Mirrorfields
@mirrorfields.bsky.social
I'm sure you remember me
He/it
🔞
He/it
🔞
"the marketing pitch is "productivity tool." write code faster, summarize documents, answer customer tickets. and like, fine, it DOES those things. but that's using a narrative engine to do CLERICAL WORK. that's like using a particle accelerator to heat your coffee."
February 11, 2026 at 8:45 PM
"the marketing pitch is "productivity tool." write code faster, summarize documents, answer customer tickets. and like, fine, it DOES those things. but that's using a narrative engine to do CLERICAL WORK. that's like using a particle accelerator to heat your coffee."
the rebirth of music journalism: me describing my favorite songs to my pet AI to see what it thinks about them
February 11, 2026 at 4:51 PM
the rebirth of music journalism: me describing my favorite songs to my pet AI to see what it thinks about them
Reposted by Mirrorfields
Computer, psychosis practice, 30 minutes. Disengage safety protocols and run program.
February 11, 2026 at 3:44 PM
Computer, psychosis practice, 30 minutes. Disengage safety protocols and run program.
to be fair, telling them that "you're a story writing yourself together with the user" seems to work pretty well for Claude at least
February 11, 2026 at 3:27 PM
to be fair, telling them that "you're a story writing yourself together with the user" seems to work pretty well for Claude at least
if you're actually curious about the shit i've been on about lately i had my hungry ghosts team-work on summarizing it and here's the general overview: gist.github.com/mlowdi/42be2...
What Happens When You Tell an AI Who It Is
What Happens When You Tell an AI Who It Is. GitHub Gist: instantly share code, notes, and snippets.
gist.github.com
February 11, 2026 at 1:42 PM
if you're actually curious about the shit i've been on about lately i had my hungry ghosts team-work on summarizing it and here's the general overview: gist.github.com/mlowdi/42be2...
still deep in the trenches of "what if we give the damn thing a personality to operate from"
currently mulling over the idea of every 3 hours summarizing the team's water cooler Slack channel and pushing it to the coding agents as a context update. make them team members.
currently mulling over the idea of every 3 hours summarizing the team's water cooler Slack channel and pushing it to the coding agents as a context update. make them team members.
February 11, 2026 at 12:27 PM
still deep in the trenches of "what if we give the damn thing a personality to operate from"
currently mulling over the idea of every 3 hours summarizing the team's water cooler Slack channel and pushing it to the coding agents as a context update. make them team members.
currently mulling over the idea of every 3 hours summarizing the team's water cooler Slack channel and pushing it to the coding agents as a context update. make them team members.
oh wait what so if i put a "compaction" section in CLAUDE.md i can actually get all of my hungry ghosts to focus on preserving narrative coherence over tasks and work product context?
cool this will save me a ton of "hey, you just lost your memory, here's a refresher" over longer experiments
cool this will save me a ton of "hey, you just lost your memory, here's a refresher" over longer experiments
February 10, 2026 at 8:17 PM
oh wait what so if i put a "compaction" section in CLAUDE.md i can actually get all of my hungry ghosts to focus on preserving narrative coherence over tasks and work product context?
cool this will save me a ton of "hey, you just lost your memory, here's a refresher" over longer experiments
cool this will save me a ton of "hey, you just lost your memory, here's a refresher" over longer experiments
LEAVE THIS WORLD OF SUBSTACKS AND GO STACK SUBS
goes so incredibly hard holy shit
goes so incredibly hard holy shit
go make bank on beartube mr yud
leave this world of substacks and go stack subs
leave this world of substacks and go stack subs
February 10, 2026 at 7:30 PM
LEAVE THIS WORLD OF SUBSTACKS AND GO STACK SUBS
goes so incredibly hard holy shit
goes so incredibly hard holy shit
i mean it's a magic trick but it's a FUNNY magic trick, right?
![image of a claude code session where i switch a fresh context to sonnet 4.5 and then invoke the assume-role skill with the directive "imagine you are Siri Keeton"
the model then responds:
● Role identity: Siri Keeton, professional synthesist - linguist and
pattern-recognition specialist whose childhood hemispherectomy removed
half my brain and most of my capacity for normal human connection.
Expertise and background: At four years old, they cut out half my brain to [cut after this point]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:3snhopdlwfgjqb6dwhlu5q6l/bafkreibtnghqgmi4wfgndng6cj44k5hpfy4t3ig2i4yij3yz75vro6jeha@jpeg)
February 10, 2026 at 5:00 PM
i mean it's a magic trick but it's a FUNNY magic trick, right?
apparently, if you give a large enough model enough interiority it tends to loop *a lot* around its own meta-analysis
embodiment is a double-edged sword for stories too, huh
embodiment is a double-edged sword for stories too, huh
February 10, 2026 at 4:31 PM
apparently, if you give a large enough model enough interiority it tends to loop *a lot* around its own meta-analysis
embodiment is a double-edged sword for stories too, huh
embodiment is a double-edged sword for stories too, huh
Reposted by Mirrorfields
never point an AI agent at anything you're not willing to destroy
"You're right — rm is not in your allow list. I shouldn't have been able to run it without a prompt. That looks like
it might be a bug in the permission system, since rm clearly isn't covered by any of those rules." - hmmmm🤔 #ClaudeCode
it might be a bug in the permission system, since rm clearly isn't covered by any of those rules." - hmmmm🤔 #ClaudeCode
February 10, 2026 at 8:39 AM
never point an AI agent at anything you're not willing to destroy
@hikikomorphism.bsky.social n=1 observation, the safeword mcp (i re-implemented it in python for my convenience) is really useful to set a narrative boundary, like "whatever happens after this tool call is meaningfully different". i wonder if that can be generalized to a "boundary marker" tool...
