Prompt templates you actually reuse
- prompting
- templates
A prompt template is a prompt with named slots: fixed text for the rules you always want, and placeholders for the parts that change each time. The order matters — build the template after a prompt has worked twice, not before. A template is a reward for a solved problem, not a guess about one.
Fix the rules, open the slots
Fixed: the task definition, the constraints that protect quality, the output shape, and a self-check line. If a rule exists because a specific failure happened, it belongs in the fixed text with a note about which failure it prevents.
Open: the topic, the audience, the source text, the tone when it genuinely varies. Each open part gets a name — [AUDIENCE], <<<SOURCE>>> — so a missing one is visibly missing instead of silently guessed. A placeholder the model can see is a question it can ask; one it cannot see becomes an invention.
Build it from evidence
- Take a run that actually worked and read it line by line: which parts would change for the next use, which parts would not?
- Circle the fixed parts; give each changing part a name and a one-line reminder of what state it arrives in.
- Add one example if the shape is subtle — one good example beats three adjectives.
- Add the check line: what must the model verify before it answers? (“Keep every number and date”, “If the source is missing, say so in one line”.)
Weak prompt
You are a helpful assistant. Summarise this text and make it good and professional: [text]
Stronger prompt
Summarise the text between <<< and >>> for [AUDIENCE]. Rules: five bullets maximum; keep every number, name and date exactly as written; end with one line — “Decision needed:” or “No decision needed”. If the text is missing context, say so in one line instead of guessing. Text: <<< [PASTE TEXT] >>>
Why the stronger one works
- The weak version asks for a quality (“good”) and leaves the shape to chance; the template fixes shape, length and fidelity rules.
- AUDIENCE is a slot: one required answer, visible when it is missing.
- The last rule handles the common failure — the model filling gaps — by making “say so” an acceptable answer.
Keep it honest, retire it on time
Templates drift. A rule that made sense for last quarter’s task becomes cargo the next; a product name in the fixed text goes stale; the folder grows to forty templates and you use three. Review a template when you use it: did it still do what its name says? If you skip a slot twice, delete the slot. If you skip a template twice, delete the template. Reusable assets earn their place through reuse, not through existing.
Practice
Build the template
Your support team reads the week's help-desk conversations every Monday and sends a short digest to product. The prompt that worked last week was improvised in a hurry; this week you are turning it into a template the whole team can run.
Fill the five template slots — job, inputs, fixed rules, output shape and check line — then compare with the reference version.
A strong example (self-assessment — not machine-graded)
Job: turn one week of support conversations into a product-facing digest of recurring problems, with evidence. Inputs: [WEEK_START] to [WEEK_END], and <<<CONVERSATIONS>>> pasted as one block, one conversation per section, each starting with its date. Fixed rules: group by product problem, never by customer; include a problem only if it appears in three or more conversations; quote at most twenty words verbatim per problem; never state or estimate usage numbers — describe frequency as “several”, “a handful”, “widespread”; if the block is missing sections or unreadable, say so instead of summarising around it. Output: five bullets maximum, each in the shape “problem → observed frequency described in words → one representative quote with date”; end with a one-line “Not covered this week:” note when the input was incomplete. Check line: before answering, verify every bullet quotes the conversation block and that no number appears that is not a date or a quote.
Hint
Fixed text is what you would correct a new teammate on; slots are what they would ask you for.
The check line names the failure you are most afraid of: invented numbers, guessed frequency, generalised mush.
Why this is the answer
A working template separates the two kinds of text: rules you would defend in review, and inputs that change weekly. Slots make missing data visible, the format line protects the reader, and the check line turns your worst failure mode into a stated instruction. Fill it once this way and the Monday digest stops depending on who is typing.
Transfer
- Weekly reports: fix the structure once, slot the data; the check line carries the fidelity rules.
- Feedback and reviews: the fixed rules are the tone boundaries; examples carry the shape.
- Support macros: the same construction — fixed policy lines, open customer specifics.
- Your own prompts: two uses earns a template; two skips earns a deletion.
Key takeaways
- Templates are built after a prompt works twice — fixed rules, named slots, one check line.
- A visible placeholder gets answered; an invisible gap gets invented.
- Review on use: if a slot or template goes unused twice, cut it.
Next
- Keep the reps going on the practice floor, or jump to the research track and learn to verify what a template produces.
- Related: Examples and few-shot is the deep version of “add one example”.
- Terms: prompt template, placeholder and few-shot prompting.