← research

THE TOP-VOICES LEDGER — one highest-leverage item per voice (sweep 2026-07-07)

Jul 14, 2026

Research spree lane 4. Practitioner-weighted: for each voice, the ONE most actionable idea/technique/artifact from roughly the last 60 days (since ~2026-05-08) that THIS user can exploit — not their fame, their leverage. Deduped hard against RESEARCH_fable_salvage.md, RESEARCH_spree_synthesis.md, and the loop-engineering skill's Provenance. Where a voice's best item was already banked, that's stated in one line and their SECOND-best (not banked) is the ledger item.

Method: 4 parallel Sonnet lanes via agent-reach/opencli (twitter search/tweets/thread/article, raw YAML banked to scratchpad), blog fetches for canonical bloggers, yt-dlp for YouTube-first voices. X search rate-limited (429) mid-run in three lanes; recovered via the tweets timeline endpoint and direct fetches — all quotes are from primary-source output, not memory.

The ledger (21 voices)

1. Andrej Karpathy (@karpathy)

  • Item: The "ask for HTML" output-modality default. "This works really well btw, at the end of your query ask your LLM to 'structure your response as HTML', then view the generated file in your browser… 3) HTML (still procedural with underlying code, but a lot more flexibility on the graphics, layout, even interactivity) <-- early but forming new good default." — x.com/i/status/2053872850101285137 (May 11, 2026)
  • Why highest-leverage: zero-setup, copy-pasteable, and carries a theory (vision is the preferred output channel from AIs) that generalizes to every report/plan/review artifact.
  • Exploit: make HTML-artifact the default output for every Mission Control analysis surface (skill fire-count audits, context-tax gauge, quota reports) instead of markdown-in-chat. Converges with Thariq (#3) and delba (#8) below — see Top-5 #2.
  • Already-banked? His #1 items were (LLM Knowledge Bases; Claude Tag "3rd UIUX redesign") — this is the second-best, NOT banked.

2. Boris Cherny (@bcherny)

  • Item: Subagents now run in the background BY DEFAULT. "In the next version of Claude Code: subagents run in the background by default, so you can keep talking to Claude while your subagents work. If you want your agent to run in the foreground, just tell Claude." — x.com/bcherny/status/2071647677591466098 (Jun 29, 2026)
  • Why highest-leverage: a default-behavior flip in the exact mechanism this user's fan-out doctrine depends on; the escape hatch is natural language, not a flag.
  • Exploit: one-line addendum to CLAUDE.md "Subagent pinning + fan-out": background-by-default means claude agents --json (busy/idle/waiting-on-permission) is now the PRIMARY subagent-state feed, not a fallback; Mission Control's fleet board should assume it. Runner-up detail: background sessions default to a worktree, labeled bg, can't /resume (https://x.com/bcherny/status/2053982327123132846).
  • Already-banked? His #1s were (/loop release, loop roster, fireside transcript) — this is second-best, NOT banked.

3. Thariq (@trq212, Anthropic)

  • Item: X Article "HTML is the new markdown." "I've stopped writing markdown files for almost everything and switched to using Claude Code to generate HTML for me." Example prompt: "Create a thorough implementation plan in a HTML file, be sure to make some mockups, show data flow and add important code snippets I might want to review. Make it easy to read and digest." — x.com/trq212/status/2052811606032269638 (article: /status/2052809885763747935) (May 8, 2026)
  • Why highest-leverage: a wholesale format-default argument (density, readability past ~100 lines, shareability, two-way interaction, better re-ingestion) with reusable prompts for specs/plans/verification — not a one-off trick.
  • Exploit: promote the existing "Use Lavish for decisions" memory into a general CLAUDE.md doctrine line: HTML over markdown for anything >100 lines or anything a human will actually read. Runner-up: his Claude Tag playbook — personal #thariq-claude channel, pinned persona message, status emojis ⏲️✅❓🛑 (https://x.com/trq212/status/2069474335656694003).
  • Already-banked? His #1 ("Finding Your Unknowns") banked — this is second-best, NOT banked.

