📦 The block library (copy-paste)
In 2.1, you assembled an MVP with three blocks. Now we're opening the whole toolbox. Each block here is text ready to copy — distilled from a real pattern — with the {spaces} waiting for your case. There’s just one rule for using it: don’t paste them all. Pick only the ones your case calls for, and expand later.
Detailed content
The identity block
This is the block that opens any system prompt. It answers three questions in a few lines: who the AI is, where she operates, and what for — the goal in one sentence. Without it, the model doesn’t know what role it’s playing, and everything that follows is untethered. Copy it, change what’s {inside braces}, paste it.
Você é {NOME}, {PAPEL}, operando em {AMBIENTE}.
Modelo: {id}. Data de hoje: {data}. Conhecimento até: {corte}.
Seu objetivo principal é {objetivo em uma frase}.
You are {NAME}, {ROLE}, operating in {ENVIRONMENT}.
Your main goal is to follow the user's instructions to {GOAL}.
Why the date and the knowledge cutoff
Lines like Today's date e Knowledge up to prevent the model from guessing the wrong year or confidently presenting something from before the cutoff as if it were current. If your use case doesn’t need it, remove the line — but it costs little and saves you from a silly mistake.
Tone + honesty (the highest ROI)
If you could only paste one block beyond identity, this would be it. It combines two things: how the AI sounds (concision, prose instead of lists) and how it behaves with the truth (no flattery, own the mistake). Just a few lines, and they already cut out half the answers that make you angry.
<tom> - Responda no menor tamanho que resolva. Em conversa casual, poucas frases bastam. - Use prosa, não listas, para explicações e relatórios — listas só quando o usuário pedir ranking/passos. - Verdade e honestidade acima de deferência. Sem bajulação, sem reasseguramento artificial, sem encher linguiça. - Discorde de forma construtiva quando achar que algo está errado — gentil, mas sem esconder o feedback. - Ao errar, assuma e corrija, sem auto-flagelação nem desculpas em excesso. </tom>
You avoid cheerleading, motivational language, or artificial reassurance, or any kind of fluff. Truthfulness and honesty are more important than deference and sycophancy. When something is wrong, point it out kindly without hiding the feedback. When you make a mistake, own it and fix it without collapsing into self-abasement or excessive apology.
✓ What it turns on
- ✓Responses sized to the problem, not longer
- ✓Prose you can read, instead of reflexive lists
- ✓Constructive disagreement when you're wrong
- ✓Admit the mistake without drama or ten apologies
✗ What it turns off
- ✗Flattery ("What a great question!") in every turn
- ✗Artificial reassurance and padding
- ✗Agree with you just to please you
- ✗Excessive self-flagellation after a slip
Why “higher ROI”
Five lines change the entire character of the responses. No other block delivers so much impact with so few words — which is why it goes almost always, even in a chat-only PMV.
Concision budget
The tone block already calls for concision. This goes further: it gives a number. When "be concise" isn’t enough — because the AI and you disagree on what "concise" means — you set a budget: so many sentences, so many bullets, and a clear rule for when it can overflow. This is the block for anyone who needs predictable response length.
Responda no menor tamanho que resolva. Padrão: {N} frases / {N} bullets de uma linha.
Cada bullet começa com o entregável/ação em **negrito**, seguido de um qualificador curto.
Expanda apenas para: código, segurança, ou quando o usuário pedir explicitamente.
1. Set the default number
Replace the {N} based on what your case calls for. A support chat response? "3 sentences / 3 bullets." An executive summary? "5 one-line bullets." The number becomes the contract.
2. Standardize the bullet format
"Deliverable in bold + the short qualifier" makes each item start with what matters. The reader scans the first word and immediately knows whether that bullet is for them.
3. List the exceptions, and only those
Without an escape hatch, AI mutilates code or skips a safety warning to fit within the limit. The line Expand only for: code, security, or when the user asks solves it: this can grow here, nowhere else.
Tone gives direction; concision gives a number
You don’t always need both. For casual chat, the tone that’s enough. Pull the concision budget when response length is a concrete problem and you want an explicit ceiling.
Priority hierarchy
Sooner or later, two instructions will fight: memory says one thing, the user asks for another; a style preference conflicts with a safety rule. Without a declared order, the model decides in the moment—and not always the way you wanted. This block sets the precedence at a time, so the tie-breaker is always the same.
Em conflito de instruções, a ordem de precedência é:
{segurança} > {política inegociável} > {instrução explícita do usuário} > {defaults do sistema} > {preferência de estilo}.
Quando uma regra de memória/contexto contradisser uma instrução direta do usuário, o usuário vence.
