Skip to main content

Apps Content Stack Blog

Prompt Architecture · Companion Tool

Sequence beats technique.

Pick prompting techniques and let Gemma 4 synthesise them into a coherent, ready-to-use prompt — not a template dump. Every prompt comes with a sequence rationale explaining why each layer sits where it does.

Task

required

What do you want the AI to do?*

Canonical stacks

quick presets

Layer selection

auto-sequenced

0 / 10 layers

Layer 6 is a reasoning scaffold — pick one, not multiple.

1

Role / Persona

Assign specific expertise or voice

Identity must precede reasoning. The same situation reads differently as a lawyer vs a poet.

Specific role

2

Context Framing

Background, audience, constraints

Context fixes the principle space before reasoning begins.

Context (audience, stakes, constraints)

3

Step-Back

Derive principle before specifics

Prevents pattern-matching to surface features. Forces the model to name the discriminating variables.

4

XML Input Structure

Tag variable inputs semantically

Claude was trained heavily on XML. Tag boundaries are reliably attended to.

Input tag name

5

Few-Shot Examples

2-5 worked examples with edge cases

Examples define target behaviour. The last example weighs most — put the trickiest case last.

Number of examples

Example input tag

6

Chain-of-ThoughtLayer 6 alt

Explicit step-by-step reasoning

Default reasoning scaffold. Cheap, reliable, well-supported.

6

DecompositionLayer 6 alt

Fixed ordered sub-steps

Beats CoT when steps are known and the failure mode is skipping or reordering them.

Sub-steps (one per line, 3-5 max)

6

Tree-of-ThoughtsLayer 6 alt

Multiple reasoning paths, then synthesise

Heavy weapon. Use for high-stakes decisions with genuine ambiguity. ~3× token cost.

Number of independent paths

6

ReAct LoopLayer 6 alt

Thought → Action → State per turn

For long iterative sessions. The <state> block is the killer feature — persistent scratchpad across turns.

7

Constitutional Rules

3-7 explicit guardrails

Falsifiable rules outperform vague guidance. Phrase as positive constraints where possible.

Rules (one per line, 3-7 max)

8

Output Contract

Format, schema, length

Most underused technique. Be specific — show the schema literally if structured.

Output format spec

9

Self-Critique

Draft → critique → revise

Reliably improves creative and analytical output. ~2× token cost.

9

Reflexion

Reflect on prior attempt before retrying

Multi-turn cousin of self-critique. Useful when iterating on hard problems across turns.

10

Prefill

Seed the first tokens of the response

Most reliable single technique for output discipline. Always last — by definition.

Prefill text (literal first tokens)

⚙ = Generate as system prompt / project context

Describe your task

plain English

What do you want the AI to do?*

Task type override (optional)

⚙ = Generate as system prompt / project context

No prompt yet

Pick techniques on the left and hit Generate. Gemma 4 will synthesise them into a coherent prompt.

No rationale yet

Generate a prompt first to see the sequence rationale.

Built on the 10-layer cognitive pipeline. Companion to the essay Prompt Architecture — Why Sequence Beats Technique.

vibelife.sg · 2026

VibeLife — a life, tuned.

vibelife.sg
Blog
Apps
Contact

© 2026 · Built in SG