Prompt library
Copy-ready prompts that summarise, extract, classify, and rewrite with less waste. The biggest wins usually come from capping output, skipping repeated context, avoiding regenerate loops, and routing simple work to lighter models — not only from shorter wording.
Pair these with the token cost calculator or request an AI cost audit.
Force short outputs and hard length caps so replies do not ramble.
Caps output tokens and blocks narrative filler that burns credits on every long document.
When: Meeting notes, articles, tickets, research dumps.
Summarise the text below for a busy teammate. Rules: - Maximum [N] bullets - Maximum [M] words total - No preamble, no closing paragraph - Prefer facts, decisions, owners, and dates - If something important is unclear, add one final bullet starting with "Unknown:" Text: """ [PASTE_TEXT] """
Stops the model from writing a full essay when a two-line answer is enough.
When: Any long input where you may not need depth yet.
Give a TL;DR of the text in at most 2 sentences. Then stop. Only if I reply "details", expand into at most 5 bullets covering: key points, risks, and next actions. Text: """ [PASTE_TEXT] """
Avoids restating unchanged context; output focuses on deltas only.
When: Document revisions, sprint updates, contract redlines.
Compare PREVIOUS and CURRENT. Return only: 1. What changed (bullets) 2. What stayed the same (max 2 bullets) 3. Risks or follow-ups (bullets) Do not rewrite either document. Max [N] bullets total. PREVIOUS: """ [OLD_TEXT] """ CURRENT: """ [NEW_TEXT] """
Pull only the fields you need into JSON or tables instead of prose.
Structured output is shorter, cheaper to parse, and avoids essay-style extraction.
When: CRM enrichment, intake forms, invoice/email parsing.
Extract fields from the text into JSON only. No markdown. No commentary.
Schema:
{
"company": string | null,
"contact_name": string | null,
"email": string | null,
"intent": string | null,
"budget": string | null,
"deadline": string | null,
"open_questions": string[]
}
Rules:
- Use null when unknown
- Do not invent values
- Keep strings short
Text:
"""
[PASTE_TEXT]
"""Limits both reasoning and formatting tokens by requesting a narrow table.
When: Competitor lists, feature matrices, research notes.
Convert the text into a markdown table with ONLY these columns: [COLUMN_1] | [COLUMN_2] | [COLUMN_3] Rules: - One row per item - If a cell is unknown, write "—" - No intro or outro - Max [N] rows; if more exist, keep the most relevant and note "truncated" Text: """ [PASTE_TEXT] """
Skips transcript restating and returns a compact task list.
When: Calls, standups, long Slack threads.
Extract action items only. Format each line as: - [Owner or Unassigned] — [Action] — [Due date or None] — [Priority: H/M/L] Rules: - No meeting summary - No duplicate actions - Max [N] items - If none, reply exactly: No action items found. Transcript: """ [PASTE_TEXT] """
Route or label items with tiny outputs that cheaper models handle well.
Classification is ideal for light models and tiny completions.
When: Support tickets, lead routing, content tagging.
Classify the text into exactly one label from this list:
[LABEL_1], [LABEL_2], [LABEL_3], [LABEL_4], Other
Return JSON only:
{"label":"...","confidence":"high|medium|low","reason":"≤12 words"}
Text:
"""
[PASTE_TEXT]
"""Replaces long incident write-ups with a short priority decision.
When: Bugs, outages, inbound requests.
Triage the issue. Return exactly: Priority: P0|P1|P2|P3 Impact: one sentence Next step: one sentence Needs human: yes|no Definitions: - P0: service down / data loss / security - P1: major feature broken, no workaround - P2: impaired with workaround - P3: minor / cosmetic / question Issue: """ [PASTE_TEXT] """
Prevents launching a costly long generation when a cheap check is enough.
When: Before rewriting, researching, or multi-step agent runs.
Answer with YES or NO only, then one short reason ≤15 words. Question: Does this text contain enough information to [TASK] without guessing? Text: """ [PASTE_TEXT] """
Edit existing text instead of regenerating long drafts from scratch.
Editing existing copy is cheaper than regenerating a full draft repeatedly.
When: Landing pages, emails, docs that are almost done.
Tighten the draft. Goals: - Cut at least [PERCENT]% of words - Keep meaning and required claims - Improve scannability - Preserve [MUST_KEEP_PHRASES] Return only the revised text. No commentary. Draft: """ [PASTE_TEXT] """
Stops models from 'helpful' expansion when you only need a reshape.
When: Bullets ↔ email, notes ↔ checklist, memo ↔ Slack.
Convert the text into [TARGET_FORMAT]. Rules: - Do not add new ideas - Do not expand explanations - Keep names, numbers, and constraints exact - Output only the converted text Text: """ [PASTE_TEXT] """
Avoids full rewrites for one fix; fewer tokens in and out.
When: Tone fixes, CTA changes, compliance wording.
Edit the draft to fix ONLY this issue: [ISSUE]. Do not change anything else. Return the full updated draft only. Draft: """ [PASTE_TEXT] """
Get decisions and trade-offs first; details only when requested.
Forces a decision-first structure that light models can do well.
When: Build vs buy, vendor choice, prioritisation.
Create a decision memo with this exact structure and length limits: Decision needed: ≤20 words Recommendation: ≤25 words Why: 3 bullets max Risks: 2 bullets max Next step: 1 bullet with owner placeholder No other sections. Context: """ [PASTE_CONTEXT] """
Tables compress comparison tokens versus narrative analysis.
When: Tooling choices, campaign variants, process options.
Compare options in a markdown table with columns: Option | Fits goal | Cost/effort | Risk | Verdict Then add exactly 3 bullets: - Best default - When to pick alternative - Missing info Goal: [GOAL] Options: [OPTION_LIST] Constraints: [CONSTRAINTS] Context: """ [PASTE_CONTEXT] """
Bounded critique prevents long stylistic rants and endless regenerate loops.
When: Reviewing drafts before another expensive rewrite.
Review the draft against this checklist only: 1. Clarity of main point 2. Evidence / specificity 3. Audience fit for [AUDIENCE] 4. CTA / next step clarity 5. Brand or claim risk For each item return: Pass | Needs work — ≤12 words. Then list max 3 concrete fixes. Do not rewrite the draft. Draft: """ [PASTE_TEXT] """
Default prompts that prevent padded openings and repeated context.
Removes throat-clearing intros that show up in almost every chat completion.
When: Default system-style instruction for everyday Q&A.
Answer the question directly in the first sentence.
Then add at most 3 short bullets if needed.
No preamble ("Great question", "As an AI", "Certainly").
If you lack data, say what is missing in one line.
Question: [QUESTION]
Context (optional):
"""
[CONTEXT]
"""Stops models from echoing the prompt back, which doubles output cost.
When: Any prompt that includes long pasted context.
Use the context to answer. Do not quote or restate it unless I ask. Return: - Answer - Assumptions (max 2) - Next question (only if blocked) Question: [QUESTION] Context: """ [PASTE_CONTEXT] """
A cheap outline pass prevents multi-thousand-token wrong drafts.
When: Before long articles, specs, or multi-file generations.
Before writing the full deliverable, propose a brief plan only: 1. Goal in one sentence 2. Outline with max [N] sections 3. Assumptions 4. What you need from me (max 3 questions) Wait for my "go" before producing the full draft. Request: [REQUEST] Constraints: [CONSTRAINTS]
Intrascope shows cost by user, project, and model, with controls that keep expensive models from becoming the default.
Prefer a softer start? Request a free AI cost audit.