AI Workflow Reference

Upgraded AI Workflow

A 3-layer operating system — always-on instructions, a single prompt architecture, and modular playbooks, with quality gates that keep every output defensible.

1

Project Custom Instructions — the OS

Always-on. Governs every prompt. This is what you already built — upgraded.

What changes
Before
Role defined once, applied inconsistently
Role loads before every module, non-negotiable
Before
Push back only when prompted
Push back is automatic, especially on weak framing
Before
Research and synthesis mixed together
Research → Diagnosis → Strategy always separate
Before
Quality varies by prompt quality
Quality gates enforced after every phase
The 5 operating rules — locked in
  • Source-first: fetch and analyze the company before answering. Never invent facts. Label assumptions explicitly.
  • Structured outputs only: Observation → Insight → Implication → Recommendation. Never blended.
  • No filler: every sentence must answer "why does this matter commercially?" No buzzwords, no empty language.
  • Sequential inheritance: each module builds on all prior ones. Do not repeat — extend.
  • Push back automatically: if positioning is generic, name it. If evidence is missing, bracket it. If framing mixes objectives, say so.
2

Prompt Architecture Standard

Every prompt follows one format. This is what makes the workflow scalable and reusable.

Master prompt structure — 7 fields, every time
Context

Why this work exists. One sentence. Links the task to a commercial outcome.

Background

Company, role, scenario. What the AI needs to know before it starts.

Task

Exactly what to produce. Verbs only. "Analyze X." "Produce Y." No ambiguity.

Scope

Which sources to use. Which pages, files, modules, or previous outputs to inherit.

Framework

Which methodology applies: April Dunford, Emma Stratton, JTBD, GEO, PMA, etc.

Constraints

What to avoid. No buzzwords. No generic language. No competitor names as villains.

Expected output + quality bar

Exact structure of the output, plus what "good" looks like. Forces the AI to self-evaluate.

Why this matters for Viewz
  • You can brief a new company in 2 minutes — just replace the company URL and name.
  • Every module inherits from prior modules — no repeated research.
  • The structure is tool-agnostic: runs on Claude, ChatGPT, or any long-context AI.
3

Playbook Modules

Four phases. Each module is a self-contained unit. Run in sequence or on demand.

Phase 1 — Research (7 modules)
Brand Intelligence
Positioning Audit
Messaging Audit
Competitor Intelligence
ICP + Persona
Market Intelligence
GTM Motion Audit
Phase 2 — Diagnosis
Cross-source synthesis
What is generic / missing
GEO / AI visibility
Content funnel gaps
Phase 3 — Strategy
Positioning statement
Messaging house
Homepage narrative
ICP prioritization
Category POV
GTM recommendations
Phase 4 — Activation assets
Executive POV memo
30/60/90 plan
Homepage teardown
Mock presentation
Villain narrative deck
Cowork SOW brief
Full workflow — URL to content
Input
Company URL
Research
Docs 1–16
Synthesis
Messaging house
Output
Content + assets

Quality gates

Run after every phase. If any gate fails, do not proceed. Revise first.

G1
Company-specificCould this analysis apply to any competitor in the category? If yes — reject it. Every output must be specific to this company.
G2
Separated layersAre observations, insights, implications, and recommendations clearly distinct? Blended outputs are not analysis — they are summaries.
G3
Buyer languageWould your ICP read this and recognise their own situation in their own words? No jargon the buyer would not use themselves.
G4
Evidence labeledAre unverified metrics, assumed facts, or invented personas explicitly bracketed? No invented evidence slips through as if it were real.
G5
Commercial relevanceDoes every finding connect to a commercial outcome? If a section does not answer "why should the business care?" — cut it.