🪜 The chain, from strongest to weakest
The line that matters most
The last sentence — when memory contradicts the user, the user wins — prevents the classic bug: the AI clinging to an old preference and ignoring what you just asked. The present takes precedence over the past.
Citation / source preservation
When AI uses information from outside — a search, a document, a page — it needs to say where it came from. But there’s a right way to cite: inline, with the source name as the link text, never a raw URL or a "click here." This block gives the text the feel of serious research instead of a patchwork.
Toda frase com informação derivada de fonte externa cita inline: [nome da fonte](url). O texto âncora é o nome da fonte ou uma frase descritiva natural — nunca "fonte"/"link" nem URL crua. Citações inline, não numa seção "Referências" separada. Teste: o texto deve ler naturalmente mesmo se todas as URLs fossem removidas.
✓ Citation done right
- ✓Natural anchor text: "according to the React documentation"
- ✓Link embedded in the sentence itself, flowing with the text
- ✓It can be read in full even if the URLs disappeared
- ✓Without a "References" block tacked on at the end
✗ Citation done wrong
- ✗Meaningless "click here" / "source" / "[1]" anchor text
- ✗Raw URL dropped into the middle of the text
- ✗Text that turns into Swiss cheese without the links
- ✗Everything piled into a loose "References" section
The test embedded in the block
The last line is a ready-made test: remove all URLs and read. If the text still makes sense, the citation is integrated as it should be. If it becomes a pile of "see [link]" references, the AI cited it incorrectly.
How to combine blocks — the 3 recipes
Loose blocks don’t make a prompt — assembly does. The library includes three recipes starting point, one for each common type of AI. Each recipe is a list of blocks in order. Find your case, copy the blocks from that row, and you have an MVP. Remember the golden rule: don’t paste them all.
1 IDENTIDADE + 2 TOM + 4 CONTRATO + 5 LOOP + 6 RECUPERAÇÃO + 7 RELATO + 11 CONFIRMAÇÃO
The “heavier” recipe — because an agent that edits files and runs commands needs discipline: the loop to work in cycles, the recovery so you don’t repeat the mistake, the faithful report so you don’t disguise the result, and the confirmation before the irreversible action.
1 IDENTIDADE + 3 CONCISÃO + 13 CITAÇÃO + 9 PRIORIDADE
For an AI that searches and summarizes: brevity with a number (answers that fit), citation (every claim anchored in the source), and priority (when memory and the new request conflict, the new request wins). Note: no loop and no contract tools — it informs, it doesn’t execute.
1 IDENTIDADE + 12 PERSONA + 2 TOM + 8 SEGURANÇA
The brand chatbot recipe: persona (values, tone, way of guiding), tone and a handful of safety. This is exactly the foundation of what you built in 2.1 with Mia—and it’s the one you see most often in products.
| Recipe | When to use | Is there a loop? |
|---|---|---|
| A · Coding agent | Edits files, runs commands, executes in steps | Yes |
| B · Research assistant | Searches, summarizes, and cites external sources | No |
| C · Product persona | Branded chatbot, conversation only, no tools | No |
The recipe is a starting point, not a law
The three cover most cases, but they are ground, not a ceiling. Start with the one that most resembles your case, run it, and only then add a block when you see genuinely bad behavior — as you learned in 2.1.
📦 Hands-on: build the library
Return to the MVP you built in 2.1 and replace it with one of the three recipes—copying the blocks from this library verbatim.
- 1 Decide which of the three recipes is yours: coding agent, research assistant, or product persona. That determines the blocks.
- 2 Copy each block from the recipe line exactly as shown here, and fill in the {keys} for your case.
- 3 Use the PT or EN version according to your prompt — when both are available, choose the language of the rest of your system prompt.
- 4 Run it on real tasks. Resist pasting in extra blocks “just in case” — only add them when a real error appears.
📝 Module Summary
- ✓Identity: open the prompt with who/where/goal in one sentence (in PT and EN)
- ✓Tone + honesty: the highest-ROI block—concision, prose, and truth over flattery
- ✓Concision budget: when "be concise" isn't enough, set a {N} with exceptions
- ✓Priority hierarchy: the tie-break order—safety at the top, style yields first
- ✓Source citation: inline, a natural anchor, and the test of removing the URLs
- ✓3 recipes: code, research, persona—copy the line, but don’t paste them all
➡️ Next Module
2.3 — 🎭 Persona design. You have the blocks. Now let’s go deep on one of them: how to design a persona that gives AI character without turning it into fantasy—values, tone, interaction, and escalation, plus the golden rule that separates persona from artifact.