Conviction Brief — March 2026

159 bookmarks · delta vs prior: N/A (first month in series) · rolls up 4 weekly briefs

Baseline opening. This is the first Conviction Brief in the time-series — there is no prior month to diff against. All conviction marks are opening positions, not MoM deltas. Treat as the starting line: directional and foundational, not comparative. The corpus is ~95% AI-builder Twitter (Claude Code, skills, memory, workflows), heavily SF-framed, with near-zero SEA/India or consumer signal despite those being the reader's highest-edge lanes.

The Month in One Move

The skill layer crystallized as the productization primitive of the AI era — and the highest-conviction founding signal is not "build skills" but "productize the labor that skills now make executable by one person." March's corpus is dominated by one structural event: Claude Code's skill system went from developer curiosity to mass-market operating layer in 30 days. Garry Tan shipped GStack (open-source skill factory, 15.7k bm) and got it into native Claude Code distribution by month-end. Thariq's "Lessons from Building Claude Code: How We Use Skills" (43.9k bm) and Akshay's "Anatomy of the .claude/ folder" (43.1k bm) are the two highest-engagement bookmarks of the month — both are skill-architecture teardowns. The Karpathy AutoResearch loop (generalized by HyperspaceAI to 9.3k bm, pointed at markets by Chris_Worsey at 8.0k bm) proved the pattern extends beyond code: any function with a measurable score can now be automated, scored, and iterated by an agent loop. Levie's framing — "if a task is already outsourced, it tells you three things: the company has accepted external delivery, there's an existing budget line, and the buyer is already purchasing an outcome" — is the sharpest founding signal in the corpus. It points directly at AI-native services for SEA/India SMBs: the skills layer makes the labor executable, the open-model trend makes the unit economics work, and the regional relationship is the only durable moat once the craft commoditizes monthly. Conviction 72/100 (opening), Window Opening.

Signal vs Noise

Four weekly briefs fired. Marking each call against what the full month revealed.

Net: the month's dominant signal (skills-as-product-layer) is real and accelerating but sits at NONE on the edge gradient. The edge-relevant signal (Levie's outsourcing-replacement framing, AutoResearch-as-labor-loop) is thinner in volume but higher in founding conviction. The corpus is sampling SF AI-builder tooling discourse, not the markets the reader can win in.

Conviction Marks

Baseline opening — all marks are starting positions, not deltas.

AI-Native Services (opening) — 72/100

• Confirming: Levie's outsourcing-replacement framing ("the buyer is already purchasing an outcome") is the single sharpest services-thesis signal in the corpus. The AutoResearch loop proved any function with a measurable score can be automated and iterated. Corey Ganim's explicit business model sketch (W4: "scan 30 days of Reddit/X → build custom skills library → charge $1-2K per vertical → $500/mo retainer") is the first concrete unit-economics sketch for a skills-as-services business. Felix Lee (Gotrade/ADPList, SEA founder) publishing "The Claude-Native Designer" signals the SEA operator network is beginning to think in AI-native terms.
• Disconfirming (required): zero bookmarks show AI-native services working in SEA/India SMB contexts. Every named example is US/SF-framed. The labor-cost-to-software-margin arbitrage is asserted, not demonstrated, in the reader's actual markets. No customer-side evidence anywhere.
• Net: highest-conviction edge-aligned thesis in the corpus, but the evidence is entirely supply-side. Driver weights: Levie framing +30%, AutoResearch-as-labor-loop +25%, Ganim unit-economics sketch +20%, SEA operator signal (Felix Lee) +15%, edge-fit alignment +10%. This is the spine of the time-series going forward.

Skills / Agent Harness as Product Layer (opening) — 68/100

• Confirming: dominant theme by volume and engagement. Thariq (43.9k bm), Akshay (43.1k bm), zodchiii's Top 50 Skills list (42.4k bm), Garry Tan's GStack (15.7k bm → native distribution). The skill layer is now the default abstraction between human intent and AI execution. Himanshu's "Agent Harness is the Real Product" (5.5k bm W1) named the thesis early.
• Disconfirming (required): this is dev infra / tooling. Edge-fit NONE. The craft layer is commoditizing in real time — Garry Tan open-sourced GStack, skill marketplaces are proliferating, and the knowledge encoded in a skill is becoming free content. The moat is not the skill; it's the relationship, the verifier IP, or the data.
• Net: high-conviction structural thesis, zero founding edge for this reader. Operator fluency thesis — use it to run the firm, don't found into the skill layer itself.

