Cywareness Simulation Archive
Production security-awareness content at volume, and the pipeline that now builds it.
Hand-authoring calibrated sims across brands, languages, and tiers stopped scaling somewhere in the hundreds. The pipeline's split matters for maintenance: the deterministic Python packagers produce byte-identical output run to run, while the model-driven agents are the only part that improves when a better model ships. One named seam, one place to upgrade. The original work belongs to the employer, so what is public is a sanitized rebuild of the pipeline: the seven skills, the eight-step orchestrator, and the twelve sub-agents, with none of the client templates, brands, or captured pages. It is linked below.
Volume was the easy part. The output had to upload into a platform that expects exact structure, stay reviewable by trainers who do not read code, and hold a tier calibration steady enough that Beginner means the same thing in Manila and in Warsaw. The work itself belongs to the employer, so what I can show is the pipeline and the method, not the client files.
A security-awareness platform producing sims across many brands and markets.
The depth behind the sim production line above.
- Skills hold the rulesClaude Code skills encode the indicator rules, tier spreads, and the sim-pipeline contract.
- The orchestrator runs the contractAn eight-step orchestration turns one brief into a full tiered package.
- Agents do the judgmentSub-agents author the tiers and the indicator realism, the part that is genuinely hard.
- Python packages it, deterministicallyIndicator sheets and upload ZIPs are built by Python that does not change run to run.
- Upgradeable at the seamOnly the model-driven agents get better with a better model. The Python packagers stay byte-for-byte identical. Now that Fable 5 exists, the indicator-realism agent is the highest-upside swap.
Where to put the seam between the model and the code. The obvious build is one agent that authors a sim and emits the finished package. I split it instead: sub-agents author, where judgment is the whole job, and Python packages, where judgment is a liability. The payoff arrives on upgrade day. When a better model ships, the packagers are byte-for-byte unchanged and need no re-verification, so the only thing to re-test is the half that was supposed to improve. A single-agent build would have turned every model upgrade into a full regression of the entire output.
Every one is a hand-built vanilla HTML/CSS recreation of a real sign-in or verify page, used as the click target in an authorized training sim. These are screenshots of my rebuilds. Click any to open it full size.
Each is a full HTML email built to a brand's real house style, the message half of a tiered sim. Consumer brands, SaaS tools, and plain internal-comms pretexts. Click any to open it full size.
