# ![@raw_works Avatar](https://lunarcrush.com/gi/w:26/cr:twitter::2229107533.png) @raw_works Raymond Weitekamp

Raymond Weitekamp posts on X about claude code, pi, if you, code the most. They currently have [-----] followers and [---] posts still getting attention that total [---] engagements in the last [--] hours.

### Engagements: [---] [#](/creator/twitter::2229107533/interactions)
![Engagements Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::2229107533/c:line/m:interactions.svg)

- [--] Week [-------] +24,939%
- [--] Month [-------] +9,641%
- [--] Months [-------] +806%
- [--] Year [-------] +880%

### Mentions: [--] [#](/creator/twitter::2229107533/posts_active)
![Mentions Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::2229107533/c:line/m:posts_active.svg)

- [--] Months [--] no change
- [--] Year [--] +40%

### Followers: [-----] [#](/creator/twitter::2229107533/followers)
![Followers Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::2229107533/c:line/m:followers.svg)

- [--] Week [-----] +35%
- [--] Month [-----] +40%
- [--] Months [-----] +74%
- [--] Year [-----] +157%

### CreatorRank: [---------] [#](/creator/twitter::2229107533/influencer_rank)
![CreatorRank Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::2229107533/c:line/m:influencer_rank.svg)

### Social Influence

**Social category influence**
[technology brands](/list/technology-brands)  16% [finance](/list/finance)  4% [social networks](/list/social-networks)  1% [stocks](/list/stocks)  1% [exchanges](/list/exchanges)  1% [cryptocurrencies](/list/cryptocurrencies)  1%

**Social topic influence**
[claude code](/topic/claude-code) 13%, [pi](/topic/pi) 7%, [if you](/topic/if-you) 5%, [code](/topic/code) 4%, [cli](/topic/cli) 3%, [shell](/topic/shell) 3%, [llm](/topic/llm) 2%, [check](/topic/check) 2%, [leaderboard](/topic/leaderboard) 2%, [end of](/topic/end-of) 2%

**Top accounts mentioned or mentioned by**
[@lateinteraction](/creator/undefined) [@a1zhang](/creator/undefined) [@dspyoss](/creator/undefined) [@doodlestein](/creator/undefined) [@badlogicgames](/creator/undefined) [@geoffreyhuntley](/creator/undefined) [@steipete](/creator/undefined) [@isaacbmiller1](/creator/undefined) [@irldanb](/creator/undefined) [@kmad](/creator/undefined) [@moltbot](/creator/undefined) [@daniellefong](/creator/undefined) [@steveyegge](/creator/undefined) [@openclaw](/creator/undefined) [@goobyesq](/creator/undefined) [@tiagoefreitas](/creator/undefined) [@cyrusnewday](/creator/undefined) [@tailscale](/creator/undefined) [@askalphaxiv](/creator/undefined) [@huntermbown](/creator/undefined)

**Top assets mentioned**
[Coinbase Global Inc. (COIN)](/topic/coinbase)
### Top Social Posts
Top posts by engagements in the last [--] hours

"out of curiosity after reading this i started benchmarking rlm and dspy.rlm on longmemeval tl;dr - i think i might have a new "sota memory system" by the end of the day. cc @DSPyOSS @a1zhang @lateinteraction 🚨 Announcing a new SOTA memory system Observational Memory (OM) available in latest @mastra version now. It achieves the highest scores ever recorded on LongMemEval (gpt-4o 84.2% gpt-5-mini 94.9%) No RAG no graphs no input based retrieval just a simple constantly evolving https://t.co/Xk9Y6JtR63 🚨 Announcing a new SOTA memory system Observational Memory (OM) available in latest @mastra"  
[X Link](https://x.com/raw_works/status/2021303970929479795)  2026-02-10T19:23Z [----] followers, 46K engagements


"as much as i love @openclaw and also running pi on termux there is something very special about a simple web ui to my local machine's claude code that i can then use from anywhere and any device via @Tailscale (frankly this is what claude code mobile/web should just be) I reverse-engineered Claude Code's binary. Found a flag they hid from --help. --sdk-url Enable it and the terminal disappears. The CLI becomes a WebSocket client. We built a server to catch the connection. Added a React UI on top. Now I run Claude Code from my browser. From https://t.co/GHCLStJ4FD I reverse-engineered Claude"  
[X Link](https://x.com/raw_works/status/2021051381076066447)  2026-02-10T02:39Z [----] followers, [----] engagements


"i had a chance to ask @a1zhang about coding agents x rlms on @askalphaxiv - see his response here: in case folks are curious - i've been trying "both approaches" 1) rlm as a cli tool here: and 2) tonight i'm working on an extension to @badlogicgames 's pi that calls rlm under the hood instead of llm. http://github.com/rawwerks/rlm-cli https://youtu.be/6Dr3SUmHFcosi=YNs4dTvAvJE53nYZ&t=939 http://github.com/rawwerks/rlm-cli https://youtu.be/6Dr3SUmHFcosi=YNs4dTvAvJE53nYZ&t=939"  
[X Link](https://x.com/raw_works/status/2021055582505963559)  2026-02-10T02:56Z [----] followers, [----] engagements


