, it must not appear in the email.
LISTING COPY
TITLE: Cold Emails That Invent Nothing (31)
DESCRIPTION: Write a short cold email to one named prospect that uses only the facts you give it. Fill in who you are writing to, what you sell and one result you can prove, and get a friendly email under 120 words ending in a 15-minute call ask. No made-up first names, no invented company news, no stretched stats. Every claim comes with a note saying which input it came from, so you can check it before you hit send. ✉️ (410)
INSTRUCTIONS
- Fill in the five lines under INPUTS by replacing the bracketed text on each line with your own value. Leave every {{NAME}} reference in the body exactly as it is. RECIPIENT_ROLE is a job title, for example head of marketing. PROSPECT is one sentence about the company: its name, what it does, where it is and its size.
- SENDER_PITCH is your company name followed by one sentence on what you help with, for example: Northwind Analytics; we help retailers see which promotions actually make money.
- PROOF_POINT is one result you can stand behind, written exactly as you want it quoted. Type none if you have no result to share, and the email will make no numeric claim at all.
- TONE takes one of friendly, warm-professional, direct or playful. Leave it empty to get friendly.
- Tip: the email can only be as specific as your PROSPECT line. If you want it to mention something about the company, put that fact in the input, because the template forbids the model from adding details of its own. Read the Facts used list before sending, and check the word count with a counter.
USAGE NOTES: There are five variables:
- RECIPIENT_ROLE (required) is who the email is addressed to.
- PROSPECT (required) is everything the email may say about the target company.
- SENDER_PITCH (required) is your company name and offer.
- PROOF_POINT (optional) is your one result, quoted word for word, or "none" to drop it.
- TONE (optional) is one of friendly, warm-professional, direct or playful. The default is friendly.
To use it, paste the whole block into ChatGPT with the INPUTS lines filled in. If you put it in a custom GPT's instructions, leave browsing off or keep the no-search rule. The 120-word limit and the 15-minute call are fixed in the body, so edit those lines to change them.
Three things you have to check yourself
The proof point. The template can stop the model stretching the 18%, but it can't check that the 18% is true. You also need the bakery chain's OK to mention it, even without its name.
- The word count. The model reports its own count, so check it with a counter.
- The sign-off. Swap YOUR NAME for your real name and add any opt-out line your outreach rules need.
Would you like to generate 4 example outputs? I will ask you for a set of variable values, run the template with them, and show you the real output. We will do this 4 times.
Subscription · always the latest version
- Secure checkout via Stripe
- Cancel anytime
- Delivered live via MCP
About this skill
Builds and audits reusable prompt templates. It pulls out the values that change between uses, declares them once in an inputs block, and puts them back where they belong. Then it does the part most template tools skip: it checks the prompt for instructions that make a model sound more certain than it is.
What it does
It works in three modes: build a template from scratch, fix an existing one, or add variables to a plain prompt. Variables get a single naming convention, with fixed choices and defaults where free text would be ambiguous (up to five). The audit removes false-confidence lines such as "double-check your facts", which make output sound verified without verifying anything, and replaces them with rules that help: use only the facts given, and list which input each claim came from. It flags anything the model would have to invent (names, details, numbers), checks the template's structure, applies the target platform's conventions, and ends with the finished template and a summary of every change.
Try it first
Paste a prompt you wrote for one case and say "turn this into a reusable template and check it for anything that would make the model overconfident". Check what it removed and why.
What it needs
Nothing to install: no scripts, no network access and no API keys. It targets ChatGPT, Claude, Gemini, n8n, API pipelines and prompt marketplaces.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Prompt & Skill Engineering.
It audits the wording, not the model's behaviour, so the template still needs a test run on real inputs; it offers to run example fills. It caps a template at five variables to keep it usable. It won't make up facts to fill a slot, and it won't fact-check your inputs. Word counts the model reports about its own output are flagged as unreliable.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 27 days ago
Permissions
Network Access is declared because the skill's text says "single curly braces", which automated scans read as curl; the skill makes no network requests.