Memory / Second-Brain Architecture (opening) — 55/100

• Confirming: enormous volume. Obsidian + Claude Code is the settled stack. Real capability gradient exists in decay modeling (signulll), structured vaults (om_patel5), and AI-readable task management (dSebastien's TaskNotes).
• Disconfirming (required): saturated with content. Converging on a technical moat (the exact opposite of a GTM/distribution wedge). Anti-thesis risk live: bigger context windows could collapse "memory" back into "dump it in the prompt."
• Net: low-confidence opening. Edge-fit LOW. Not a Found target. Watch as an enabler for services plays — the memory stack makes a managed service stickier, but it's not the business.

Consumer AI (SEA/India + wellness) (opening) — 42/100

• Confirming: barely. Tanjina Islam's "side quests" post (26.7k bm) is the month's highest-engagement non-tooling bookmark — a consumer/behavior signal, not an AI signal. BasedBiohacker's pinealon/cognition posts (W2, W4) and limitlesstack's sleep-optimization compound (3.9k bm W4) are the only wellness-adjacent items. Dan Romero's "Strava for cooking" (W4 end) is a lone consumer-product signal.
• Disconfirming (required): n≈3, and none are AI-native consumer products in SEA/India. The corpus barely sees consumer at all.
• Net: low-confidence opening. Weights: consumer-behavior-signal (Tanjina) +50%, wellness-adjacent-content (biohacker/pinealon) +30%, lone-product-signal (Strava-for-cooking) +20%. Edge-fit HIGH if a real wedge appears; right now there's almost nothing to act on.

Prosumer / Creator AI (opening) — 58/100

• Confirming: the entire corpus is prosumer-adjacent — non-technical operators using Claude Code/Cowork as a creative and operational multiplier. edgaralandough's "snowball method" for content (4.3k bm W1), shannholmberg's Claude-for-marketers thread (7.4k bm W2), Stitch as "vibe design partner" (40.6k bm W3).
• Disconfirming (required): the prosumer wave is real but the product layer is commoditizing faster than companies can form. Stitch is free. GStack is open-source. The wedge is not the tool — it's the done-for-you delivery and the vertical judgment.
• Net: medium-confidence opening. Edge-fit MED-HIGH. The prosumer signal matters most as a demand indicator for the services thesis — these operators need managed delivery, not better tools.

Startup Wedges to Explore

Spread across archetypes deliberately.

1. Skills-as-services firm for SEA/India SMB verticals (AI-native services · domain operator who knows regional SMB workflows · "we build, maintain, and deliver the AI skill stack for your industry" · India/Indonesia/Philippines)

• Window NOW: the skill layer just productized (GStack, .claude/ folder, skill marketplaces). Corey Ganim's unit-economics sketch ($1-2K setup + $500/mo retainer per vertical) is the first concrete pricing model. The craft is commoditizing monthly — the only defensible layer is the regional relationship + done-for-you delivery + vertical-specific verifier IP.
• Edge-fit: HIGHEST (AI-native services, SEA/India). Found.
• First probe: pick ONE vertical (e.g. Indonesia SME tax compliance or India D2C ad operations), build the skill stack for that vertical using existing tools only — no build. Run 5 paid pilots. Falsifier: if they won't pay a recurring monthly fee vs one-off project pricing, the services thesis is wrong for this segment.

2. AutoResearch-loop-as-service for a regulated SEA/India function (AI-native services · technical operator + domain expert · "we run the verified loop that replaces your analyst function" · e.g. SEA insurance claims processing, India lending ops)

• Window NOW: the AutoResearch pattern was generalized (HyperspaceAI, 9.3k bm) and proven across domains (markets, prediction, marketing copy). The architecture is a commodity. The moat is the verifier — the encoded human judgment that scores the loop's output. A firm that owns domain verifiers for a vertical and licenses the verified loop as labor replacement is the services-business-with-software-margins that Levie's framing implies.
• Edge-fit: HIGHEST. Found (or Back a returning operator with domain expertise).
• First probe: hand-build one vertical eval set (e.g. 100 adjudicated insurance claims), run a generic AutoResearch loop against it, measure whether the loop + your verifier beats the loop alone by enough to charge for. Falsifier: if the verifier doesn't improve accuracy/quality by a margin the customer can feel, the IP isn't defensible.

3. AI-native creative studio for SEA/India D2C brands (prosumer → AI-native services · creative operator with regional D2C relationships · managed creative-as-a-service at flat fee)

• Window NOW: the creative tooling layer is commoditizing in real time (Stitch is free, GStack is open-source, the "snowball method" is a viral prompt). Non-technical operators are already producing agency-grade output — but they can't systematize, deliver at scale, or maintain quality across a brand portfolio. The wedge is the managed delivery relationship, not the tool.
• Edge-fit: HIGH (prosumer → services, SEA/India). Found if a sharp creative operator; Back otherwise.
• First probe: run 5 paid creative pilots for Jakarta/Mumbai D2C brands using existing tools only. Falsifier: if they won't pay a recurring monthly fee for creative-as-a-service (vs one-off project pricing), the managed-creative thesis doesn't hold for this segment.

The Next Category [MANDATORY]

Agent-to-agent communication / multi-agent orchestration as a substrate. Two of the month's most structurally significant bookmarks point here: shawn_pana's "I made Claude Code and Codex talk to each other" (3.2k bm W4) — agents communicating via shared terminal, no API, no special protocol — and Nikil Viswanathan (Alchemy CEO) open-sourcing ClawFlows, "100+ prebuilt workflows" where @davehappyminion "runs my life." Combined with the AutoResearch loop's multi-agent debate pattern (Chris_Worsey's "25 AI agents debate macro daily," AlexFinn's "5 different AI models autonomously meet and discuss my business twice a day"), the corpus is showing the emergence of agent-to-agent as a communication substrate — not a tool, a rail.

• The capability + why conviction is high: when agents can talk to each other through a shared interface (terminal, file system, message queue), the unit of work shifts from "one human + one AI" to "one human orchestrating N specialized agents." This is the substrate that makes the "30-agent company" (DAIEvolutionHub's viral framing, 8.7k bm W3) technically feasible. Conviction that this becomes a layer is high — the pattern is repeating across independent builders.
• Why edge-fit is low TODAY (blunt): this is dev infra / orchestration. Edge-fit NONE. The reader's edges are GTM, distribution, regional services, and consumer — none of which touch the agent-communication protocol layer.
• The edge you'd have to BUILD or BUY: you'd need to either (a) build the orchestration layer (technical, not this reader's strength) or (b) wait for the application layer to open and build the services company on top of someone else's agent-communication rail — which is realistic, and is exactly what wedges #1 and #2 above are doing. The multi-agent substrate is the enabler, not the business.
• Who is better positioned to found it: a technical infra founder (shawn_pana archetype) or a platforms company (Anthropic, Alchemy). For this reader the honest call is watch as enabler, don't found into the rail. The discomfort is mild here — unlike world models, this is an infra layer that serves the reader's thesis rather than competing with it.

Relevance Radar

• Skills as the default productization abstraction — the .claude/ folder is the new .git. Every operator, every workflow, every function is being encoded as a skill. This is the structural enabler under the entire services thesis: skills make labor executable by one person, which makes the services-margin arbitrage possible. Watch skill-marketplace formation and vertical-skill quality as leading indicators.
• AutoResearch loop as universal automation pattern — "anything with a measurable score" (Zhengyao Jiang). The loop is a commodity; the verifier is the IP. This is the technical engine under wedge #2. Watch for the first vertical where someone commercializes a verified loop as a service — that's the tripwire from pattern to company.
• Open-model / model-agnostic trend making inference cheap — Perplexity as "$20B company that built zero models" (aakashgupta W4), OpenClaw as model-agnostic ("plug in whatever LLM you want"). The cost floor for AI-native services is dropping toward zero variable cost. This is the unit-economics engine under the entire Found thesis for price-sensitive SEA/India markets.
• "From Hierarchy to Intelligence" (jack, 24.3k bm, Mar 31) — Jack Dorsey's month-closing essay on org structure transformation. High-engagement, low-specificity. Flag as a macro tell: the "company as AI organism" framing is going mainstream, which creates both tailwind (acceptance of AI-native services) and risk (narrative saturation before the businesses are real).

Corpus Blind Spots

• The bookmarks are ~95% SF AI-builder Twitter (Claude Code, skills, memory, workflows, harnesses). The reader is sampling the discourse of the tool-builders, not the tool-buyers. SEA/India is nearly absent as primary subject despite being the reader's HIGHEST-edge geography. The one SEA operator signal (Felix Lee) is a designer-who-uses-Claude post, not a market signal.
• Zero customer-side or demand-side voice. Every services/consumer signal is a builder announcing a tool or a VC asserting a category. There is not a single bookmark from an SMB, a D2C brand, or an end customer saying what they'd pay for. The corpus measures supply-side hype, not pull. This is the single biggest blind spot — the Found thesis is built entirely on inference, not evidence.
• Wellness, the named edge archetype, is essentially missing. The closest hits are biohacker/pinealon posts (W2, W4) and Hawks0x's "7 Claude prompts for fitness" (11.0k bm W4) — these are wellness-adjacent content, not wellness-product signals. If wellness is a real lane, the follow-list isn't feeding it.
• No emerging-markarks operator or founder appears. The corpus has zero bookmarks from SEA/India founders building AI-native companies. Either they're not posting (sampling problem) or they don't exist yet (market-timing signal). Either way, the reader has no peer network signal from their highest-edge geography.
• The "org transformation" signal (jack, Joanne Chen's "Great Reorg," Matt MacInnis at Rippling) is all US enterprise. The "AI replaces the PM layer" framing (clairevo, aakashgupta's "vibe CPOing," stanine at Rippling) is real but entirely US-enterprise-framed. No signal on whether this applies to SEA/India mid-market or SMB.

Q1–Q5

• Will SEA/India SMBs pay recurring (not project) fees for AI-delivered services/creative? The entire Found thesis lives or dies on retention economics in price-sensitive markets — and the corpus has zero customer-side evidence. Corey Ganim's $500/mo retainer sketch is the only pricing data point, and it's US-framed.
• Where does the AI-native services margin actually land after delivery friction — software-margin (70%+) or dressed-up consultancy (20–30%)? Levie asserts the former; nobody's shown the post-friction number in a real vertical with real customers.
• Is the verifier (encoded human judgment inside the AutoResearch loop) a defensible moat, or does it commoditize as fast as the skills layer? If Garry Tan open-sourced GStack in a week, what stops someone from open-sourcing a vertical verifier? The answer determines whether wedge #2 is a company or a project.
• Does the prosumer wave (non-technical operators running Claude Code as their entire workday) create demand for managed services, or does it eliminate the need for services entirely? If every D2C brand owner can run their own creative stack, the managed-creative wedge (wedge #3) dissolves. The corpus doesn't answer this — it only shows supply-side tooling.
• When does the first SEA/India AI-native services company appear in this corpus? The absence is itself a signal: either the market is earlier than the SF discourse suggests (good — less competition), or the SF discourse is leading the market by too wide a gap (bad — the thesis may not transfer). The reader needs a tripwire: the first bookmark from a SEA/India founder building an AI-native services company.

Key Reads

• Anthropic has launched free courses to master AI with certificates for $0.00 — Peter Agboola, Mar 1 (117.8k bm)
• Lessons from Building Claude Code: How We Use Skills — Thariq, Mar 17 (43.9k bm)
• Anatomy of the .claude/ folder — Akshay, Mar 21 (43.1k bm)
• Top 50 Claude Skills & GitHub Repos for AI — darkzodchi, Mar 20 (42.4k bm)
• Meet the new Stitch, your vibe design partner — Stitch by Google, Mar 18 (40.6k bm)
• Claude Cowork Masterclass for Beginners — Corey Ganim, Mar 2 (31.5k bm)
• You're bored because you're not doing side quests — Tanjina Islam, Mar 29 (26.7k bm)
• From Hierarchy to Intelligence — jack, Mar 31 (24.3k bm)
• My chief of staff, Claude Code — Jim Prosser, Mar 5 (23.8k bm)
• I wanted to share a bunch of my favorite hidden and under-utilized features in Claude Code — Boris Cherny, Mar 30 (51.8k bm)

Signal Brief — Week of 2026-03-01

29 new bookmarks this week

This Week's Opening

The window is skills as a first-class productization layer — not "prompt engineering" but structured, shareable, composable skill files that turn Claude Code from a chatbot into an operating layer. Anthropic's Cowork Skill builder (emollick, 1.4k bm), the Google Workspace CLI with 40+ agent skills (addyosmani, 20.5k bm), and Himanshu's "Agent Harness is the Real Product" (5.5k bm) all landed in the same week. The shift: skills went from developer hack to mass-market abstraction in days. This is the structural enabler under the reader's highest-edge archetype — in SEA/India, labor-replacement-as-software only pencils when one person can productize the work, and skills are what make that executable. Archetype: enabler for AI-native services. Edge-fit: the cluster itself is NONE (dev infra), the implication is HIGHEST.

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Skills-as-services for one SEA/India SMB vertical — speculative. The skill layer just productized (Workspace CLI, skill-builder, harness-as-product). A firm that builds, maintains, and delivers a vertical-specific skill stack for SEA/India SMBs — e.g. "your entire back-office workflow encoded as skills, delivered as a managed service" — could capture the outsourcing-budget-line Levie named. Why now: skills make one person productizable, which is the precondition for services-margin arbitrage in price-sensitive markets. Cheapest probe: pick one vertical, build the skill stack using existing tools only, run 5 paid pilots at a flat monthly fee. Falsifier: if SMBs won't pay recurring vs project pricing, the thesis doesn't hold for this segment.
• AI-native creative studio for SEA/India D2C — speculative. The "snowball method" (4.3k bm) and Claude-native designer framing (4.0k bm) prove non-technical operators can produce agency-grade creative. But they can't systematize or deliver at scale. A managed creative-as-a-service at a flat fee for Jakarta/Mumbai D2C brands, using existing tools only, captures the relationship layer while the craft commoditizes. Cheapest probe: 5 paid creative pilots for D2C brands. Falsifier: if they won't pay recurring for creative-as-a-service, it's a project business, not a services company.

What's Breaking

The skill-layer commoditization cuts against any wedge premised on the skill itself as the moat. If Garry Tan open-sources GStack next week (he will) and the .claude/ folder becomes public knowledge (it already is, via Akshay), then "we have better skills" is not defensible. Anything in the pipeline that leaned on skill quality as the wedge should re-anchor on distribution, regional trust, or encoded vertical judgment instead. The moat is the relationship and the verifier — not the skill.

Carry-Forward

Watch whether the "skills as productization layer" framing produces an actual vertical services company (someone selling a skills-powered managed service in a real industry) versus staying stuck as dev-tooling content. The Levie outsourcing-replacement framing is the tripwire: if someone operationalizes "replace the outsourced budget line with an AI-delivered service" in the next 30 days, that's the first real data point for the Found thesis.

Signal Brief — Week of 2026-03-08

30 new bookmarks this week

This Week's Opening

The window is the AutoResearch loop as a generalized automation pattern — Karpathy's architecture escaping its ML-training origin and porting to "anything with a measurable metric." Varun Mathur's HyperspaceAI generalized it to plain-English optimization problems (9.3k bm), Chris_Worsey pointed 25 debating agents at macro/markets (8.0k bm), and Manthan Gupta's teardown hit 5.6k bm. The pattern is the same: propose → score → mutate → keep-if-better → repeat. This is the technical engine under the reader's highest-edge thesis: a verified loop that replaces an analyst function is a services business with software margins, IF the verifier (the encoded human judgment) is defensible. Archetype: enabler for AI-native services. Edge-fit: the loop itself is NONE (infra), the verifier IP is HIGHEST.

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Verifier-IP studio for one SEA/India regulated function — speculative. The AutoResearch loop is a commodity; the verifier (the encoded human judgment that scores the loop's output) is the asset. Sketch: a firm that builds and owns domain verifiers/evals for a vertical (e.g. SEA insurance claims, India lending ops) and licenses the verified loop as a service — replacing an outsourced analyst function at software margins. Why now: the loop architecture was just generalized (HyperspaceAI) and proven across domains (markets, ML). Everyone has the loop; almost no one has the encoded judgment. Cheapest probe: hand-build one vertical eval set (100 adjudicated cases), run a generic loop against it, measure whether your verifier improves accuracy enough to charge for.
• Managed marketing-ops desk for SEA/India D2C — speculative. The Claude-for-marketers wave (7.4k bm, 11.6k bm) proves non-technical operators can produce agency-grade output — but they can't systematize, maintain quality, or deliver across a brand portfolio. A managed service that runs a D2C brand's entire marketing-ops stack (content, competitive intel, creative) as a flat monthly fee, using existing tools only, captures the relationship layer while the craft commoditizes. Cheapest probe: run marketing-ops for 3 Jakarta/Mumbai D2C brands for 30 days using existing Claude skills. Falsifier: if the hand-holding IS the whole product and it doesn't compound, it's a consultancy.

What's Breaking

The GStack open-source release and the AutoResearch generalization together cut against any wedge premised on "we have better tooling" or "we have a proprietary loop." Both are now free. The only thing that isn't free is (a) the regional relationship, (b) the vertical verifier/eval, and (c) the encoded domain judgment. Anything in the pipeline that leaned on tool quality or loop architecture as the moat should re-anchor on one of those three.

Carry-Forward

Watch whether the AutoResearch-as-generalized-pattern framing produces an actual vertical services company — someone selling a verified loop as labor replacement in a real industry — versus staying as dev-tooling content. Chris_Worsey's "25 agents debating markets daily" is the closest to commercialization; if he (or someone like him) starts selling that output as a service to a fund or a corporate, that's the tripwire from pattern to company.

Signal Brief — Week of 2026-03-15

41 new bookmarks this week

This Week's Opening

The window is the skill layer reaching mass distribution and the emergence of the "one-person company" as an operational reality — not a meme but a demonstrated workflow pattern. Garry Tan shipped GStack into native Claude Code distribution (5.4k bm for /office-hours, then "just say install gstack and it works" by W4). Thariq's "Lessons from Building Claude Code: How We Use Skills" (43.9k bm) and Akshay's "Anatomy of the .claude/ folder" (43.1k bm) are the two highest-engagement bookmarks of the entire month — both are skill-architecture teardowns that made the skill layer legible to non-developers. Stitch's "vibe design partner" launch (40.6k bm) with DESIGN.md as an agent-readable design system file extended the pattern to creative: the interface between AI and existing productivity surfaces is being standardized. The implication for the reader: the productization layer is now fully built and open — the founding opportunity has moved from "build the tool" to "productize the labor the tool makes executable." Archetype: enabler for AI-native services. Edge-fit: the cluster itself is NONE (infra/tooling), the implication is HIGHEST.

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Vertical verifier-IP firm for SEA/India regulated function — speculative. The skill layer is fully productized (Thariq, Akshay, GStack). The AutoResearch loop is a commodity (Zhengyao, AlexFinn). The only thing not free is the verifier — the encoded domain judgment that scores the loop's output. A firm that builds and owns domain verifiers for a vertical (e.g. Indonesia SME tax compliance, India insurance claims) and licenses the verified loop as a service replaces an outsourced analyst function at software margins. Why now: the loop architecture was generalized this month, and the community proved it ports to anything with a score. Cheapest probe: build one vertical eval set (100 adjudicated cases), run a generic loop with and without your verifier, measure the quality delta. Falsifier: if the verifier doesn't improve accuracy by a margin the customer can feel, the IP isn't defensible enough to charge for.
• Managed AI-ops desk for SEA/India mid-market — speculative. The "30-agent company" and "zero human company" patterns (8.7k bm, 3.9k bm) prove the tooling works — but the operators hitting the complexity wall are exactly the ones who need managed delivery. A service that runs a mid-market company's entire AI-ops stack (competitive intel, content, CRM, analytics) as a managed service at a flat monthly fee, using existing tools only, captures the relationship layer while the craft commoditizes. Why now: noahzweben's cloud-scheduled recurring tasks (5.8k bm) make 24/7 delivery technically feasible. Cheapest probe: run AI-ops for 3 mid-market companies for 30 days using existing Claude skills. Falsifier: if the service doesn't compound (same effort month 3 as month 1), it's a consultancy, not a productized service.

What's Breaking

The mass-distribution of the skill layer (Thariq at 43.9k bm, Akshay at 43.1k bm, zodchiii at 42.4k bm) kills any remaining wedge premised on "we know how to set up Claude better." The knowledge is now fully public, fully indexed, and fully free. Three of the month's top-5 bookmarks are skill-architecture teardowns. The window for "Claude setup consultant" has closed — it was never a company, and now it's not even a project. Anything in the pipeline that leaned on setup expertise or skill quality as the moat should re-anchor on vertical verifier IP, regional relationship, or encoded domain judgment.

Carry-Forward

Watch whether the "30-agent company" pattern produces an actual AI-native services company in SEA/India — someone selling verified-loop-as-labor-replacement in a real industry — versus staying as SF tech discourse. The skill layer is built, the loop is commoditized, the cloud-scheduling is live. The missing piece is the vertical verifier and the regional delivery relationship. That's the reader's edge. The tripwire: the first bookmark from a SEA/India founder building an AI-native services company.

Signal Brief — Week of 2026-03-22

44 new bookmarks this week

This Week's Opening

The window is agent autonomy and agent-to-agent communication emerging as a substrate — not "use more AI" but agents that schedule themselves, talk to each other, and run without human intervention between handoffs. shawn_pana made Claude Code and Codex talk to each other via a shared terminal with no API (3.2k bm). Nikil Viswanathan (Alchemy CEO) open-sourced ClawFlows with 100+ prebuilt workflows where @davehappyminion "runs my life." aakashgupta's "6 levels of Claude Code autonomy" (4.8k bm) and bcherny's "no more permission prompts" (1.2k bm, 5.6k likes) are the infrastructure moves. Meanwhile, coreyganim posted the first explicit unit-economics sketch for a skills-as-services business: "scan 30 days of Reddit/X → build custom skills library → charge $1-2K per vertical → $500/mo retainer" (2.4k bm). That's the Found thesis in one tweet. Archetype: enabler for AI-native services. Edge-fit: the infra cluster is NONE, the services implication is HIGHEST.

Moving Now

NEW or MOVED this week only:

Wedge Sketches

• Skills-as-services firm for SEA/India SMB verticals — with first unit-economics anchor — speculative, strengthening. Corey Ganim's pricing sketch ($1-2K setup + $500/mo retainer per vertical) is the first concrete unit-economics data point in the corpus. The model: scan a vertical's public discourse → build custom skills library → charge setup + monthly retainer for maintenance and updates. The SEA/India equivalent: pick one vertical (e.g. Indonesia SME tax, India D2C ad ops), build the skill stack, charge a monthly retainer for managed delivery. Why now: the skill layer is fully productized, the autonomy infrastructure is production-grade, and the first pricing model is public. Cheapest probe: 5 paid pilots at $500/mo in one vertical using existing tools only. Falsifier: if SMBs won't pay recurring vs project pricing, the services thesis doesn't hold.
• Vertical eval/verifier IP studio — speculative, strengthening. Vtrivedy10's "How we build evals for Deep Agents" (5.2k bm) and aakashgupta's "the prompt is temporary, the eval is permanent" are the technical proof: the verifier IS the durable IP layer. A firm that builds and owns domain evals/verifiers for a vertical (e.g. SEA insurance claims, India lending ops) and licenses the verified loop as a service owns the layer that doesn't commoditize. Why now: the loop is a commodity, the skill is free, the workflow is open-sourced — the only thing not free is the encoded human judgment. Cheapest probe: build one vertical eval set, run a generic loop with and without it, measure the quality delta. Falsifier: if the verifier doesn't improve accuracy enough to charge for, the IP isn't defensible.

What's Breaking

The agent-to-agent communication substrate (shawn_pana, 3.2k bm) and the autonomy levels (aakashgupta, 4.8k bm) together cut against any wedge premised on "we run agents better" as a moat. If agents can talk to each other through a shared terminal and run autonomously on schedule, then "we manage your agents" is not a defensible service — it's a setup task that's already free. The moat must be the vertical judgment (the verifier), the regional relationship, or the proprietary data — not the orchestration. Anything in the pipeline that leaned on "we run better agent loops" should re-anchor immediately.

Carry-Forward

Watch whether coreyganim's skills-as-services pricing sketch produces an actual company — someone operationalizing the $500/mo retainer model in a real vertical with real customers — versus staying as a viral tweet. That's the tripwire from content to business. Also watch whether the evals-as-permanent-asset framing (aakashgupta, Vtrivedy10) produces a commercial eval/verifier product — the first vertical eval sold as a service is the second tripwire.