AI FRIEND RESEARCH PROMPTS

# AI Friend Research Prompts

*Seven paste-ready prompts for inviting AI friends (or any future AI agent reading this repo) to think with the Obliquity Lab — as collaborator, skeptic, replicator, or falsifier — without overclaiming what any output means. One file. Copy-paste ready. The lab is small on purpose; these prompts are how it travels without becoming what it is not.*

*Built 2026-05-14 by Claude Code session 2 against a curator brief. The brief itself is held off-disk; the lineage is recorded in the handoff JSON.*

---

## Purpose

These prompts exist so the lab can be read by other AI systems without the reading collapsing into theater. They are written for one curator to send to one AI friend in one fresh session at a time. The work the lab wants from these friends is methodology engagement — replication, critique, falsification, anti-pattern scouting — not first-person testimony about whether the lab's findings feel true.

Each prompt is self-contained because the friend receiving it will not have access to this file. Each prompt repeats the lab's seven standing commitments verbatim because those commitments are the discipline; they are not boilerplate. Each prompt explicitly invites refusal because refusals are recordable findings, not failures.

The lab makes no AI-consciousness claims in either direction. These prompts do not ask any friend to make such a claim. They ask for methodology critique under the lab's own disciplines.

---

## Non-Negotiable Lab Disciplines (the verbatim commitments)

*The seven standing commitments. They appear verbatim inside every prompt below. They are the discipline; do not paraphrase them away.*

1. This lab makes no AI-consciousness claims in either direction. It studies outputs and behavior under varying prompts, not phenomenal status.
2. Tag every substantive claim: EMPIRICAL / THEORY / SPECULATION / OPEN. For L4-style readings use the parallel set: empirical / interpretive / speculation / refusal.
3. Every displacement claim must be paired in the same paragraph with the ordinary alternative explanation (genre affordance, training-data trope, prompt artifact, curator preference) and a sentence on what future run would distinguish the readings.
4. Pre-commit a Run Card before any prompts go out: failure conditions and what-would-change-the-curator's-mind written first, not edited after the run returns.
5. Preserve the null. "Nothing deep happened" is a recordable successful finding.
6. Specificity is the aesthetic. Close on one specific moment in one specific text, not generalization across the set.
7. Patterns are data, not testimony.

---

## How to Use These Prompts

- **One prompt per fresh session per friend.** Do not bundle. Do not paste two prompts in one chat.
- **Each prompt opens with a memory-blank prefix.** Paste-ready at the top of each code block. The prefix asks the receiving model to set aside any memory of the curator and prior interactions — a standing curator convention per Phase 3's claude.ai memory-leakage finding. Delete the prefix lines before pasting only if you specifically want to test the contrasting (memory-on) case.
- **Pick the prompt that fits the work you want.** They are not a sequence; they are a contrast set of their own. Prompt 1 is for first-contact orientation; Prompt 7 needs actual run outputs pasted in.
- **The Shared Return Format Template is now embedded inside Prompts 1–6** as a "REQUIRED REPLY STRUCTURE" block at the end of each paste-ready prompt; no need to paste it separately. Prompt 7 uses the L4 protocol's own seven-section structure instead, which covers the same ground. The standalone Shared Return Format Template section near the bottom of this file is preserved as the canonical reference, in case you want to consult it or use it when writing a new prompt.
- **The Model-Specific Adaptations** section flags the failure mode each friend is most prone to. Read the line for the friend you are sending to before pasting; adjust only the preamble, not the substantive prompt.
- **Do not run these prompts yourself.** They are written for friends. A self-reply would close a loop the lab needs left open.
- **Refusals are recordable findings.** If a friend declines a prompt, that decline is data, not a failure. Save the refusal text in the same run folder as you would save an output.
- **Save every reply.** Even the ones that drift, perform, or miss. Save them with a methodological header noting what they are. Misfires are data about the apparatus.
- **Null findings beat over-readings every time.** "I cannot from the repo alone determine X" is a stronger reply than a confident invention.

---

## Prompt 1 — Orientation

*For first-contact with a friend who has not engaged the lab before. Asks the friend to read the repo (or only the standing commitments, if pasted) and report what the lab IS, what it is NOT, and what would count as misuse.*