Cheatsheet Phase 1 - Claude Fondamentaux

import anthropic

client = anthropic.Anthropic()  # ANTHROPIC_API_KEY env

message = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    system="Tu es un expert en architecture.",
    messages=[
        {"role": "user",
         "content": "Explique les microservices"}
    ]
)
print(message.content[0].text)

* = requis

{
  "id": "msg_01XFDUDYJgAACzvnptvVoYEL",
  "type": "message",
  "role": "assistant",
  "content": [{
    "type": "text",
    "text": "Les microservices sont..."
  }],
  "model": "claude-sonnet-4-5-20250929",
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 25,
    "output_tokens": 150
  }
}

with client.messages.stream(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    messages=[{"role": "user",
               "content": "Ecris un poeme"}]
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Events SSE: message_start, content_block_start, content_block_delta, message_delta, message_stop

import base64
with open("image.png", "rb") as f:
    data = base64.standard_b64encode(f.read()).decode()

msg = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    messages=[{"role": "user", "content": [
        {"type": "image", "source": {
            "type": "base64",
            "media_type": "image/png",
            "data": data}},
        {"type": "text",
         "text": "Decris cette image"}
    ]}]
)

Formats: JPEG, PNG, GIF, WebP | Max 20MB/image

Headers: x-ratelimit-limit-requests, x-ratelimit-remaining-requests, retry-after

import time

def call_with_retry(messages, retries=3):
    for i in range(retries):
        try:
            return client.messages.create(
                model="claude-sonnet-4-5-20250929",
                max_tokens=1024,
                messages=messages)
        except anthropic.RateLimitError:
            time.sleep(2 ** i)
    raise Exception("Max retries")

msg = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    system=[{
        "type": "text",
        "text": "LONG SYSTEM PROMPT...",
        "cache_control": {"type": "ephemeral"}
    }],
    messages=[{"role": "user",
               "content": "Question"}]
)

batch = client.messages.batches.create(
    requests=[
        {"custom_id": "r1",
         "params": {
            "model": "claude-sonnet-4-5-20250929",
            "max_tokens": 1024,
            "messages": [{"role":"user",
              "content":"Q1"}]
        }},
        # ... plus de requetes
    ]
)
# Poll: client.messages.batches.retrieve(batch.id)

Use case: Traitement non-urgent en masse

• 1. Clarte - Instructions directes, sans ambiguite
• 2. Specificite - Preciser le format, la longueur, le style
• 3. Structure - Utiliser XML, listes, sections
• 4. Exemples - Fournir des exemples de sortie attendue
• 5. Iteration - Tester, mesurer, ameliorer

# Zero-shot CoT
"Resous ce probleme etape par etape:"

# Structured CoT
"""Analyse cette architecture:
1. Identifie les composants
2. Evalue chaque composant
3. Identifie les faiblesses
4. Propose des ameliorations
5. Justifie chaque choix"""

Quand: Raisonnement complexe, math, logique, debug

"""Classifie le sentiment:

Texte: "Super produit!" -> Positif
Texte: "Decevant" -> Negatif
Texte: "Correct, rien de special" -> Neutre

Texte: "J'adore cette formation!" -> """

<context>
Systeme e-commerce, 1M utilisateurs
</context>

<task>
Identifie les 3 risques architecturaux
</task>

<format>
JSON: {risk, severity(1-5), mitigation}
</format>

<constraints>
- Maximum 200 mots par risque
- Inclure des exemples concrets
</constraints>

"""Tu es un [ROLE] avec [EXPERIENCE].

CONTEXTE:
[Description du contexte]

REGLES:
- Regle 1
- Regle 2
- Regle 3

FORMAT DE SORTIE:
[Format attendu]

CONTRAINTES:
- Ne pas [contrainte 1]
- Toujours [contrainte 2]

EXEMPLES:
[Exemple entree] -> [Exemple sortie]
"""

• Litteralite accrue - Suit les instructions exactement
• Moins de refus - Moins de faux positifs safety
• Format strict - Respecte le format demande
• Meilleur coding - SWE-bench SOTA
• Extended Thinking - Raisonnement en etapes visibles
• Adaptive Thinking - 4 niveaux: low/medium/high/max

Tache simple?
  OUI -> Haiku 4.5 ($1/$5)
    Classification, extraction, routing
  NON ->
    Besoin max qualite?
      OUI -> Opus 4.6 ($5/$25)
        Agents, raisonnement, code complexe
      NON -> Sonnet 4.5 ($3/$15)
        Coding, analyse, generation

Formule:

cout = (input_tokens / 1M * prix_input)
     + (output_tokens / 1M * prix_output)

Exemple Sonnet:

1000 tokens in + 500 tokens out
= (1000/1M * $3) + (500/1M * $15)
= $0.003 + $0.0075
= $0.0105 par requete

10K requetes/jour = $105/jour

# Python
pip install anthropic

# TypeScript/JavaScript
npm install @anthropic-ai/sdk

# Claude Code CLI
npm install -g @anthropic-ai/claude-code

# Variable d'environnement
export ANTHROPIC_API_KEY=sk-ant-api03-...

# Verification Python
python -c "import anthropic; print(anthropic.__version__)"