PUMA Vault

RCOIF — Role·Context·Objective·Instructions·Format

Se lee en 3 min

RCOIF — Role · Context · Objective · Instructions · Format

Atomic Claim

Overview

Structuring every AI prompt into five components (Role, Context, Objective, Instructions, Format) consistently produces higher-quality, more predictable outputs than unstructured prompting.

💡 The Framework

RCOIF is a structured prompting methodology that decomposes every prompt into five mandatory components:

ComponentPurposeKey question
R — RoleDefine AI’s persona and expertiseWho are you in this interaction?
C — ContextProvide all relevant backgroundWhat does the AI need to know?
O — ObjectiveState the desired outcome preciselyWhat exactly do we want to achieve?
I — InstructionsStep-by-step task breakdownHow should the AI proceed?
F — FormatSpecify the output structureWhat form should the answer take?

The RCOF variant omits the Objective as a separate component (merging it with Instructions), used for simpler tasks.

🔬 Why It Works

The framework exploits three properties of large language models:

  1. Role anchoring — Priming the model with a specific role activates domain-relevant knowledge and reduces generic responses
  2. Context windows are attention windows — What you put in context shapes what the model attends to
  3. Format specification — Explicit output format constraints reduce hallucination and improve parsability

📋 Template

ROLE:
You are a [specific expert type] with expertise in [domain].
Your approach is [methodological stance].

CONTEXT:
[Background information: project, constraints, prior work, definitions]
[What has already been established]
[What the user/researcher brings to the interaction]

OBJECTIVE:
[Single, precise goal statement]
[Success criterion]

INSTRUCTIONS:
1. [First action]
2. [Second action — reference CONTEXT where needed]
3. [Third action]
[Continue until complete]

FORMAT:
- Structure: [e.g., markdown with H2 headers per section]
- Length: [e.g., 500-800 words]
- Language: English
- Include: [mandatory elements]
- Exclude: [what to omit]
- Tone: [academic | technical | conversational]

🧩 Application to PUMA

In PUMA, RCOIF is used at two levels:

Level 1 — Research assistance prompts (Claude, Perplexity, DeepSeek)

  • Role: domain expert in software engineering / empirical research
  • Context: PUMA project specifics, dataset properties, prior findings
  • See: PT-Claude-RCOIF-Research

Level 2 — Experiment prompts (Ollama Llama3.2, Mistral7B)

  • These are the prompts the benchmark itself uses for triage and estimation
  • The prompting strategy IS the independent variable being tested
  • See: 60 - Resources/61 Prompts/PT-PUMA-Triage-ZeroShot

🔗 Connected Ideas

Extends: PN-CoT-FewShot-Prompting (CoT adds reasoning steps to the I component) Used in: PN-EGI-Framework · PN-AMI-DRCA-IIPR-Frameworks Contrasts with: Zero-shot unstructured prompting Applied to PUMA prompts: PT-Claude-RCOIF-Research · PT-PUMA-Experiment-Prompts Experiment context: PR-PUMA-Ch3-Methods (§3.4) · EX-Hypotheses-H1-H2 MOC: MOC-Methods-Frameworks

⚠️ Caveats

  • RCOIF improves consistency, but does not guarantee accuracy — outputs still require human validation
  • Over-specified Format constraints can suppress creative or unexpected useful responses
  • For very simple one-shot queries, the full framework adds unnecessary friction — use RCOF

Vista Gráfica

Retroenlaces