From Prompt Slop to Skill Engineering: A Practitioner's Playbook

The shift. Paul Bakaus argues that effective AI-assisted development has moved beyond prompt engineering into what he calls 'skill engineering' — treating skills as harness extensions rather than packaged prompts. His Impeccable project, originally a personal design skill for Cursor and Claude Code, now spans over a hundred files, scripts, hooks, and model-specific builds. The core claim: prompting alone cannot overcome a model's gravitational pull toward median outputs, no matter how many lines of instruction you write.

Why bans fail. Bakaus demonstrates that banning undesirable outputs — like purple gradients or system fonts — merely relocates the model to the next cluster in its latent space. 'Slop is a moving target,' he notes. Instead, he introduces 'anti-attractors': mechanisms that force divergence by introducing random seeds from scripts, user input, or creative prompts like 'what would Beyoncé look like as a shader?' The technique prevents the convergence that naive prompting produces.

Sub-agent architecture. The most practically useful technique involves spawning multiple sub-agents that argue blindly. For critique and review, the model evaluating its own work anchors on its prior output. Impeccable solves this by having one sub-agent act as a design director (heuristic analysis) and another run a deterministic detector, then synthesizing both outputs. Two blind opinions beat one confident guess — applicable to code review, security audits, and plan critiques.

Harness differences matter. Codex, Claude Code, Cursor, and Copilot have different permission models, tool availability, and model behaviors. Sub-agent spawning works differently across them. Background jobs, user-question tools, and hooks all vary. Bakaus ships five separate builds of Impeccable (Claude Code, Cursor, Codex, Copilot, Gemini) with model-specific rules. The investment is substantial but necessary for cross-platform reliability.

The script hook technique. A standout insight: instructions delivered via script standard output are followed far more reliably than rules buried in system prompts. By calling a context.mjs script that produces structured JSON with precise next steps, the model obeys with higher fidelity. The trade-off is that dynamic script output breaks prompt caching — acceptable for interactive sessions but suboptimal for high-volume batch use.

The evaluation gap. Bakaus built a custom evaluation harness that runs end-to-end tests across five models, using an LLM as a simulated user for interactive flows. He does ablation testing on individual lines (each tagged with a unique ID) to measure impact. But he admits taste evaluation remains unsolved — models are maximalists, and inverting their aesthetic judgments sometimes works better than trusting them.