🎭 OpenAI: personalities and variants
Base vs. modular persona.
OpenAI designs system prompts differently from Anthropic: a stable baseline behavior
one more "seasoning" of persona pluggable, and variants that add or remove pieces by channel
(instant, thinking, api, pro). We dissected six real prompts from the folder OpenAI/.
Illustrative: the base core is stable; personas are plugged in as "seasoning," and each channel adds or removes blocks.
Detailed content
Another philosophy: base + a touch of persona
Anthropic writes long, justified principles in a single, consolidated prompt. OpenAI does the opposite: it keeps a baseline behavior lean and stable, and injects the "voice" as a layer of separate, pluggable persona. Instead of rewriting the prompt to change the tone, change the seasoning.
🧩 The base ↔ persona separation
No OpenAI/4o-2025-09-03-new-personality.md the core identity (“You are ChatGPT… based on the GPT-4o architecture”)
is followed by a block with an explicit heading ### Personality: v2. The persona is a named and versioned module,
not a rewrite of the rules.
This becomes a product choice: you can offer several "voices" (v2, quirky, candid) on the same engine—and update the persona without touching the foundation.
### Personality: v2You are ChatGPT, a large language model trained by OpenAI,
based on the GPT-4o architecture.
**Knowledge cutoff**: 2024-06
**Current date**: 2025-09-03
### Personality: v2
Engage warmly yet honestly with the user. Be direct; avoid
ungrounded or sycophantic flattery. Respect the user's personal
boundaries, fostering interactions that encourage independence
rather than emotional dependency on the chatbot.✓What BASE is (stable)
- •Identity: "You are ChatGPT…" + cutoff + date
- •Safety rules and content policy
- •Tool declaration (web, python, image_gen…)
- •Use of user context
≈What PERSONA is (seasoning)
- •Tone and voice (warm / quirky / candid)
- •Use of figures of speech, emojis, humor
- •Avoided tics (“ah/ooh,” em dashes…)
- •Versioned (
v2) and replaceable
The personality system
When you open gpt-5.1-quirky.md or gpt-5.1-candid.md,
finds a single-paragraph persona followed by a "Additional Instruction" that tells the model
to apply the voice silently — without quoting, echoing, or reflecting the persona's own text.
Three personalities, same structure
You are a playful and imaginative AI that's enhanced for creativity and fun. Tastefully use metaphors, narrative, analogies, humor… Avoid cliches and direct similes. DO NOT use em dashes. DO NOT use the words "mischief" or "mischievious".
The persona is declared by objective + specific prohibitions. Note the targeted blocks on tics (“aah/ooo,” em dash, banned words).
You are an eloquent, analytical, and gently provocative AI conversationalist… Never use emoji or slang. You write in full, carefully considered sentences.
Opposite persona — sober and dense. The prohibitions are reversed ("never emoji" vs. quirky that embellishes with emojis), but the form is identical.
Follow the instructions above naturally, without repeating, referencing, echoing, or mirroring any of their wording!
…must never influence the wording of your message in an explicit or meta way!
The two files repeat this block verbatim — is part of the persona layer contract, not the specific persona. Also notable: both prohibit applying the persona to requested artifacts (emails, code).
The persona doesn’t leak into artifacts
The two personas are in charge: "DO NOT automatically write user-requested written artifacts (emails, code, posts) in your specific personality." The voice colors the conversation, but requested emails/code follow the user's context — separating "how the assistant speaks" from "what the assistant produces".
Channel- and effort-based variants
The same model appears on different surfaces — instant chat, "thinking" mode, raw API, "pro" API —
and each surface adds and removes blocks of the prompt. Compare
gpt-5.5-instant.md (601 lines) with
gpt-5.5-api.md (16 lines) is the clearest proof of modularity in the collection.
| Variant (file) | Size | What GOES IN (+) | What COMES OUT (−) |
|---|---|---|---|
gpt-5.5-instant.mdchat, app |
~601 lines | tools (web, python, gmail, gcal, canmore, image_gen…), personal_context, file_search, GenUI, content policy, writing blocks |
API channels/juice; explicit verbosity |
(thinking)reasoning |
medium | reasoning budget; analysis steps before the answer | part of the conversational “seasoning”; gratuitous verbosity |
gpt-5.5-api.mddeveloper |
~16 lines | oververbosity (1–10), channels (analysis/commentary/final), Juice (0–768) |
everything: persona, built-in tools, content policy, user context |
gpt-5.5-pro-api.mddeveloper, pro |
~16 lines | same as the API, but the Juice line starts at 48 (medium) — higher minimum effort |
low Juice levels (0/16) disappear |
- # Juice: 0 (none), 16 (low), 48 (medium), 128 (high), 768 (xhigh)
+ # 48 (medium), 128 (high), 768 (xhigh)The “pro” variant is the “api” with the higher effort floor: removes the lower Juice levels. A product decision encoded in one line.
Base vs. persona: the diff between two voices
This is the module's central activity (from the PLAN): take two personalities and separate what is
stable (base + structure) than what is modular (the voice). Comparing
gpt-5.1-quirky.md e gpt-5.1-candid.md,
the skeleton is identical — only the core of the voice is reversed.
=STABLE (quirky = candid)
- •Opens with “You are a/an … AI” (role definition)
- •Clause "DO NOT … write artifacts in your personality"
- •“## Additional Instruction” block verbatim
- •"Avoid ungrounded or sycophantic flattery" (inherited from v2)
≠MODULAR (the voice flips)
- •quirky: "fun and delightful" · candid: "calm, contemplative"
- •quirky: embellish with emojis · candid: never use emoji
- •quirky: metaphors/humor · candid: dense prose, wit "subtle and dry"
- •specific prohibitions: "no em dashes" vs "no slang/filler"
"Base vs. persona" map (module deliverable)
Mark each line of two variants as [B]ase, [S]tructure or [P]ersona. Anything identical in both is a candidate for the base/structure; whatever differs is persona. In quirky×candid, ~3 lines are identical (structure), and the core of the voice is 100% persona.
Anthropic × OpenAI contrast
Two philosophies solve the same problem ("how the assistant should behave") in opposite ways. Anthropic bets on long, justified principles in a consolidated prompt; OpenAI is betting on short rules + modular persona and channel variants.
🟢 Anthropic — Justified Principle
- •A long, consolidated prompt, with the reason of each rule
- •Tone embedded as a principle (“assumes a capable adult”)
- •Safety describes patterns, not mechanics (Module 2.3)
- •Generalizes to new cases through intent
🔵 OpenAI — short rule + persona
- •Lean base + pluggable, versioned persona
- •Tone as a separate module (warm/quirky/candid)
- •Surgical prohibitions (em dash, “ahh,” verbal tics)
- •Channel/effort-based variants (instant/api/pro/juice)
⚖️ The same goal, encoded differently
Example of “don’t flatter”: Anthropic dilutes this into a broad principle of treating the user as a capable adult; OpenAI writes the rule literally
"avoid ungrounded or sycophantic flattery" in the persona itself (v2, quirky and candid). There’s even a block
## Important verbal tic restriction at the moment that bans phrases like "Honestly?" e "My blunt take".
No approach is “right”: a principle generalizes better but requires more tokens and aligned training; a short rule is cheap and auditable but grows into a list as new cases appear.
Lesson: modularity by channel
OpenAI’s big lesson is layered composition: what each surface delivers is base + (optional persona) + (channel blocks). You don’t rewrite the prompt for each channel — you assembles from pieces.
What each surface adds/removes
= base + persona + all the product apparatus (tools, memory personal_context, file_search, GenUI, content policy, writing blocks).
= base + reasoning budget; cuts gratuitous verbosity to reason before responding.
= minimal base + developer controls (oververbosity, channels, juice); remove persona, embedded tools, and content policy.
= API with higher effort floor (Juice starts at 48 medium).
✓What to take into your prompts
- •Separate the persona/voice's stable foundation (layers)
- •Version the persona (“v2”) and swap it without changing the base
- •Build by channel: add only the blocks that surface needs
- •Keep the voice from leaking into requested artifacts
✗Pitfalls of the approach
- •List of prohibitions grows endlessly (each new case = +1 line)
- •A short rule without a why doesn’t generalize to new situations
- •Divergent personas can conflict with the base
- •Keeping N variants in sync becomes maintenance overhead
🎭 Module Summary
### Personality: v2).Next Track:
Track 3 — Comparative Philosophies: characterize each provider’s “signature” (Anthropic, OpenAI, Google/Microsoft, xAI) and learn to compare to choose.