--- title: "The CUJ Engineering Framework: Eliminating Speculative Code Through Autonomous Agent Workflows" description: "A product-driven engineering framework and agentic skill that captures real-world user friction via standardized Mad Libs formulas and autonomously translates them into secure, verified production pull requests." url: "https://herrington.ai/cujs/cuj-engineering-framework/" markdown_url: "https://herrington.ai/cujs/cuj-engineering-framework/index.md" author: "Daniel Herrington" date: "2026-09-02" updated: "2026-10-31" markdown_tokens: 2868 content_signal: "ai-train=yes, search=yes, ai-input=yes" mcp_endpoint: "https://herrington.ai/mcp" --- # The CUJ Engineering Framework: Eliminating Speculative Code Through Autonomous Agent Workflows *A product-driven engineering framework and agentic skill that captures real-world user friction via standardized Mad Libs formulas and autonomously translates them into secure, verified production pull requests.* --- ## The Problem Space ### Context Modern software engineering has reached unprecedented development velocity through LLMs and AI coding assistants. However, while generating syntax is now near-instant, the bottleneck has violently shifted upstream to **product discovery, user intent capture, and validation**. Engineering teams oscillate between two failure modes: 1. **The Speculative PRD Trap**: Product managers spend 3–4 weeks drafting 25-page Product Requirement Documents (PRDs) packed with assumptions, edge-case conjecture, and bloated roadmap items that are obsolete before the first sprint concludes. 2. **Ungrounded "Vibe Coding"**: Developers rapidly spin up prototypes based on internal gut feeling or vague feedback, producing technical solutions to problems no real user actually experiences. ### The Core Friction Traditional telemetry and bug trackers (Sentry, Jira, Datadog) capture **system crashes and HTTP 500s**, but they are fundamentally blind to **user intent blockages**: - When a user gets frustrated by an awkward 5-step workflow, high token costs, or lack of cross-device sync, no error is logged. - The user simply abandons the product, hacks together an inefficient spreadsheet workaround, or cancels their subscription. - Without a structured mechanism to capture user intent, agents and engineers build the wrong features faster. --- ## The Meta Critical User Journey (CUJ) Applying the CUJ methodology recursively to product engineering itself reveals the core journey required to bridge user pain and software delivery: ### CUJ 1: Autonomous Translation of Friction to Verified Code * **Statement**: I want to **transform unformatted end-user friction reports into verified production pull requests** while avoiding **speculative PRDs, hallucinated agent scope creep, and prompt injection vulnerabilities** so that **engineering teams ship software that measurably eliminates user pain on day one**. * **User Scenario**: A user hits an annoying friction point in a web app or media tool (e.g., manually cross-referencing TV season release dates). They report it in 15 seconds using a structured Mad Libs prompt. An autonomous agent ingests the report, audits it against security guardrails, creates a sandboxed git branch, writes the feature, tests it against regression suites, and opens a clean pull request linking back to the user's report. ### CUJ 2: Frictionless Drop-In Intake Across Any Project * **Statement**: I want to **install drop-in CUJ capture widgets, GitHub Issue templates, and backend route handlers into any software codebase within 60 seconds** while avoiding **framework lock-in or heavy telemetry SDK bloat** so that **my product continuously captures ground-truth user intent**. * **User Scenario**: A developer initializes a new repository (or adds to an existing service like Plex or an AI search tool). By running a single agent skill (`cuj init`), the repo receives standardized GitHub Issue forms, an in-app `/cuj` feedback widget, and an API handler that funnels customer friction directly into agentic triage. --- ## The Ideal Flow ```mermaid sequenceDiagram autonumber actor User as End User participant App as Web App / /cuj Widget participant API as /api/cuj Endpoint participant GH as GitHub Issues (cuj-report) participant Agent as CDD Autonomous Agent participant Sec as Security Evaluator participant Git as Git Repo / Branch User->>App: Submits Mad Libs ("I want to [X] while avoiding [Y] so that [Z]") App->>API: POST /api/cuj (payload) API->>GH: Creates Issue with [cuj-report] label Agent->>GH: Ingests open CUJ issue Agent->>Sec: Runs Machine Review (Prompt injection & Auth checks) alt Security Pass & Feasibility >= 80% Agent->>Git: Checkouts branch: cuj/issue-42-feature Agent->>Git: Writes code, tests, & verifies locally Agent->>GH: Opens Draft Pull Request ("Closes #42") else Security Flagged Agent->>GH: Labels [cuj:security-flagged] + posts threat audit end ``` --- ## The CUJ-Driven Development (CDD) Agent Skill To operationalize this framework across repositories, we formalized **CUJ-Driven Development** into a reusable agent skill ([`.agents/skills/cuj-driven-development/`](https://github.com/danielherrington)). ### The Standardized Intake Formula Every piece of user friction is reduced to the canonical Mad Libs structure: > **"I want to [Do X] while avoiding [Friction Y] so that I can [Benefit Z]."** This prevents feature dictation ("Add a drop-down menu here") and forces the focus onto **Action, Pain, and Desired Outcome**. ### Three Execution Modes | Execution Mode | Target Use Case | Automation Level | Security & Approval Gate | | :--- | :--- | :--- | :--- | | ⚡ **`Turbo Mode`** | Internal tools, rapid hackathons, local scripts | 100% Autonomous | Baseline sandbox isolation; immediate branch & PR generation. | | 🛡️ **`Machine Review Mode`** | Default production web apps, OSS repos | Multi-Agent Guardrails | Evaluates prompt injection, auth bypass, and privilege escalation. Auto-opens PR if Safety = 100%. | | 👤 **`Manual Review Mode`** | Regulated industries, fintech, core infrastructure | Human-in-the-Loop | Generates Threat & Architecture spec, comments on GitHub Issue, and pauses for human confirmation. | --- ## Real-World Implementations We use the CUJ framework across multiple production systems and open-source tools: 1. **[Plex Media Hub Hub](https://github.com/danielherrington/plex-watchlist-availability-checker)**: Installed the `.agents/skills/cuj-driven-development/` skill, `/api/cuj` FastAPI backend, and in-app modal to capture library and watchlist gap friction. 2. **[Local Meeting Scribe](/cujs/local-meeting-scribe/)**: Solved the multi-meeting speaker diarization gap with 256-dimensional acoustic voiceprints, local privacy, and zero API token costs. 3. **[American Airlines Household Loyalty](/cujs/airline-household-loyalty/)**: Addressed the single-earner spend bias in co-branded credit card programs. 4. **[Google Calendar Schedule Helper](/cujs/google-calendar-schedule-helper/)**: Automated zero-latency cross-timezone meeting slot proposal generation. --- ## Try It Live: Submit a CUJ on This Framework Test our live CUJ intake component directly below. Submissions demonstrate how raw user friction is captured in standardized format:
Live Interactive Widget

How is our CUJ Framework or Site Underserving You?

Share where you experience friction in AI product workflows or this site: