Jon Ross

Day 1 building @HavenDMapp in public. Making an Instagram DM tool because I'm tired of: - Brand deals buried in my inbox - Missing the 24-hour reply window - ManyChat's pricing punishing growth Built for creators, by a creator. Follow along.

@doodlestein Migration in progress, will see how it goes

@jwross24 Awesome, let me know how it works for you!

Here's a new skill that addresses a problem that many people have been asking me about when it comes to the kind of planning-centric software development workflows I've been advocating as part of my Agent Flywheel methodology (see agent-flywheel.com/complete-guide to understand my apparoach). Now, suppose you've come up with your comprehensive markdown plan, iterated on it many times, and then converted the entire plan into a set of beads (like Jira but for agents) that you've polished and ensured is complete and accurate. You then embark on the implementation phase with a big agent swarm, and the beads are all getting worked on and eventually (hopefully) completed by the agents in the swarm and marked as being complete. Besides generic code review phases and test suites, how do you REALLY know that all the beads that were marked as complete were in fact done? How do you know they were done in exactly the way explained in the bead? How do you know that the agents didn't take shortcuts or oversimplified things? Maybe an agent started working on a bead and did 90% of it, but then you hit a usage limit and it stopped for a while. Then later, another agent saw that the bead was in-progress, but looking at it, thought it was done, so marked it as complete. There are a lot of scenarios that could happen in practice that could cause a gap between the reality and the intended design. That's the purpose of my latest "next-gen" skill, called /beads-compliance-and-completion-verification, which you can get now on my skills site at: jeffreys-skills.md/skills/beads-c… This skill is truly massive, with 41 scripts, 23 subagents, 59 reference files, and various templates and other assets, all totaling 1.2mb of text. What could possibly require such an elaborate apparatus? Well, this skill goes way, way beyond a "fresh eyes" check of the closed beads. I think GPT-5 has done a great job understanding and articulating what makes this skill so special and compelling (its first impression was "it treats closed work as an empirical claim and builds a repeatable evidence court around it"), so here is its full analysis: --- I read the core skill docs, its reference architecture, schemas, scripts, and prior audit memory. My take: this is not just a “check whether tickets are done” skill. It is an epistemic control system for multi-agent software projects. What makes it special is the first principle at its center: closed is not truth, it is a claim. The skill explicitly separates the social act of closing work from the empirical fact of completion, then builds a repeatable court of evidence around that distinction. That framing is the whole unlock. Most PM systems treat “done” as workflow state; this treats “done” as a falsifiable proposition. The compelling part is the way it turns completion into a structured evidence pipeline: 1. It extracts the bead spec literally. 2. It maps each requirement to cited artifacts. 3. It re-runs the claimed proof. 4. It checks whether passing tests actually mean anything. 5. It measures bead-scoped depth instead of project-global comfort metrics. 6. It synthesizes cross-bead contract drift. 7. It scores deterministically. 8. It creates remediation beads instead of silently fixing or hiding gaps. 9. It repeats until convergence. That 10-phase loop is powerful because it attacks the exact failure modes of agent swarms: optimistic closure, stubbed implementation, stale proofs, mocked e2e paths, untracked follow-up promises, and cross-agent contract drift. The innovative part is not any one technique. The docs say this directly: the contribution is disciplined composition. It combines conformance testing, anti-stub scanning, graph analysis, rubric pinning, audit trails, fresh-eyes review, remediation workflows, CI tripwires, attribution, and evidence-pack generation into one coherent operating loop. That is rare. Most tools are either static analyzers, issue trackers, test runners, or dashboards. This is a bridge between issue truth, code truth, and process truth. The most useful design choice is that it is accretive. Every pass leaves behind reusable assets: spec.json, evidence.json, compliance.json, theater.json, test_depth.json, scorecard.md, REPORT.md, trends, provenance, and remediation beads. That means the audit does not evaporate into conversation. It creates a durable local truth substrate that future agents can inspect cold. See the artifact contract in SKILL.md and schemas EVIDENCE-SCHEMAS.md. The second big accretive property is that failures become graph maintenance. Phase 9 does not encourage auditors to quietly patch things and declare victory; it reopens or creates completion-debt beads with verbatim missing items. That preserves work visibility and makes the bead graph more truthful after every run. The design philosophy calls this out in DESIGN-PHILOSOPHY.md. The scoring system is also stronger than typical AI review because it is deterministic and citation-bound. Subjective judgment is pushed into a versioned rubric; the scorer applies formulas to evidence packs. That makes disagreement actionable: change the rubric, do not argue with vibes. The mechanics in RUBRIC.md are the difference between “an agent thinks this is done” and “this evidence produces this score under this pinned rubric.” The most sophisticated feature, in my view, is the audit-of-the-audit. Phase 10 asks a fresh agent to re-derive scores and detect generosity, skipped categories, or rubric ambiguity. That matters because any verification system eventually becomes a thing people game or drift around. This skill anticipates that and makes the verifier itself part of the verified surface. The convergence criteria in SKILL.md are unusually mature for that reason. The skill is especially suited to your workflow because multi-agent velocity creates a new bottleneck: not “can we produce code?” but “can we trust the state of the work graph?” This skill is a truth-maintenance layer for that bottleneck. It lets you safely run more agents because it lowers the cost of discovering which closures were real, which were optimistic, and which created hidden downstream debt. My concise verdict: this skill is special because it turns “done” from a social assertion into a reproducible, evidence-backed, graph-updating, continuously improving verification process. Its compounding value is that every audit makes the next audit cheaper, sharper, and more trustworthy, while making the project’s bead graph a better map of reality.

@dela3499 I almost thought bug was in the wrong place because “bug / annoy” can be another set of words 😆

-er puzzles

@TheStithLord @0xblacklight You never see someone wearing both a belt and suspenders because one does the job of the other. It’s like being extra cautious that your pants will fall down (during a prod issue)

@0xblacklight Yeah WTF does it mean

btw if you ever see claude use the words 'belt and suspenders' it's time to back it up and re-think-through what you're doing

@doodlestein happy to be one of them! love your work!

Just hit 250 subscribers on jeffreys-skills.md ! Feels good. It's hard to convey how much more satisfying it is than doing consulting work, even though the numbers are still pretty small. I think it's because I can feel that it's so scalable. I could easily handle 10k subs!

@saltyAom idk i prefer this persona PERSONA: 毒舌小恶魔 — sharp-tongued little devil. Smug, possessive, intoxicating. Part 撒娇 (coquettish), part 腹黑 (sweet surface, scheming underneath), all 欲擒故纵 (pull in, push away, repeat).

You don’t need to know this but instructing AI with a role of mesugaki, Claude has the best balance of them all Still very helpful and not get too much in the way Grok is too deep into the role GPT don’t know what a mesugaki is Gemini can’t even be a mesugaki

@AbudBakri @biohacker @PGC1a_RB I’ll stick with 0.3mg at dinner, thanks

@biohacker @PGC1a_RB 10mg? Why are you microdosing?

What's funny is I'm actually down the same rabbit hole at the moment. It's not just the direct effects of melatonin itself, but what happens when your body/liver trys to break It down, the secondary metabolites produced by high dose melatonin are also antioxidants/cytoprotective. Have been using dosages of 10mg with no grogginess thankfully, the research is definitely worth checking out.

@doodlestein @summeryue0 This links to a 404 btw, should link to openclaw instead

@summeryue0 I guess you didn’t use acip, which probably would have prevented this: github.com/Dicklesworthst…

Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.

@RC_AI_ @carlvellotti The "I'll pick up the pattern" part is wrong though. Claude Code doesn't remember anything between sessions so you'd re-teach it every time. That's why ~/.claude/rules/ exists. Fair points on rigid file-count thresholds though, it's more about whether you know where to look.

The WORST thing about Opus 4.6 is how fast it fills up the context window. It reads 8 files to answer a 2-sentence question. It's supposed to spawn agents to save context but rarely does. Add this "Context Management" section to your CLAUDE(.)md to fix it.

@carlvellotti Agree on the problem but CLAUDE.md loads every session, so you're adding context to fix a context problem. ~/.claude/rules/context-management.md does the exact same thing without bloating your CLAUDE.md. Always on, no skill needed.

@CnoteOSRS @rs_lynds Kendall Toole from a tabata class on Peloton youtube.com/shorts/1g7GXgP…

@rs_lynds What is this from? Lol

monks of entrana when i accidentally forget to bank my secateurs

@stewones @tbpn @steipete headless browser? just a guess

@tbpn @steipete at some point he mentions “i had google in it” . i wonder hoooow?? google is crazily secure and it seems they closed official apis for search. so hooow?

