LN: Shum (2025) — Pensar con Prompts

Bibliographic Reference

Citation: Shum, J. (2025). Pensar con prompts: La guía definitiva de la ingeniería de prompts. Independently published. Note: Independently published in 2025; available via major online book retailers. No official author website confirmed at time of writing. Verify current availability via retailer search.

Metadata Note

This book is self-published. Author details and edition information reflect vault metadata. The Spanish-language original title is preserved; the summary below is in English per PUMA convention.


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategoryPractical guide + methodology compendium
ContextSpanish-language comprehensive reference for prompt engineering techniques; fills a gap in accessible Spanish-language AI resources
CorrectnessSynthesizes published techniques (CO-STAR, CoT, Self-Consistency, ReAct, etc.) with practical examples
Contributions(1) CO-STAR framework formalization in Spanish; (2) Curated taxonomy of 20+ prompting techniques; (3) Applied examples for business and research contexts
ClarityExcellent. Structured for both beginners and practitioners.

Relevance: ⭐⭐⭐⭐

The CO-STAR framework in this book is the primary structural guide for PUMA’s prompt design. The taxonomy of techniques maps directly to PUMA’s experimental prompt strategy matrix.


Pass 2 — Content

CO-STAR Framework (Detailed)

Shum operationalizes CO-STAR as a six-section prompt template:

[C] CONTEXTO: Quién eres y cuál es el trasfondo de la tarea
[O] OBJETIVO: Qué quieres que el modelo haga exactamente
[S] ESTILO: Cómo debe escribir o responder
[T] TONO: El registro emocional o actitudinal de la respuesta
[A] AUDIENCIA: A quién va dirigida la respuesta
[R] RESPUESTA: El formato exacto que esperas

Key principle: The [R] section (Response format) is the most critical for programmatic use. Specifying JSON schema in [R] reduces format errors by 60–80% compared to unstructured prompts.

Prompting Technique Taxonomy

TechniqueCategoryWhen to UsePUMA Status
Zero-ShotBasicSimple, well-defined tasksH1/H2 baseline
Few-ShotBasicPattern matching, classificationH1/H2 primary
Zero-Shot CoTReasoningMulti-step inferenceH1 augmented
Few-Shot CoTReasoningComplex estimation with examplesH2 primary
Self-ConsistencyReliabilityHigh-stakes, ambiguous casesH2 production
ReActAgenticTool use + reasoningStage 4+
COSTARStructureAny LLM taskAll PUMA prompts
Role PromptingPersonaSpecialized expert behaviorPM agent persona
Meta-PromptingAdvancedSelf-improving promptsResearch extension
Tree of ThoughtsAdvancedComplex multi-path decisionsFuture work

Structured Output Best Practices

  1. Always include a concrete JSON example in the prompt — not just a schema description
  2. Specify field constraints explicitly: "priority": "one of [Critical, High, Medium, Low]"
  3. Include a “rationale” field even if not consumed — forces model to reason before classifying
  4. Add an escape hatch: "confidence": 0–1; if confidence < 0.6, set type to 'uncertain'

Prompt Anti-Patterns (to avoid in PUMA)

  • Vague instructions: “Analyze this issue” → “Classify this issue as Bug, Feature, Task, Improvement, or Sub-task”
  • Missing format specification: No JSON schema → inconsistent outputs, parse failures
  • Role without persona: “You are an AI assistant” → “You are a senior project manager with 10 years of Agile experience”
  • No examples for classification: Zero-shot on subjective tasks → high variance outputs
  • Overlong prompts: >4K tokens for simple classification → model loses focus on task

Human-AI Co-Creation and Generative Cognition

Beyond the technical taxonomy, Pensar con Prompts advances a broader thesis about what prompting represents epistemologically: it is a form of generative cognition — thinking through and with an AI system rather than simply instructing it.

Key claims:

  • Prompts are not commands to a tool; they are the articulation of thought in a form that another mind (the model) can extend
  • The act of writing a good prompt requires the human to clarify their own thinking: vague prompts reflect vague intentions
  • The output of a well-designed prompt session is not just the AI’s response — it is the human’s clarified understanding plus the AI’s contribution

This framing positions prompt engineering as a cognitive practice, not merely a technical skill. The quality of human-AI interaction depends as much on the human’s ability to think clearly as on the model’s capability.

Implications for Human-AI Collaboration

Shum argues that the most valuable prompt engineers are not those who memorise technique lists — they are those who develop a genuine understanding of how language models “think” (statistically, contextually, pattern-matching on training distributions). This understanding enables:

  • Anticipating model failure modes before they occur
  • Diagnosing why a prompt produces poor results
  • Iterating deliberately rather than randomly
  • Designing prompts that surface the model’s genuine capabilities rather than averaging over its patterns

Generative AI and Cognitive Extension

The book closes with a philosophical claim: LLMs are the first general-purpose cognitive extension tools available to everyone. Unlike calculators (which extend numerical cognition) or search engines (which extend memory), LLMs extend the full range of linguistic-reasoning cognition. The implications are systemic: how humans think, write, and decide will change fundamentally in a world where generative AI is ubiquitous.


PUMA Integration

  • Prompt design: CO-STAR structure is the template for all PUMA H1/H2 prompts
  • Technique selection: Shum’s taxonomy provides vocabulary for PUMA’s experimental conditions
  • Anti-pattern checklist: Review all PUMA prompts against Shum’s anti-pattern list before experiment execution
  • Spanish-language resource: Relevant for PUMA academic context (Spanish university TFG)
  • Generative cognition framing: Supports PUMA’s claim that prompt engineering is a methodology, not just a parameter

MOCs