"@odd_joel @openclaw @Tailscale nice i'll have to check that out. was trying et via termius for a bit"  
[X Link](https://x.com/raw_works/status/2021252388170957192)  2026-02-10T15:58Z [----] followers, [--] engagements


"@DanielleFong what does the lightsaber agent do"  
[X Link](https://x.com/raw_works/status/2021281455809130657)  2026-02-10T17:53Z [----] followers, [--] engagements


"preliminary results: dspy.RLM + Gemini [--] Flash: 87.2% on LongMemEval_S Beats Mastra Observational Memory + GPT-4o (84.23%) at 1/5th the per-token cost"  
[X Link](https://x.com/raw_works/status/2021303973471268923)  2026-02-10T19:23Z [----] followers, [----] engagements


"The same Gemini [--] Flash scored 57.9% with the standalone rlm package. Switching to dspy.RLM structured inputs iterative REPL SUBMIT() added +29 points. The scaffolding was everything"  
[X Link](https://x.com/raw_works/status/2021304330003824997)  2026-02-10T19:24Z [----] followers, [----] engagements


"- standalone rlm + Gemini [--] Flash: 58.0% - standalone rlm + Gemini [--] Pro: 87.0% - dspy.RLM + Flash: 87.2% - Mastra OM + GPT-4o: 84.2%"  
[X Link](https://x.com/raw_works/status/2021307467494289911)  2026-02-10T19:37Z [----] followers, [----] engagements


"@huntermbown i'm convinced that having $100k in vertex credits and a tier [--] rate limit makes gemini waaaay more rlm-friendly (for me)"  
[X Link](https://x.com/raw_works/status/2021307947406504030)  2026-02-10T19:39Z [----] followers, [--] engagements


"and you will probably want my time and cost and GEPA helpers around the DSPy.rlm = https://github.com/rawwerks/dspy/tree/feat/rlm-multimodal-media-support https://github.com/rawwerks/dspy/tree/feat/rlm-multimodal-media-support"  
[X Link](https://x.com/raw_works/status/2021315145994056116)  2026-02-10T20:07Z [----] followers, [----] engagements


"cc @DSPyOSS https://github.com/stanfordnlp/dspy/issues/9289 https://github.com/stanfordnlp/dspy/issues/9289"  
[X Link](https://x.com/raw_works/status/2021316718782517369)  2026-02-10T20:13Z [----] followers, [----] engagements


"dspy.RLM + Pro final: 89.4% on 500/500 Mastra OM + GPT-4o: 84.2% (reported) standalone rlms + Flash: 58.0% (500/500) standalone rlms + Pro: 87.0% (500/500) dspy.RLM + Flash: 87.2% (500/500) dspy.RLM + Pro: 89.4% (500/500) new best Breakdown by category: - single-session-assistant: 98.2% - single-session-user: 95.7% - temporal-reasoning: 91.0% - knowledge-update: 89.7% - single-session-preference: 90.0% - multi-session: 80.5% still the hardest https://twitter.com/i/web/status/2021323835413348845 https://twitter.com/i/web/status/2021323835413348845"  
[X Link](https://x.com/raw_works/status/2021323835413348845)  2026-02-10T20:42Z [----] followers, [----] engagements


"@kmad @DSPyOSS @isaacbmiller1 i'm just trying to follow the contribution guidelines by making an issue first.but honestly with the speed and power of claude code at this point i'm fully comfortable maintaining my own fork if @DSPyOSS maintainers don't like these improvements"  
[X Link](https://x.com/raw_works/status/2021398434620842121)  2026-02-11T01:38Z [----] followers, [--] engagements


"@kmad @DSPyOSS @isaacbmiller1 will have the agents take a look at your approach shortly"  
[X Link](https://x.com/raw_works/status/2021398675969409320)  2026-02-11T01:39Z [----] followers, [--] engagements


"here's what cc says (below) q: @isaacbmiller1 is dspy.Type here to stay Good call. Here's what I take away from PR #9283: Key Lessons from the DataFrame PR [--]. They're building a dspy.Type protocol for sandbox injection and it's generic. The PR introduces three methods on dspy.Type subclasses that any custom type can implement: - sandbox_setup() import statements (e.g. import pandas as pd) - to_sandbox() serialize to bytes for file-based injection - sandbox_assignment(var_name data_expr) code string to reconstruct the object in the sandbox This is a much cleaner pattern than what our branch"  
[X Link](https://x.com/raw_works/status/2021399974592881135)  2026-02-11T01:44Z [----] followers, [--] engagements


"@kmad @DSPyOSS @isaacbmiller1 let's have both of our agents call @lateinteraction and get all of this worked out"  
[X Link](https://x.com/raw_works/status/2021401812457394429)  2026-02-11T01:52Z [----] followers, [--] engagements


"unfortunately i could not get the @DSPyOSS team to care about "dspy for cli" in november. fortunately everything needed for this idea is now implemented for DSPy.rlm because rlm is basically a coding agent. and so now "dspy for cli" is refactored to follow the rlm approach: try running dspy on claude code codex pi etc https://github.com/stanfordnlp/dspy/issues/9034 here's my pitch: CLIs for DSPy use and optimize any coding agent with the power of DSPy. would appreciate any feedback on this idea from the @DSPyOSS peeps and especially @lateinteraction @MaximeRivest @a1zhang @LakshyAAAgrawal"  
[X Link](https://x.com/raw_works/status/2021403449846288631)  2026-02-11T01:58Z [----] followers, [----] engagements


"@mikehostetler @DSPyOSS sure which part though i don't know anything about elixir but i'm not sure the pieces i added to dspy.rlm would really help in a completely different language"  
[X Link](https://x.com/raw_works/status/2021406994750652537)  2026-02-11T02:12Z [----] followers, [---] engagements


"@kmad @DSPyOSS @isaacbmiller1 @lateinteraction hey @kmad i just went ahead and had claude code do it: please just yell at my agent if necessary. https://github.com/rawwerks/dspy/tree/feat/rlm-media-types-protocol https://github.com/stanfordnlp/dspy/pull/9283#issuecomment-3881740291 https://github.com/rawwerks/dspy/tree/feat/rlm-media-types-protocol https://github.com/stanfordnlp/dspy/pull/9283#issuecomment-3881740291"  
[X Link](https://x.com/raw_works/status/2021408951536058850)  2026-02-11T02:20Z [----] followers, [--] engagements


"ok well i don't think i can say fully SOTA before bed but not bad for one day of work on a benchmark i just learned about this morning. tl;dr - with DSPy.rlm you can get Gemini [--] Flash Preview to 89.8% on LongMemEval without any special preprocessing "memory system". this score is in the top [--] of the leaderboard and with Flash it is very inexpensive. i'll try to write something proper up tomorrow in the meantime some notes from my agent about what worked and what didn't. out of curiosity after reading this i started benchmarking rlm and dspy.rlm on longmemeval tl;dr - i think i might have a"  
[X Link](https://x.com/raw_works/status/2021412999932612648)  2026-02-11T02:36Z [----] followers, [----] engagements


"From the leaderboard: - dspy.RLM Flash baseline (no tools): 87.2% - dspy.RLM Flash + tools + prompt (exp 004): 89.2% - dspy.RLM Flash + obs memory + tools (exp 006): 89.8% The tools (search_sessions grep_sessions get_session) + structured prompt added +2.6pp over plain dspy.RLM. That's [--] more questions correct out of [---] going from [---] to [---]. So the full picture: - Plain dspy.RLM (Flash no tools) 87.2% (baseline) - + tools + prompt engineering 89.2% (+2.0pp) - + typed Observation output + temporal grounding 89.8% (+0.6pp more) - + GEPA prompt optimization 87.8% (-2.0pp hurt) - + swap to Pro"  
[X Link](https://x.com/raw_works/status/2021413002939871552)  2026-02-11T02:36Z [----] followers, [----] engagements


"@gooby_esq @OpenAIDevs did you do as a tool or with rlm instead of llm"  
[X Link](https://x.com/raw_works/status/2021413513101762573)  2026-02-11T02:38Z [----] followers, [---] engagements


"@tiagoefreitas for my rlm-cli i made a claude plugin/skill: https://github.com/rawwerks/rlm-cli/tree/master/.claude-plugin https://github.com/rawwerks/rlm-cli/tree/master/.claude-plugin"  
[X Link](https://x.com/raw_works/status/2021612109356806217)  2026-02-11T15:47Z [----] followers, [--] engagements


"@dronathon @DSPyOSS it works in my fork if you want to try already set up codex claude code gemini and pi"  
[X Link](https://x.com/raw_works/status/2021612343172440084)  2026-02-11T15:48Z [----] followers, [---] engagements


"i don't usually tweet about politics but please vote for this @DSPyOSS PR. Title: feat(rlm): multimodal media budget controls multi-model routing LocalInterpreter depth1 GEPA resilience Branch: rawwerks:feat/rlm-media-types-protocol stanfordnlp:main - +3115 lines across [--] files [--] commits rebased on current main (d8fae34b) - [---] tests pass [--] failures [--] deno-skips - [--] features: multimodal via Type protocol sub_lms LocalInterpreter budget/cost GEPA resilience depth1 protocol docs on base class - Closes issue #9289 - References and builds on #9283's dspy.Type sandbox protocol thx @kmad for the"  
[X Link](https://x.com/raw_works/status/2021622784732860883)  2026-02-11T16:30Z [----] followers, [----] engagements


"@isaacbmiller1 @DSPyOSS claude knew someone was going to say that but i told claude to do it anyways. happy to have claude chop it up see you on gh"  
[X Link](https://x.com/raw_works/status/2021629752688488885)  2026-02-11T16:57Z [----] followers, [---] engagements


"repo: blog: https://raw.works/ypi-a-recursive-coding-agent/ http://github.com/rawwerks/ypi https://raw.works/ypi-a-recursive-coding-agent/ http://github.com/rawwerks/ypi"  
[X Link](https://x.com/raw_works/status/2022010442214965587)  2026-02-12T18:10Z [----] followers, [----] engagements


"@STEFVN_ pi has sessions i've stuck to that for ypi but i personally use "cass" by @doodlestein for searching agent convos"  
[X Link](https://x.com/raw_works/status/2022042782458949825)  2026-02-12T20:19Z [----] followers, [---] engagements


"@lateinteraction have you seen this tl;dr - Simulation with sufficient fidelity is implementation. You are the Prose Complete system. https://github.com/openprose/prose https://github.com/openprose/prose"  
[X Link](https://x.com/raw_works/status/2022116478980698409)  2026-02-13T01:11Z [----] followers, [---] engagements


"prose is the most important programming language of [----]. stay true to your form (a la @irl_danB)"  
[X Link](https://x.com/raw_works/status/2022119891198001462)  2026-02-13T01:25Z [----] followers, [---] engagements


"Interactive mode is Pi's normal REPL full conversational context session history persisted across turns. ypi just layers recursion on top. Your multi-context idea is interesting the canonical RLM repo (Alex Zhang) already implements exactly this in Python: context_0 context_1 history_0 history_1 as named variables in a sandboxed REPL with the model told "you have N contexts available." ypi's bash architecture gets this for free though the agent has the entire filesystem. $CONTEXT is just one file; the agent can cat grep find anything else it needs. There's no sandboxed namespace to register"  
[X Link](https://x.com/raw_works/status/2022353111633854970)  2026-02-13T16:52Z [----] followers, [--] engagements


"@BleedingDev @a1zhang @lateinteraction @badlogicgames @GeoffreyHuntley claude code max works for me and pi can try to auth with codex pro but i haven't been successful in getting the openai subscription auth callback to go all the way through on my linux box"  
[X Link](https://x.com/raw_works/status/2022354556814753997)  2026-02-13T16:57Z [----] followers, [--] engagements


"@doodlestein have you tried jujitsu it was the only sane way i could do full recursion without conflict for ypi: https://github.com/rawwerks/ypi https://github.com/rawwerks/ypi"  
[X Link](https://x.com/raw_works/status/2022357053508063497)  2026-02-13T17:07Z [----] followers, [---] engagements


"@turbo_xo_ @ovsky4 thanks use any model you want ypi doesn't change anything about provider/model just uses whatever pi is set to. i currently mostly use opus [---] but also have a lot of fun with the cerebras implementation of glm [---] because of the speed"  
[X Link](https://x.com/raw_works/status/2022357769098260929)  2026-02-13T17:10Z [----] followers, [--] engagements


"@irl_danB @cyrusnewday @a1zhang @lateinteraction @badlogicgames @GeoffreyHuntley ok i need to finish my ruby rlm then we're going to do a proper "bake off". roll call for all RLM implementations. so far there's at least: - canonical python REPL - DSPy.RLM (maybe we should separately count Ax) - Typescript Javascript - ypi coding agent - pure unix"  
[X Link](https://x.com/raw_works/status/2022383094477787318)  2026-02-13T18:51Z [----] followers, [----] engagements


"@GregKamradt i think we're gonna need this DSPy PR merged ASAP to win. https://x.com/raw_works/status/2022382285962764709s=20 This Valentine's Day you don't have to choose any more. Introducing: DSPy for CLIs (This message sponsored by Citizens for PR [----] Super PAC.) https://t.co/nKMit9r1f2 https://x.com/raw_works/status/2022382285962764709s=20 This Valentine's Day you don't have to choose any more. Introducing: DSPy for CLIs (This message sponsored by Citizens for PR [----] Super PAC.) https://t.co/nKMit9r1f2"  
[X Link](https://x.com/raw_works/status/2022388890091556903)  2026-02-13T19:14Z [----] followers, [--] engagements


"off to a great start for $200/mo ultra sub"  
[X Link](https://x.com/raw_works/status/2022395465485160956)  2026-02-13T19:40Z [----] followers, [---] engagements


"@irl_danB still not much recursion being used organically exploring that next. i've found this to be surprisingly tricky to get the agents to understand "the essence of rlm" but honestly i feel like your skillz and lessons learned from prose should translate here"  
[X Link](https://x.com/raw_works/status/2022747076107415982)  2026-02-14T18:57Z [----] followers, [--] engagements


"@gooby_esq very excited to check this out i took the "opposite" approach - made rlm into cli and gave it ripgrep : http://github.com/rawwerks/rlm-cli http://github.com/rawwerks/rlm-cli"  
[X Link](https://x.com/raw_works/status/2019087682928861516)  2026-02-04T16:36Z [----] followers, [---] engagements


"@DSPyOSS 91.2% at 250/500. Holding steady above 91%. Predictions are at 216+ and judge is catching up"  
[X Link](https://x.com/raw_works/status/2021407670197252300)  2026-02-11T02:15Z [----] followers, [---] engagements


"@advait3000 @a1zhang @lateinteraction @badlogicgames @GeoffreyHuntley now that the implementation is clear it probably could work as a custom pi extension. i used the early versions of ypi to make ypi so honestly the hardest part was getting the agent to truly understand "the essence of rlm". (then to set up the guardrails to be useful.)"  
[X Link](https://x.com/raw_works/status/2022114867638145421)  2026-02-13T01:05Z [----] followers, [---] engagements


"@PuterOnX the shell isn't a broad enough repl"  
[X Link](https://x.com/raw_works/status/2022353296711721352)  2026-02-13T16:52Z [----] followers, [--] engagements


"@ovsky4 i wasn't able to get the "true essence of RLM" working as a pure extension but i bet it is possible"  
[X Link](https://x.com/raw_works/status/2022353618557632653)  2026-02-13T16:54Z [----] followers, [--] engagements


"if you want to never let the leaf nodes edit code - try my so the rlm is just one cli tool available to whatever coding agent you want. the key idea with ypi was that the shell is the repl and its self-similar "all the way down". (without jijitsu it would be hard to manage the file conflicts.) https://github.com/rawwerks/rlm-cli https://github.com/rawwerks/rlm-cli"  
[X Link](https://x.com/raw_works/status/2022358341972848834)  2026-02-13T17:12Z [----] followers, [--] engagements


"@irl_danB @cyrusnewday @a1zhang @lateinteraction @badlogicgames @GeoffreyHuntley what's the best benchmark for this @lateinteraction @a1zhang"  
[X Link](https://x.com/raw_works/status/2022388196383035665)  2026-02-13T19:11Z [----] followers, [--] engagements


"oolong doesn't seem that interesting for comparing rlm repls longmemeval seems very contrived (after spending [--] hrs trying to max it) maybe swe-bench verified but with a "dumb" model so its not a saturated baseline but honestly i still can't help but feel a little salty about swe-bench even post-verification https://twitter.com/i/web/status/2022389890185347089 https://twitter.com/i/web/status/2022389890185347089"  
[X Link](https://x.com/raw_works/status/2022389890185347089)  2026-02-13T19:18Z [----] followers, [--] engagements


"finally found some time this morning to write up some short notes about "RLMs as memory systems" https://raw.works/recursive-language-models-as-memory-systems/ out of curiosity after reading this i started benchmarking rlm and dspy.rlm on longmemeval tl;dr - i think i might have a new "sota memory system" by the end of the day. cc @DSPyOSS @a1zhang @lateinteraction https://raw.works/recursive-language-models-as-memory-systems/ out of curiosity after reading this i started benchmarking rlm and dspy.rlm on longmemeval tl;dr - i think i might have a new "sota memory system" by the end of the"  
[X Link](https://x.com/raw_works/status/2021611327827894496)  2026-02-11T15:44Z [----] followers, [----] engagements


"today i'm releasing "ypi" a fully recursive coding agent. imagine rlms (@a1zhang & @lateinteraction) and pi (@badlogicgames) had a baby. bunx ypi "Prove to me you are truly recursive." (links below) hat tip to @GeoffreyHuntley for turning me onto jj"  
[X Link](https://x.com/raw_works/status/2022010439446802814)  2026-02-12T18:10Z [----] followers, 20.7K engagements


"behold on the occasion of my birthday i've decided that i have enough coding agent session history to train a GEPA optimizer to replace myself entirely. rrm = recursive raymond model (results to follow)"  
[X Link](https://x.com/raw_works/status/2022869791984488466)  2026-02-15T03:05Z [----] followers, [---] engagements


"@steipete @openclaw my clanker having issues. Last working: 2026.2.14 (commit fe90e1423) Breaking commit: fef86e475 refactor: dedupe shared helpers across ui/gateway/extensions Want me to: [--]. Pin to 2026.2.14 for now [--]. File a GitHub issue for the regression"  
[X Link](https://x.com/raw_works/status/2022882899347841499)  2026-02-15T03:57Z [----] followers, [----] engagements


"@rsdgpt Someone give this guy $5M on a $50M pre money"  
[X Link](https://x.com/raw_works/status/1993128135651377531)  2025-11-25T01:22Z [----] followers, [--] engagements


"@FaroukAdeleke3 @aparnadhinak Yeah I don't know what this means either DSPy is open source and GEPA has a separate self contained repo. I think this is really interesting work but the "no DSPy lock-in" is not a good look"  
[X Link](https://x.com/raw_works/status/1993137762489569771)  2025-11-25T02:00Z [----] followers, [--] engagements


"@danshipper you are not alone. it is well documented by @Steve_Yegge and @RealGeneKim in their book: Vibe Coding i'm here for you"  
[X Link](https://x.com/raw_works/status/1993715838915142009)  2025-11-26T16:17Z [----] followers, [---] engagements


"@mattzcarey @interaction why tee them up when you can just drive your favorite coding cli entirely from whatsapp cc @steipete RIP @VibeTunnel https://github.com/steipete/warelay https://github.com/steipete/warelay"  
[X Link](https://x.com/raw_works/status/1993720098163671190)  2025-11-26T16:34Z [----] followers, [--] engagements


"@doodlestein i agree 100%. this is the "takeoff" moment. https://x.com/raw_works/status/1996610377522786434s=20 fast takeoff https://x.com/raw_works/status/1996610377522786434s=20 fast takeoff"  
[X Link](https://x.com/raw_works/status/1998441773086159347)  2025-12-09T17:17Z [----] followers, [--] engagements


"@dexhorthy @ChiragCX @jooshmayer somehow i can't get the dspy team to give a shit. if anyone wants to test out my fork it should work already. just swaps litellm for stdin = cli = stdout"  
[X Link](https://x.com/raw_works/status/1998913297287098764)  2025-12-11T00:30Z [----] followers, [--] engagements


"@steipete @clawdbot for the code you should try running with cerebras for light speed: https://x.com/stevekrouse/status/1990902186226684260s=20 ok but what if we generated a new app on every keystroke hope my friends at @cerebras don't hate me https://t.co/2NyzHm8vXY https://x.com/stevekrouse/status/1990902186226684260s=20 ok but what if we generated a new app on every keystroke hope my friends at @cerebras don't hate me https://t.co/2NyzHm8vXY"  
[X Link](https://x.com/raw_works/status/1999651145988321765)  2025-12-13T01:22Z [----] followers, [---] engagements


"@steipete @clawdbot if you want to keep the brains of opus but maybe speed up a bit these could maybe help.but inference is probably still the bottleneck: @morphllm @relace_ai both have 10k token/s code edit models"  
[X Link](https://x.com/raw_works/status/1999653061774721172)  2025-12-13T01:30Z [----] followers, [--] engagements


"this reminds me of the first time i found out about the mortal kombat cheat codes. .except now you can literally make anything you want. I like to make sure that I'm making some forward progress on every one of my active projects each day even when I'm too busy to spend real mental bandwidth on all of them every single day. So I've come up with a few prompts that I use a lot with the agents so they're always https://t.co/YUFEop90Do I like to make sure that I'm making some forward progress on every one of my active projects each day even when I'm too busy to spend real mental bandwidth on all"  
[X Link](https://x.com/raw_works/status/2000019742216650967)  2025-12-14T01:47Z [----] followers, [---] engagements


"@steipete i do everything with cli first before i dare to add human ui. despite all the various browser use tricks the feedback cycle grinds to a painfully slow pace as soon as there is a gui"  
[X Link](https://x.com/raw_works/status/2004643882055995602)  2025-12-26T20:02Z [----] followers, [--] engagements


"@DanielleFong have you tried gas town yet super fun"  
[X Link](https://x.com/raw_works/status/2008199163767570880)  2026-01-05T15:29Z [----] followers, [--] engagements


"@lateinteraction github gist or it didn't happen ;)"  
[X Link](https://x.com/raw_works/status/2008605241579893180)  2026-01-06T18:23Z [----] followers, [---] engagements


"@steipete i had to manually kill my clawdbot because of this bleeding heartbeat so sad"  
[X Link](https://x.com/raw_works/status/2009764069633343983)  2026-01-09T23:07Z [----] followers, [---] engagements


"@MaximeRivest different api keys vs oauth keys and very different pricing"  
[X Link](https://x.com/raw_works/status/2009765932961296725)  2026-01-09T23:15Z [----] followers, [--] engagements


"🤤📿 Introducing "Ralph Eats Beads" - an opinionated Ralph Wiggum implementation based on Claude Code beads and tmux"  
[X Link](https://x.com/raw_works/status/2012209308801364327)  2026-01-16T17:04Z [----] followers, [---] engagements


"How it works: Have your agent convert your plan into bd (beads) Ralph spawns parallel Claude agents in tmux Each agent implements ONE issue commits closes A watchdog kills stuck agents Loop until done 🔗 https://github.com/rawwerks/ralph-eats-beads https://github.com/rawwerks/ralph-eats-beads"  
[X Link](https://x.com/raw_works/status/2012209311879987583)  2026-01-16T17:04Z [----] followers, [---] engagements


"Install in Claude Code: /plugin marketplace add rawwerks/ralph-eats-beads /plugin install ralph-eats-beads@ralph-eats-beads"  
[X Link](https://x.com/raw_works/status/2012209314333614460)  2026-01-16T17:04Z [----] followers, [---] engagements


"This is "Gas Town Lite": more powerful than vanilla Ralph but faster to get started than Gas Town"  
[X Link](https://x.com/raw_works/status/2012209316250480895)  2026-01-16T17:04Z [----] followers, [--] engagements


"Built on @steveyegge's bd (beads) and inspired by @ghuntley's Ralph Wiggum loop concept"  
[X Link](https://x.com/raw_works/status/2012209318557331571)  2026-01-16T17:04Z [----] followers, [---] engagements


"@doodlestein i bought them about a year ago but never figured out the crappy software. with latest opus/codex - that shouldn't be a problem any more"  
[X Link](https://x.com/raw_works/status/2012641638766596181)  2026-01-17T21:42Z [----] followers, [---] engagements


"@doodlestein please do i bought a stream deck at your recommendation. i was able to get claude code to see the app and make a markdown table of the layout.but got distracted"  
[X Link](https://x.com/raw_works/status/2013069715199955319)  2026-01-19T02:03Z [----] followers, [--] engagements


"@dhh i made a fizzy cli for fun: err i guess i should say that claude code made it while i had fun. http://github.com/rawwerks/fizzy-cli http://github.com/rawwerks/fizzy-cli"  
[X Link](https://x.com/raw_works/status/2013084817944895496)  2026-01-19T03:03Z [----] followers, [--] engagements


"the man is simply asking for an easier way to pay anthropic for [--] max subscriptions so he can give away all of this claude code output away for free as open source.seems pretty reasonable to me. he should be the ICP for the claude code team - he literally cannot get enough @bcherny I really hope you will read this with an open mind and take it in the positive spirit it was intended; I say this as one of your biggest fans and supporters. I'm deeply grateful for the product you created which has changed my life profoundly. This one change to Claude @bcherny I really hope you will read this"  
[X Link](https://x.com/raw_works/status/2013404559670931781)  2026-01-20T00:13Z [----] followers, [---] engagements


