Architecting Prompt Workflow Systems
Given a rough draft prompt or raw concept, don't immediately expand it. First run the Post-Mortem Evaluation (see Workflow Step 1) to score it and expose structural gaps. Only after that, build the 5-Phase Prompt System (Steps 2-6). Output as chunked, sequential deliverables if the full system would exceed context — state "Remaining pending responses: N of N" at the end of each chunk.
Example invocation pattern:
Input: "Generate a market forecast using all data sources, be accurate and novel,
list what's missing from this prompt too."
Output: Scoring matrix + gap analysis + tactical rewrite recommendations
→ then, on request, the full 10-prompt phased system
Progress:
- Step 1: Post-mortem score the seed prompt/concept
- Step 2: Phase 1 — Ingestion & Strategic Scoping
- Step 3: Phase 2 — Human-in-the-Loop Triaging & Filtering
- Step 4: Phase 3 — Adversarial Stress-Testing (Red Team)
- Step 5: Phase 4 — Product/Concept Conceptualization & Blueprinting
- Step 6: Phase 5 — Execution & Continuous Monitoring
- Step 7: Advanced Concept Matrix (emergent/second-order effects)
Step 1: Post-Mortem Evaluation (always do this first)
Diagnose the seed prompt for conflated meta-goals — the most common failure is commanding the AI to execute a task and audit itself in the same breath. This causes execution drift.
Produce:
- Quantitative Scoring Matrix — score 5-6 dimensions (e.g., Methodological Intent, Data Silo Architecture, Variables & Modularity, Operational Feasibility, Prompt Telemetry & Safety) on a 0-10 scale, each with a one-line bottleneck note. Compute a composite score.
- Structured Gap Analysis — numbered list, each gap has: Why it's essential + Mechanism of Failure (how, mechanically, the LLM breaks without it).
- Tactical Recommendations — concrete rewrites: bracketed variable blocks, explicit negative constraints ("strictly forbid X"), and a two-pass architecture split (meta-prompt generator vs. execution prompts) if goals were conflated.
Never skip straight to "building the system" without this diagnostic — it anchors every downstream phase's guardrails.
Step 2-6: Build the 5-Phase System
Each phase = 2 prompts. Each prompt follows this exact template shape:
System Role: [Persona + one-sentence task]
Variables to Insert: [BRACKETED_TOKENS] with example values
Objective: [one sentence, outcome-oriented]
Deterministic Logic Chain: [3 numbered steps, each a concrete operation not a vague goal]
Output Blueprint: [named artifacts, e.g., "Scorecard Table", "Forensic Report" — never "a summary"]
Phase 1 — Ingestion & Strategic Scoping
- Contextual Anchor: converts org constraints (budget, moat, timeline, exclusions) into a Boundary Matrix + Moat Alignment Index.
- Exhaust Map Validator: audits candidate data channels for noise-to-signal ratio, manipulation risk (bots/SEO/PR), and proxy validity before any pattern-matching begins.
Phase 2 — Human-in-the-Loop Triaging
- Signal Sieve: scores raw AI-discovered anomalies against human conviction input; outputs Proceed/Pivot/Kill decisions via a Velocity ÷ Friction index.
- Bricolage Recombinator: forces pairing of Signal A (mechanic from unrelated domain) with Signal B (underserved pain point) into a novel hybrid concept.
Phase 3 — Adversarial Stress-Testing
- False Positive Dissector: actively tries to debunk the trend — checks for bot/seasonal/echo-chamber artifacts, outputs an "Artificially Inflated" probability score.
- Narrative Counter-Weight: builds the Bear Case — incumbent asymmetry, fatal barriers, and a falsification thesis (the exact metric that would kill the idea).
Phase 4 — Product Conceptualization & Blueprinting
- Meso-Scale Translator: bridges macro trend + micro signal into functional specs, respecting MVP constraints; outputs a Fragmented Task Log.
- Value Proposition Matrix: mirrors target-user vocabulary exhaust, contrasts against bloated incumbents, anchors pitch in a non-speculative value metric.
Phase 5 — Execution & Continuous Monitoring
- Exhaust Tracking Dashboard Spec: defines Safe/Warning/Critical thresholds per data channel; detects signal decay via rolling-window comparison to baseline.
- Feedback Loop Calibration: audits variance between projection and reality, isolates where the original prompt logic over-fit or hallucinated, outputs corrected prompt fragments to paste back into Phase 1-2.
Step 7: Advanced Concept Matrix (optional extension)
After the lifecycle system is specified, generate emergent/second-order effects: what new properties appear once all 5 phases run continuously (e.g., cross-domain epistemological bridges, asynchronous multi-agent decomposition, compounding calibration effects). Present as a short causal chain diagram + 3 numbered "Emergent Property → Second-Order Effect" entries.
Example 1: Input: A single dense paragraph prompt asking an AI to "find profitable trends using all data and also tell me what's wrong with this prompt." Output: Composite score 6.0/10, primary bottleneck = "Decoupled Meta-Diagnostic Framework" (execution + self-audit conflated), tactical fix = split into Pass 1 (meta-prompt generator) / Pass 2 (execution prompts), plus bracketed variable injection.
Example 2: Input: A concept document about treating craft/textile patterns as state machines. Output: Full 5-phase system where Signal = quantified loop/stitch instructions, Noise = narrative/style filler; Red Team objections include "spatial disconnect" and "logic inflation"; Monitoring phase tracks checkpoint error rate, latency, and reproduction velocity with numeric trigger thresholds.
- One abstraction layer per prompt. Never ask a model to simultaneously execute a task and audit its own instructions — split into passes.
- Bracket every variable.
[TARGET_INDUSTRY],[TIME_HORIZON]— never bury parameters in prose; this is what makes the system reusable/modular. - Anti-bloat clause is mandatory. Explicitly forbid marketing listicles, SEO fluff, and speculative press releases in any discovery prompt — otherwise web-search-enabled models default to superficial summaries.
- Every threshold must be a number. "Search volume acceleration >40% WoW," not "significant increase." Vague thresholds can't be automated or falsified.
- Close the loop. The system isn't done until Phase 5 outputs feed back into Phase 1-2 prompt text — otherwise it's a one-shot report, not a workflow.
- Chunk long builds. If generating all 10 prompts exceeds a reasonable response, split by phase and explicitly state remaining chunk count.
- Jumping straight to prompt-writing without scoring the seed draft first — you'll bake in the same structural gaps.
- Letting "novelty" and "big picture" language (golden goose, blue ocean, north star) stand unchecked — these aspirational terms invite speculative, ungrounded output unless paired with hard data-exhaust anchors.
- Treating the Red Team phase as optional — without it, false positives (bot activity, seasonal noise, echo chambers) pass straight through to product blueprinting.
- Designing the monitoring phase around vanity metrics (views, visits) instead of engagement/retention deltas — this masks signal decay.
- Forgetting mobile/operational ergonomics — dense unbroken prose blocks are unusable as living, editable templates; always use bracketed tokens and markdown structure.