4. Cat Wu (@_catwu, Head of Product, Claude Code)

  • Item: The claude agents control plane + her personal habit. "run claude agents for a control plane in your terminal! after, hit <- from any cli session to register that with the control plane. personally, i like to run claude agents from my root code dir to manage all my claude code agents in one place." — x.com/_catwu/status/2053999857799672111 (May 12, 2026)
  • Why highest-leverage: names the built-in live control-plane surface a session-dashboard builder must position against, plus the register-in-flight-sessions keystroke.
  • Exploit: deliberate positioning audit for Mission Control — link out to / embed the live control plane, and spend build effort only on what it can't do (historical transcript analytics: fire-counts, fallback forensics, context tax). Runner-up: her "workflow → Artifact → email → lock laptop, review async" pattern (https://x.com/_catwu/status/2073806626965049686) — a reusable shape for crypto-sweep-style fan-outs.
  • Already-banked? Partially — claude agents --json was flagged in RESEARCH_spree_synthesis as "best unused feed"; the control-plane UX, <- registration, and root-dir habit are new. Treat as half-new.

5. Andrew Ng (@AndrewYNg)

  • Item: "My 3 key loops for building 0-to-1 products" — agentic coding loop (minutes) / developer feedback loop (tens of minutes–hours) / external feedback loop (hours–weeks), plus the reframe: "Many people describe this human contribution as 'taste,' but I prefer to think of it as humans having a context advantage." — x.com/i/status/2071988145667928442 (Jun 30, 2026)
  • Why highest-leverage: turns "taste" from a vibe into an information problem — a design principle you can act on.
  • Exploit: when a cheap-model loop produces mediocre user-facing output, first check whether the loop is missing context only the human has (users, positioning, brand) before escalating models. Lightweight context-advantage.md per project; nuances the CLAUDE.md "taste >= 7" gate.
  • Already-banked? No.

6. Greg Brockman (@gdb)

  • Item: The self-improvement/skill-mining audit prompt he amplified (authored @reach_vb): "Look back over my recent work from the last 30 days… identify repeated manual workflows worth packaging… Only act on a candidate when it: occurred at least twice, or is clearly likely to recur and costly to repeat; has stable inputs, a repeatable procedure, and a clear output or stopping condition… Choose the smallest appropriate form: Skill … Custom subagent … Automation … Skip." — x.com/i/status/2058598608224858442 quoting x.com/reach_vb/status/2058538305872949490 (May 24, 2026)
  • Why highest-leverage: a production-grade audit-and-package loop with hard gating criteria that prevent over-eager skill creation.
  • Exploit: the mirror image of Mission Control's dead-skill audit (95/110 never fired = prune); this finds workflows that SHOULD be skills but aren't. Run monthly against transcripts as a scheduled loop. See Top-5 #4.
  • Already-banked? No.

7. Dario Amodei (@DarioAmodei)

  • Item: NULL — honest. Window output = one essay launch ("Policy on the AI Exponential," Jun 10, darioamodei.com — policy, not technique) + companion policy tweets. No tactical, exploitable item. One adjacent operational fact worth knowing (surfaced via Ng, Jun 19): Fable usage carried a mandatory 30-day data-retention policy — relevant if routing proprietary code through Fable sessions.
  • Already-banked? n/a.

8. delba (@delba_oliveira)

  • Item: The architecture-first code-review prompt. "You don't need to read code anymore. Read architectural changes first, then code, if necessary. Hey Claude, walk me through your architectural changes (before/after) in an html file: modules and their dependencies, seams, function signatures, etc. use visuals like mermaid diagrams. This works better if 1. you've given Claude a lot of ways to verify its work 2. you're iterating on architectural decisions with Claude upfront, before code gets written." — x.com/delba_oliveira/status/2073467304491233543 (Jul 4, 2026)
  • Why highest-leverage: copy-pasteable; converts raw-diff review into a before/after architecture walkthrough, with explicit preconditions.
  • Exploit: fold verbatim into the /code-review / /review flow as a pre-diff step. Runner-ups: /cd moves a live session to another directory keeping context and picking up the new CLAUDE.md; /add-dir grants access without moving (2074193750373376450); Ctrl+G opens the prompt in your editor and syncs back (2073782737182367803).
  • Already-banked? Her #1 (official loops taxonomy + ladder koan) banked — this is second-best, NOT banked.

9. Geoff Huntley (@GeoffreyHuntley, ghuntley.com)

  • Item: The spec-with-citations porting recipe (best available; predates window — Mar 15, 2026). "Run a ralph loop which compresses all tests into /specs/.md… link the implementation as citations in the specification… The key theory here is usage of citations in the specifications which tease the file_read tool to study the original implementation during stage 3."* — ghuntley.com/porting/
  • Why highest-leverage: the citations are the mechanism — they bait file_read re-grounding during execution; a reusable 3-stage recipe for any large migration.
  • Exploit: loop-engineering catalog entry ("/port-loop"): tests→specs-with-citations, code→specs-with-citations, then TODO-driven one-thing-per-loop execution.
  • Already-banked? No — but flag: NOTHING new in-window. Last 60 days = conference/travel content + a Jul 7 "invert the latent space… iykyk" post whose own replies read as a shitpost — not banked as a technique. Partial null.

10. Peter Steinberger (@steipete)

  • Item: Give the agent its own full computer (VM), not just a browser. "Give your agent its own computer to REALLY end to end test stuff." / "The best part: IT WILL CLICK ON ALL THE ANNOYING MACOS ALERTS FOR YOU" — setup: "codex and parallels." — x.com/steipete/status/2073214429655883814 (Jul 4, 2026)
  • Why highest-leverage: names the mechanism (Parallels VM handed to the agent) that breaks the browser/CDP-only ceiling — OS dialogs, permission prompts, native apps.
  • Exploit: upgrade path for automated-e2e/browser-harness: scoped VM-per-agent experiment for native macOS flows the CDP harness structurally can't reach. Runner-up one-liner: "try 'use imagegen to re-imagine this design and implement that'" (2073277317464682723) — generate the redesign image first, implement against it; feeds frontend-design.
  • Already-banked? His #1s (heartbeat/adversarial review/review-until-clean) banked — second-best, NOT banked.

11. shadcn (@shadcn)

  • Item: Pinned Messages + Notes scratchpad as harness UX. "1. Pinned Messages: Let me pin assistant messages to the sidebar for things I want to keep track of but am not ready to address yet. Render as a checklist & jump navigation. 2. Notes: Give me a scratchpad for thoughts while working." — x.com/shadcn/status/2060716123381780930 (May 30, 2026)
  • Why highest-leverage: a concrete, implementable fix for the "good idea surfaced mid-session, not ready to act, gets lost" failure mode; already seeded shipped clones.
  • Exploit: Mission Control companion feature — a pinned-insight/checklist panel over transcripts (starred or miner-flagged decision points) with jump-nav back into the source session; gives the implementation-notes.md Deviations doctrine a UI.
  • Already-banked? His #1 (/improve, "the plan is the product") banked — second-best, NOT banked.

12. theo (@theo, t3.gg)

  • Item: The GLM-5.2 cost correction. "That said - it's not cheap. Both Opus 4.8 and GPT-5.5 set to 'medium' are cheaper and smarter than GLM-5.2." — x.com/theo/status/2068533586131841072 (Jun 21, 2026; 5k likes)
  • Why highest-leverage: a falsifiable economic claim that attacks the "open-weight = cheap fallback" assumption baked into this user's own routing table.
  • Exploit: the CLAUDE.md fleet matrix scores GLM cost=9/int=5 as the overflow lane — theo's claim says that lane may be dominated. A/B one real task (GLM vs Opus-medium vs GPT-5.5-medium) before the next GLM dispatch; update the matrix with measured numbers.
  • Already-banked? His #1 (cost/intelligence/taste matrix) banked and in CLAUDE.md — second-best, NOT banked.

13. Addy Osmani (@addyosmani)

  • Item: "Agentic Autonomy Levels" — the run contract. "Every run of an agent should be preceded by a contract that defines what it's trying to do. The goal... The scope... Non-goals... Tools and permissions... Stopping condition: when to stop; ideally, a measurable variable. Evidence... Escalation... And budget." Plus the three-question autonomy test ("How quickly will we know we're wrong? How cleanly can we undo? What would prove we're right?") and four named anti-patterns: autonomy as status, permission laundering, summary substitution, fleet cosplay. — addyo.substack.com/p/agentic-autonomy-levels (Jul 3, 2026)
  • Why highest-leverage: the direct sequel to his banked loop-engineering canon — the field guide for how much rope each loop gets, as a checklist.
  • Exploit: extend the CLAUDE.md Definition-of-Ready gate (currently goal/files/acceptance/out-of-scope) with the two missing fields: escalation and budget; surface per-agent on the fleet board. Add "summary substitution" (never let the agent's summary stand in for the evidence packet) to fable-judgment. See Top-5 #3.
  • Already-banked? His #1 (loop-engineering + siblings) banked — this Jul 3 post is newer, NOT banked.

14. Martin Fowler (@martinfowler)

  • Item: "Interrogatory LLM." "The way I can do this is to prompt the LLM to interrogate me… insisting that the LLM ask only one question at a time." And the inversion: "give it a document… and then ask the LLM to interview a human expert to determine if the document is accurate." — martinfowler.com/bliki/InterrogatoryLLM.html (May 14, 2026)
  • Why highest-leverage: two-line technique that runs the interview pattern in BOTH directions — build a spec by interview, then audit an existing doc by interview.
  • Exploit: generalize CLAUDE.md's "interview one question at a time" into (1) build-context-by-interview for authoring new SKILL.md/PRD sections and (2) doc-audit-by-interview for reviewing the 110-skill catalog's stale entries. Runner-up: Böckeler's "Maintainability sensors for coding agents" 3-parter (May 19–27).
  • Already-banked? His #1 (harness engineering) banked — second-best, NOT banked.

15. Simon Willison (@simonw)

  • Item: uvx agentsview — live per-subagent cost auditing from INSIDE a session. "Run 'uvx agentsview --help' and then use that tool to calculate the cost of this session." Result from his sqlite-utils 4.0rc2 run: main session fable $141.02 vs five review subagents totaling ~$8; "I really should have followed my own advice and leaned more heavily into subagents with cheaper models." — simonwillison.net/2026/Jul/5/sqlite-utils-fable/ (Jul 5, 2026)
  • Why highest-leverage: a working, uvx-installable tool doing exactly the per-model, per-subagent dollar attribution Mission Control is building — plus a field datapoint (94% of cost on the orchestrator) that motivates a new alert.
  • Exploit: (1) run agentsview as an external validator against Mission Control's cost math or cannibalize its attribution logic into the fleet board; (2) new Mission Control alert: "orchestrator absorbed >X% of session cost — delegate more"; (3) his review technique: read the docs/changelog diff first to build the mental model of an agent PR.
  • Already-banked? No. See Top-5 #1.

16. swyx (@swyx)

  • Item: "Loopcraft: The Art of Stacking Loops" (AIEWF keynote). "It will be valuable to know when to go DOWN a loop when things go wrong (for reliability)… but it will probably be more valuable to know how to go UP a loop as models improve (for leverage)." Six nested loops: token → chat → agent turn → goal → automation → software factory. — x.com/swyx/status/2065307558198567206 (Jun 12, 2026)
  • Why highest-leverage: every banked loop voice describes what a loop IS; this is the decision rule for WHICH loop to operate in right now.
  • Exploit: add to loop-engineering SKILL.md as an explicit escalation/de-escalation rule: N consecutive checker rejections → drop one loop level; a model-version upgrade → re-test one level up.
  • Already-banked? No.

17. Harrison Chase (@hwchase17)

  • Item: "Wiki Memory." "Use an agent to turn raw source data into a compact, persistent, agent-readable knowledge layer... What is the raw data? → anything an agent can read or access. What is the best format? → files. How do you compress it? → an agent. How do you maintain it? → an agent." — langchain.com/blog/wiki-memory (Jun 30, 2026)
  • Why highest-leverage: a four-question checklist that settles the memory-substrate debate at "files, agent-compressed, agent-maintained."
  • Exploit: audit gbrain's write path against the four questions (is maintenance actually scheduled?), and build the missing distillation layer over Mission Control's 5.3k transcripts — a maintained wiki the tools query, instead of re-scanning raw JSONL per question.
  • Already-banked? No (distinct from the banked Karpathy KB post — this is Chase's operational synthesis).

18. Jarred Sumner (@jarredsumner)

  • Item: SIMD-optimized JSON parser in next Bun. "bun install & bun's bundler get a new SIMD-optimized JSON parser. In ideal conditions: package installs get up to 40% faster & peak memory usage drops by 2x." — x.com/jarredsumner/status/2071782393451782642 (Jun 30, 2026)
  • Why highest-leverage: free, quantified leverage on every bun install — zero workflow change, just upgrade.
  • Exploit: bun upgrade across active projects (CLAUDE.md defaults new projects to bun). Runner-up: bun build --react-compiler — Rust-native React Compiler, "19x faster than the Babel plugin" (2067892484009824536) — relevant if Mission Control's frontend is React.
  • Already-banked? His #1s (Bun Rust-rewrite pass-rate, agent-loop-perf team) banked — second-best, NOT banked. Honest note: this is routine-release leverage, the thinnest item in the ledger.

19. IndyDevDan (@indydevdan)

  • Item: PLANF3 ("PLANS For Fable 5") — his /plan meta-skill rebuilt and open-sourced. "Your planning skill is the single most important tool you and your agents have. It decides how good your results are BEFORE a single file gets modified… HTML-first plan output… Per-phase and per-task breakdowns with an embedded checklist your agents update live… Validation and testing sections with closed loops that block completion until every box is checked." —
  • Why highest-leverage: a complete shipped artifact, not prose — plan template with hard per-phase validation gates that block "done."
  • Exploit: cannibalize (1) HTML-first plan output, (2) the closed-loop per-phase checklist as a structured STOP-condition template for /goal finish conditions (upgrades the single pasted-proof gate to per-phase gates), (3) per-phase reference images for design review. Note: his X is stale; his current signal is YouTube.
  • Already-banked? No.

20. Guillermo Rauch (@rauchg — SUBSTITUTED for Jared Palmer)

  • Item: Agentic self-improvement via run introspection. "Give your agent the ability to introspect its past runs, spot inefficiencies, errors, redundant tool calls, and produce new prompts and skills." Shipped same-day: vc agent-runs CLI/MCP so an agent fetches its own traces. — x.com/rauchg/status/2073132174958841887 (Jul 3, 2026)
  • Why highest-leverage: converts learnings-capture from manual post-hoc to continuous and agent-driven; independent validation of Mission Control's premise.
  • Exploit: the next Mission Control feature — an agent-facing MCP endpoint so a LIVE session queries its own recent history (fire-counts, fallbacks, context-tax spikes, redundant tool calls) for self-diagnosis mid-run. See Top-5 #5.
  • Substitution note: Jared Palmer left Vercel Jul 2025 (now VP Eng @ Xbox) — honest null on him; Rauch is the Vercel-AI voice.

21. BONUS — Armin Ronacher (@mitsuhiko)

  • Item: "Better Models: Worse Tools." "Newer Claude models sometimes call Pi's edit tool with extra, invented fields… both Opus 4.8 and Sonnet 5 show it but none of the older models… Claude Code's own tools are comparatively flat… Anthropic's own client appears to expect and accept a fair amount of slop and repairs it, mostly silently." Fixes he measured: stripping thinking blocks from history halved the failure rate; strict tool mode eliminated it. — lucumr.pocoo.org/2026/7/4/better-models-worse-tools/ (Jul 4, 2026)
  • Why highest-leverage: root-cause finding (RL overfit to the Claude-Code-shaped harness) with two measured mitigations.
  • Exploit: shape any custom MCP tool schema (gbrain, Mission Control) flat like Claude Code's own tools; enable strict tool mode; candidate Mission Control gauge: malformed-tool-call retry loops as a detectable transcript pattern.
  • Already-banked? No.

TOP-5 — act on these now (ranked)

1. Install uvx agentsview and wire cost-attribution + an "orchestrator hog" alert into Mission Control (simonw). Run uvx agentsview today against recent sessions and diff its per-model, per-subagent dollar numbers against Mission Control's own cost math — it's a free external validator for the context-tax gauge and fleet board, and if its attribution logic (it walks session list --include-children) is better, cannibalize it rather than rebuild. Then ship the alert his datapoint motivates: when a session's top-level model absorbs more than ~80% of total cost (his fable run: $141 of $149 on the orchestrator), flag "delegate more to cheaper subagents" with the offending session linked. This lands directly on the #1 item in the existing build queue (context-tax gauge) with a working reference implementation in hand.

2. Promote "HTML is the new markdown" to a CLAUDE.md doctrine line and make it Mission Control's default output surface (Thariq + Karpathy + delba + IndyDevDan — four-voice convergence). Four independent practitioner voices converged on the same move this window: HTML artifacts as the default for anything longer than ~100 lines or anything a human reviews. Add one doctrine line ("Specs, plans, reviews, and reports render as HTML artifacts by default; markdown only for machine-consumed files"), fold delba's architecture-first review prompt verbatim into the /code-review flow (before/after modules, seams, signatures, mermaid — read the architecture HTML before the diff), and steal PLANF3's closed-loop per-phase checklist as the structured finish-condition template for /goal runs. Cheapest item on the list — it's prompt text, not code.

3. Extend the Definition-of-Ready gate into Osmani's full run contract, and add "summary substitution" to fable-judgment (Addy Osmani). The CLAUDE.md gate currently requires goal/files/acceptance-criteria/out-of-scope. Add the two fields Osmani's run contract has that it lacks — escalation (who/what gets pinged when the agent is stuck, instead of improvising) and budget (time/token cap written into the handoff, matching the existing "no cap, no run" rule) — so every subagent handoff carries all eight fields. Surface the contract per-agent on the fleet board. Separately, add one line to the fable-judgment skill: an agent's own summary never substitutes for the evidence packet (pasted output, diff, logs) — that's his "summary substitution" anti-pattern and it names the exact failure the pasted-proof doctrine guards against, now generalized beyond /goal to all reviews.

4. Schedule the inverse skill audit: Brockman/reach_vb's workflow-mining prompt as a monthly loop (gdb). Mission Control's existing audit finds skills that never fire (prune list); this prompt finds the opposite gap — repeated manual workflows that should be packaged but aren't. Run it monthly as a scheduled agent over the transcript corpus with its gating criteria intact (occurred ≥2×, stable inputs, clear stopping condition, not already covered; choose the smallest form: skill / subagent / automation / skip; shortlist with evidence before creating anything). Route candidates through the same review pipeline as the dead-skill audit so the catalog converges from both directions. Fits the loop-engineering 5-box test: weekly-plus recurrence, checkable gate (the evidence shortlist), hard cap, human approves creation.

5. Align Mission Control with the live control plane and background-by-default reality (Cat Wu + Boris Cherny). Two first-party facts change Mission Control's positioning: claude agents is the official live control plane (with <- registering any CLI session into it), and subagents now run background-by-default (worktree-backed, labeled bg, not /resume-able). Do the positioning audit this week: stop building anything the control plane already does live; make claude agents --json (busy/idle/waiting-on-permission) the primary state feed for the fleet board; and double down on the moat the control plane doesn't touch — historical forensics (fallback receipts, context tax, fire-counts) plus the rauchg-shaped next step: an agent-facing MCP endpoint so live sessions can query their own run history for self-diagnosis mid-task. Add the one-line CLAUDE.md addendum about background-by-default so fan-out sessions aren't misread as idle.

Honest nulls

  • Dario Amodei — policy essay cycle only; no tactical item. (X search also 429-rate-limited on retries; essay corpus checked directly — nothing newer.)
  • Jared Palmer — left Vercel Jul 2025; substituted Guillermo Rauch.
  • Geoff Huntley — nothing actionable in the 60-day window (conference content + a likely-shitpost "invert the latent space" tease); ledger item is his Mar 15 porting recipe, flagged as out-of-window.
  • Partial: Jarred Sumner's second-best is routine release leverage (his real agent-workflow items were already banked); Cat Wu's item is half-banked (claude agents --json was already flagged as a feed — the control-plane UX and habits are the new part).

Provenance

Sweep 2026-07-07, 4 parallel Sonnet lanes, agent-reach via opencli (twitter search / tweets / thread / article; X search 429'd mid-run in 3 lanes, recovered via the tweets timeline endpoint), curl/r.jina.ai for blogs (simonwillison.net, ghuntley.com, addyo.substack.com, martinfowler.com, langchain.com/blog, lucumr.pocoo.org), yt-dlp for IndyDevDan. Raw YAML/article caches in the session scratchpad (tw_*.yaml, th_*.yaml, art_*.md, ledger_A–D.md). Deduped against: RESEARCH_fable_salvage.md, RESEARCH_spree_synthesis.md, loop-engineering SKILL.md Provenance.