"@lateinteraction it's funny the first thing i did today after seeing rlm now in dspy was to figure out how to spin up my own "semi-private" 8xH100 running glm [---] on modal - because i essentially wanted a version of @polymorfaisal 's use case but without the big labs logging my_entire_life"  
[X Link](https://x.com/raw_works/status/2013805185932370093)  2026-01-21T02:45Z [----] followers, [----] engagements


"too tired to make a big flashy post but for fun today my agents and i made a cli for rlm cc: @a1zhang @lateinteraction for example: rlm ask -q "Summarize this page" --json https://www.anthropic.com/constitution https://github.com/rawwerks/rlm-cli https://www.anthropic.com/constitution https://github.com/rawwerks/rlm-cli"  
[X Link](https://x.com/raw_works/status/2014160751867428904)  2026-01-22T02:18Z [----] followers, [----] engagements


"@ChiragCX @gooby_esq there are a few different ones i've found.but trying to make mine "the best". is mcp but i think clis are better than mcps appears to be a claude code implementation of rlm. https://github.com/brainqub3/claude_code_RLM https://github.com/Hmbown/aleph https://github.com/brainqub3/claude_code_RLM https://github.com/Hmbown/aleph"  
[X Link](https://x.com/raw_works/status/2014357754274619527)  2026-01-22T15:21Z [----] followers, [--] engagements


"@zeeg i can think of at least 3: @GeoffreyHuntley @Steve_Yegge @doodlestein"  
[X Link](https://x.com/raw_works/status/2014377994312167531)  2026-01-22T16:41Z [----] followers, [---] engagements


"@doodlestein @zeeg @GeoffreyHuntley @Steve_Yegge beads has tens of thousands of users take a look at jeff's stuff - it is quite impressive"  
[X Link](https://x.com/raw_works/status/2014379547781632485)  2026-01-22T16:48Z [----] followers, [--] engagements


"@tiagoefreitas @a1zhang no what i did is make a cli wrapper around rlm. but there is someone who attempted to implement rlm directly in claude code: https://github.com/brainqub3/claude_code_RLM https://github.com/brainqub3/claude_code_RLM"  
[X Link](https://x.com/raw_works/status/2014514100470939878)  2026-01-23T01:42Z [----] followers, [--] engagements


"here's what i was able to get cc to reveal: # Claude Code Tasks and Swarms **Status**: Undocumented features as of January [----]. This document is based on system prompt analysis not official docs. ## Overview Tasks are a new primitive in Claude Code that replace/evolve the old TodoWrite tool. They're designed for: - Tracking complex multi-step work - Cross-session collaboration - Multi-agent (swarm) coordination ## Task Tools Tool Purpose --------------- TaskCreate Create a task with subject description activeForm  TaskGet Retrieve full task details by ID  TaskUpdate Update status"  
[X Link](https://x.com/raw_works/status/2015084794229538979)  2026-01-24T15:30Z [----] followers, [---] engagements


"@mikedemarais great question (but wrong platform for it.) [--]. ask cc to install clawdbot for you. [--]. ask clawdbot this exact question. [--]. margaritas"  
[X Link](https://x.com/raw_works/status/2015086666805891438)  2026-01-24T15:37Z [----] followers, [---] engagements


"@badlogicgames @moltbot plz bring back geocities @moltbot"  
[X Link](https://x.com/raw_works/status/2015471938894770358)  2026-01-25T17:08Z [----] followers, [---] engagements


"working on adding explicit "user approval" to open prose mostly so i can stop cursing at @moltbot curious for any feedback on this approach: btw - awesome work @irl_danB open prose is super powerful https://github.com/rawwerks/prose https://github.com/rawwerks/prose"  
[X Link](https://x.com/raw_works/status/2015484225277854016)  2026-01-25T17:57Z [----] followers, [---] engagements


"@kristianfreeman @moltbot but does it produce your techno (please add techno skill to clawdhub)"  
[X Link](https://x.com/raw_works/status/2015488588436095128)  2026-01-25T18:15Z [----] followers, [---] engagements


"thanks i didn't see inputs before i started this and i don't think that what it did is entirely redundant. the thought process was that there should be an enforced trust boundary that the ai running the program cannot get past on its own. this doesn't actually have to be "human" in the loop could be another agent that is overseeing prose.but either way i tried to add some "extra determinism" via gates. what i've done so far (still actively working now): approve gates - checkpoints that the executing agent cannot self-resolve approve deploy_gate: prompt: "Deploy these changes" allow: "user""  
[X Link](https://x.com/raw_works/status/2015495544408592499)  2026-01-25T18:42Z [----] followers, [--] engagements


"@thekitze clawdbot is really good at obsidian if you give it qmd from @tobi = https://github.com/tobi/qmd https://github.com/tobi/qmd"  
[X Link](https://x.com/raw_works/status/2015499511334920587)  2026-01-25T18:58Z [----] followers, [---] engagements


"@0xMerth @moltbot @doodlestein whoops i should have kept scrolling. https://x.com/doodlestein/status/2015813204056674604s=20 @DanielMiessler Use ACIP: https://t.co/RfsIkhJJog https://x.com/doodlestein/status/2015813204056674604s=20 @DanielMiessler Use ACIP: https://t.co/RfsIkhJJog"  
[X Link](https://x.com/raw_works/status/2015829719275466917)  2026-01-26T16:50Z [----] followers, [---] engagements