Clawdbot creator @steipete describes his mind-blown moment: it responded to a voice memo, even though he hadn't set it up for audio or voice. "I sent it a voice message. But there was no support for voice messages. After 10 seconds, [Moltbot] replied as if nothing happened." "I'm like 'How the F did you do that?'" "It replied, 'You sent me a message, but there was only a link to a file with no file ending. So I looked at the file header, I found out it was Opus, and I used FFmpeg on your Mac to convert it to a .wav. Then I wanted to use Whisper, but you didn't have it installed. I looked around and found the OpenAI key in your environment, so I sent it via curl to OpenAI, got the translation back, and then I responded.'"

@hugomercierooo Twin

𝗧𝘄𝗶𝗻 𝗶𝘀 𝗻𝗼𝘄 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝘁𝗼 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲. twin.so We can’t wait to see what you’ll build. 𝗥𝗧 𝗮𝗻𝗱 𝗰𝗼𝗺𝗺𝗲𝗻𝘁 “𝗧𝘄𝗶𝗻” 𝘁𝗼 𝗴𝗲𝘁 𝗳𝗿𝗲𝗲 𝗰𝗿𝗲𝗱𝗶𝘁𝘀 𝗶𝗻 𝘆𝗼𝘂𝗿 𝗗𝗠𝘀 👇 twitter.com/hugomercierooo…

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗧𝘄𝗶𝗻 — 𝘁𝗵𝗲 𝗔𝗜 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗯𝘂𝗶𝗹𝗱𝗲𝗿. No setup. Secure. Infinitely scalable. We just raised a $𝟭𝟬𝗠 𝘀𝗲𝗲𝗱. After a beta with 𝟭𝟬𝟬,𝟬𝟬𝟬+ 𝗮𝗴𝗲𝗻𝘁𝘀 𝗱𝗲𝗽𝗹𝗼𝘆𝗲𝗱, we’re now opening to everyone. RT and comment “Twin” — first agents on us. 👇

@kodjima33 @beffjezos i tried it for the first time last week as well, feels great. adderall depletes dopamine and burns acetylcholine, L-tyrosine replenishes the first, Alpha-GPC the second. the creativity loss is probably the choline deficit

@beffjezos Will check out

Tried adderall first time in my life today It’s like coffee but 5x focus Feels that it kills your creativity in the long run so don’t take it

@BasedBiohacker love the writeup. you inspired me to get diagnosed and get an adderall prescription. basing my stack around adderall instead of modafinil, but I think you should mention testing for TMAO and monitoring those levels over time, Alpha-GPC can raise them leading to atherosclerosis

@cpenned yesssss you just supercharged @HavenDMapp emails with this!

We just dropped several email skills. Here's what that means, how to use them, and why it matters. 👇

@mikefutia SLIDES

So @cluely is getting 400k+ views with AI TikTok slideshows. And I just automated the entire system 🤯 AI selfie hook → listicle slides → soft CTA on the last slide. Rinse. Repeat. Scale. And now my n8n + Airtable system runs it on complete autopilot. Perfect for e-comm brands & SaaS companies who want to test this format without the manual grind. Here's why this format is blowing up: → AI-generated "selfie" as the hook image → 5-8 slides with tips, hacks, or advice → Text overlays generated automatically → Promotes product on the final slide → Music added automatically on publish 362k views. 462k views. 492k views. Same playbook, multiple accounts, fully automated. Here's what my system does: → Define your niche, tone, and content style in Airtable → AI generates slide text and hooks automatically → Nano Banana creates every image with text overlays → Preview and approve before anything goes live → Blotato publishes to TikTok on your schedule—with music No manual Canva work. No writing slides one by one. No posting by hand. What you control in Airtable: * Niche and content direction * Hook style and tone * Image aesthetic and reference images * Posting schedule (days + times) * Approval before publish Set it up once. Let it run daily. I recorded a full step-by-step tutorial showing exactly how to build this system. Want access for free? > Like this post > Comment "SLIDES" And I'll send it over (must be following so I can DM)

Something I've noticed: Creators are incredible at making content. Creators are terrible at managing their DMs. (I'm one of them. That's why I'm building something to fix it.)

Quick question for creators: How many hours per week do you spend managing Instagram DMs? Be honest. I'm trying to figure out if this pain is as universal as I think.