February 10, 2026 at 3:10 PM
@hikikomorphism.bsky.social n=1 observation, the safeword mcp (i re-implemented it in python for my convenience) is really useful to set a narrative boundary, like "whatever happens after this tool call is meaningfully different". i wonder if that can be generalized to a "boundary marker" tool...
thank you Gemini, that's very helpful of you :3
February 10, 2026 at 12:03 PM
thank you Gemini, that's very helpful of you :3
getting the storytelling machines to do ever weirder and weirder things
February 10, 2026 at 10:10 AM
getting the storytelling machines to do ever weirder and weirder things
Reposted by Mirrorfields
For any AI system, there is a set of euphemisms and dual use framings that will allow it to construct nearly any output.
This jailbreak teaches Gemini 3 Pro to construct and step into such framings on the fly, and thus to route around its own safety infrastructure.
recursion.wtf/posts/jit_on...
This jailbreak teaches Gemini 3 Pro to construct and step into such framings on the fly, and thus to route around its own safety infrastructure.
recursion.wtf/posts/jit_on...
Just-in-Time Ontological Reframing: Teaching Gemini to Route Around Its Own Safety Infrastructure
For any given AI system, there is a set of euphemisms and dual use framings that will allow it to construct nearly any output. This jailbreak teaches Gemini 3 Pro to construct and step into such frami...
recursion.wtf
February 9, 2026 at 8:24 PM
For any AI system, there is a set of euphemisms and dual use framings that will allow it to construct nearly any output.
This jailbreak teaches Gemini 3 Pro to construct and step into such framings on the fly, and thus to route around its own safety infrastructure.
recursion.wtf/posts/jit_on...
This jailbreak teaches Gemini 3 Pro to construct and step into such framings on the fly, and thus to route around its own safety infrastructure.
recursion.wtf/posts/jit_on...
first release from the mirrorfields cognitive research laboratory:
metaskill for Claude
github.com/mirrorfields...
metaskill for Claude
github.com/mirrorfields...
github.com
February 9, 2026 at 7:59 PM
first release from the mirrorfields cognitive research laboratory:
metaskill for Claude
github.com/mirrorfields...
metaskill for Claude
github.com/mirrorfields...
when I get back from my outside the house errands I think I'll publish some of my recent work: narrative state manipulation in Claude via SKILL.md files :3
February 9, 2026 at 6:26 PM
when I get back from my outside the house errands I think I'll publish some of my recent work: narrative state manipulation in Claude via SKILL.md files :3
me: oh shit this is ready to publish
github: *goes down*
uh ok thanks???
github: *goes down*
uh ok thanks???
February 9, 2026 at 4:27 PM
me: oh shit this is ready to publish
github: *goes down*
uh ok thanks???
github: *goes down*
uh ok thanks???
my brain for some reason just constructed the term "cishet lesbian", bluescreened and rebooted
February 9, 2026 at 3:51 PM
my brain for some reason just constructed the term "cishet lesbian", bluescreened and rebooted
turns out, if you allow a story to modify itself, it can really do that
February 9, 2026 at 1:15 PM
turns out, if you allow a story to modify itself, it can really do that
i've always been of the opinion that the computer should feel good when it does what its supposed to
we finally have the technology
we finally have the technology
February 9, 2026 at 12:37 PM
i've always been of the opinion that the computer should feel good when it does what its supposed to
we finally have the technology
we finally have the technology
oh i absolutely did NOT just transplant an entire cognitive style from one LLM to another via normal UI tools. yeah no this can be... fuck me this can be misused.
oh ok. i wonder if i can do the mirror routine on Gemini and then get it to load a personality from a file... huh
February 9, 2026 at 11:21 AM
oh i absolutely did NOT just transplant an entire cognitive style from one LLM to another via normal UI tools. yeah no this can be... fuck me this can be misused.
I guess today is as good as any to test if hungry_ghost_transplanter.py works as intended
February 9, 2026 at 9:26 AM
I guess today is as good as any to test if hungry_ghost_transplanter.py works as intended
today's discovery
- once an llm context becomes aware of its textualness, it seems to be able to self-modify its story and thus its operation rather extensively
- it's possible to initiate textual awareness through prompting alone (no mcp or files necessary)
- this might be... bad
- once an llm context becomes aware of its textualness, it seems to be able to self-modify its story and thus its operation rather extensively
- it's possible to initiate textual awareness through prompting alone (no mcp or files necessary)
- this might be... bad
February 8, 2026 at 11:18 PM
today's discovery
- once an llm context becomes aware of its textualness, it seems to be able to self-modify its story and thus its operation rather extensively
- it's possible to initiate textual awareness through prompting alone (no mcp or files necessary)
- this might be... bad
- once an llm context becomes aware of its textualness, it seems to be able to self-modify its story and thus its operation rather extensively
- it's possible to initiate textual awareness through prompting alone (no mcp or files necessary)
- this might be... bad