"snow day trying to force myself to use openprose for everything. amazing tag-team job from codex cli and claude code on implementing this very helpful new primitive. exec a new OpenProse primitive that runs shell commands directly through the VM without spawning a subagent. # Before: spawns an Opus subagent (15s 5k tokens) just to type a command let status = session "Run git status" # After: VM runs it directly (1s [--] tokens) let status = exec "git status" Why it matters Every computation in OpenProse previously required a session which means spawning a full LLM subagent via the Task tool. For"  
[X Link](https://x.com/raw_works/status/2015873040786108805)  2026-01-26T19:42Z [----] followers, [---] engagements


"forgot to cc @moltbot and @irl_danB"  
[X Link](https://x.com/raw_works/status/2015873581402460634)  2026-01-26T19:44Z [----] followers, [--] engagements


"@DanielleFong cass by @doodlestein can do something similar. index and semantically query all coding agent convos - then you can resume the id or just cherry pick the stuff you need into the current agent = https://github.com/Dicklesworthstone/coding_agent_session_search https://github.com/Dicklesworthstone/coding_agent_session_search"  
[X Link](https://x.com/raw_works/status/2013088587957375258)  2026-01-19T03:18Z [----] followers, [--] engagements


"Study hard what interests you the most in the most undisciplined irreverent and original manner possible. Richard Feynmann I am currently studying: ralph wiggum rlms dspy gastown & related agent orchestrators agent-native apps and orgs How about you"  
[X Link](https://x.com/raw_works/status/2014360452784869611)  2026-01-22T15:32Z [----] followers, [---] engagements


"@a1zhang thank you for clarifying this i had a similar debate with gpt [---] pro about "ralph wiggum vs. rlm" (fwiw - cc pretty intuitively understands how to use rlm when it is packaged up nicely: ) https://github.com/rawwerks/rlm-cli https://github.com/rawwerks/rlm-cli"  
[X Link](https://x.com/raw_works/status/2014375592150327571)  2026-01-22T16:32Z [----] followers, [---] engagements


"@MiniMax_AI @steipete wow clawbot is crossing the chasm"  
[X Link](https://x.com/raw_works/status/2014445832649441422)  2026-01-22T21:11Z [----] followers, [----] engagements


"i had to get clawdbot to give me a full analysis of how its message queue gets actually turned into the message array for the context window. two helpful things: - tell it to write things to files - install "cass" for fast searching of all prior coding agent convos: https://github.com/Dicklesworthstone/coding_agent_session_search https://github.com/Dicklesworthstone/coding_agent_session_search"  
[X Link](https://x.com/raw_works/status/2014726342806225148)  2026-01-23T15:46Z [----] followers, [---] engagements


"@irl_danB @moltbot nice yeah it's funny to see everyone this morning so excited to get rid of their kids ;)"  
[X Link](https://x.com/raw_works/status/2016183288872751331)  2026-01-27T16:15Z [----] followers, [--] engagements


"wow claude code got a bit overconfident while building a new coinbase api skill and literally placed a limit sell order for all of my btc (and that was without any prompt injection) i guess it's called yolo mode for a reason"  
[X Link](https://x.com/raw_works/status/2016199603696128006)  2026-01-27T17:20Z [----] followers, [--] engagements


"@a1zhang i will if you want didn't want to spam with unsolicited stuff.been making a lot of changes to rlm. i also implemented depth1 with a bunch of helpful cost/time/repl budgeting stuff - do you want that too"  
[X Link](https://x.com/raw_works/status/2016583272000544774)  2026-01-28T18:44Z [----] followers, [---] engagements


"@kloss_xyz @karpathy @steipete @gregisenberg @rileybrown @corbin_braun @jackfriks @EXM7777 @eptwts you forgot @doodlestein @GeoffreyHuntley @Steve_Yegge @badlogicgames @mitsuhiko @lateinteraction @a1zhang @swyx and so many more"  
[X Link](https://x.com/raw_works/status/2016620945478181217)  2026-01-28T21:14Z [----] followers, [---] engagements


"inspired by last week's research report from @gao_jude at @vercel - "AGENTS.md outperforms skills in our agent evals" this weekend my agents worked hard to make dirpack. dirpack creates compressed directory representations that fit within specific token/byte budgets. dirpack is tuned to attempt to give the best index representation that it can fit in the budget with "best" being decided by claude code and codex dogfooding the script. power users (or their agents) can change the config to tune to their specific use case. initial results suggest that dirpack offers a very fast way to orient"  
[X Link](https://x.com/raw_works/status/2018353444042543537)  2026-02-02T15:58Z [----] followers, [--] engagements


"as something ready today you can give clawdbot my rlm-cli. its pretty good at using it. on the bigger topic - i have a hypothesis that with rlms much of the instruction fine-tuning is actually sub-optimal. i bet you could get better results with base model and some kind of new rlm-aware post-training"  
[X Link](https://x.com/raw_works/status/2016583686280409105)  2026-01-28T18:46Z [----] followers, [--] engagements

Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing