{ "flags": [ { "id": "flag-1", "meeting_id": "GPM16-03", "kind": "worth_spreading", "title": "Banning "counts" made Plus One drop numbers everywhere—and a disclaimer scored 0/3", "face": "In open Plus One PR #1755, Engineer A's branch reports that telling the agent not to relay a provider's "candidate and attempt counts" made it replace timings like 250ms, 120ms, and 8s with vague words in unrelated clean replies. Adding a disclaimer that the ban "never overrides the numbers rule above" scored worst (0/3): "Priming does not read disclaimers." Today's commit deleted all six quantity words and stopped failing replies for quoting numbers.", "context": "In open PR EveryInc/plus-one #1755 (Claude-assisted, with an afternoon commit widening the degraded-job retry gate and updating the rubric), Engineer A fixes an incident where an external copy-edit job returned HTTP 200 with a crashed stage and the agent pasted "112 candidates and kept 5." Testing a vocabulary ban against the formatting eval suite exposed a priming regression.", "moment": [ { "line": 175, "speaker": "Engineer A", "quote": "A ban mentioning counts or tallies makes the model stop carrying ordinary numbers in prose, and an explicit disclaimer in the same clause does not stop it." }, { "line": 75, "speaker": "Engineer A", "quote": "2. **Stating the precedence explicitly** — the clause literally said it *never overrides the numbers rule above*. This scored **worst (0/3)**. Priming does not read disclaimers; a reassurance names the class as loudly as a ban." }, { "line": 78, "speaker": "Engineer A", "quote": "So the variable is the **word class**, not the placement or the phrasing. Every mention of counts, tallies, candidates, attempts, figures and numbers is now deleted from the addition, in both directions." } ], "why": "This gives you a concrete, measured prompt-design rule from Plus One: when a prohibition names a broad word class like counts or figures, disclaimers and narrow scope fail to contain priming, so the prompt must ban only specific artifact strings.", "strategy_quote": null, "reaction_inferred": null, "options": [], "people": [ "Engineer A" ], "confidence": 0.91 }, { "id": "flag-2", "meeting_id": "SPM16-02", "kind": "worth_spreading", "title": "Person Q flags model drafts that drop names, numbers, and payoffs a reader can't decode", "face": "In firsthand writing spot checks, Person Q flagged the same blind spot across email and social drafts: models stay buried in the author's context instead of supplying what a cold reader needs. One email draft named "Claudie, Andy, Viktor, Fin" without explaining them, cited 44%, 85%, 80%, and "62" without baselines, and compressed the payoff into "Climbed a frame, not the framer." Even with her "Reader A lens" skill, Marzipan's pitches stayed "very navel-gazey."", "context": "In Slack, Person Q shared firsthand writing reactions from testing Polaris, Quasar, Wisp, Marzipan, and Grok on Every's writing bench and social-pitch experiments. Alongside her Marzipan social-pitch test, she posted a concrete email draft roundup to show how the model misjudges what a cold reader can understand.", "moment": [ { "line": 23, "speaker": "Person Q", "quote": "the pitches are all very navel-gazey - buried in my own context, not extracting what a reader can learn from or apply" }, { "line": 74, "speaker": "Person Q", "quote": "Passing this email draft along as an example of a problem I'm noticing with the model on the writing front: It doesn't have great judgment on the context a reader needs to make information legible to them. Roundup of things the reader won't understand:" }, { "line": 78, "speaker": "Person Q", "quote": "3. "Climbed a frame, not the framer" — the essay's whole payoff, compressed into a phrase only the author can decode." } ], "why": "Person Q's examples turn "coherence" and "audience awareness" into specific failure modes you can use in tomorrow's Vibe Check and in Checks rubrics for cold-reader legibility.", "strategy_quote": null, "reaction_inferred": null, "options": [], "people": [ "Person Q" ], "confidence": 0.88 }, { "id": "flag-3", "meeting_id": "NPM16-04", "kind": "room_texture", "title": "Likely Person B says users don't care about skills; likely Person D ties them to revenue", "face": "In the Every Agent product review, likely Person D noted that two of three workspaces reject onboarding automation suggestions and proposed turning automations into "skills on a schedule." Likely Person B pushed back that "literally people don't care about skills" and warned against "vanity work" that fakes focus. Likely Person D agreed skill metrics aren't the goal: if skill usage rises while satisfaction, revenue, and memberships don't change, "then that's a failure."", "context": "During the Every Agent product review, likely Person D and likely Person B walked through the activation roadmap and discussed whether connecting automations to a versioned skill library would actually raise the one-in-three onboarding automation acceptance rate.", "moment": [ { "line": 101, "speaker": "likely Person B", "quote": "we also just have incredibly strong usage based data that just say that like literally people don't care about skills. Now, I think that doesn't have any impact on the idea of automations and skills. Skills is just, automations just run a skill on a schedule" }, { "line": 106, "speaker": "likely Person D", "quote": "If we ship a bunch of updates to Skills and then Skill usage goes up because we like, you know, orient around that. But. like overall satisfaction, revenue. Memberships? Don't change. then that's a failure." }, { "line": 107, "speaker": "likely Person D", "quote": "What I do believe though is that skills are a useful primitive, to use your word, to as a delivery mechanism for value that the agent could provide." } ], "why": "This shows how likely Person B and likely Person D are pressure-testing the skills roadmap against real activation and membership conversion rather than treating skill usage as an end in itself.", "strategy_quote": null, "reaction_inferred": null, "options": [], "people": [ "Person B", "Person D" ], "confidence": 0.84 }, { "id": "flag-4", "meeting_id": "NPM16-05", "kind": "room_texture", "title": "UC Davis IT director wants private channels first in a ~30,000-person Slack workspace", "face": "Onboarding into a UC Davis Slack workspace he said has about 30,000 people, likely Person M (engineering IT executive director) saw only public channels in the manual picker: "I would want to add it to private channels first." Even without protected GitHub repos connected, public channels worried him: "People are people and we sometimes say, you know, hey, this student had gotten an F on this assignment that might be in a public Slack."", "context": "In a live beta onboarding call with Person B and likely Person D, UC Davis College of Engineering IT executive director Person M installed the Every Agent into the university's 10-year-old Slack workspace and walked through the manual channel-selection step.", "moment": [ { "line": 82, "speaker": "likely Person M", "quote": "Where at every works, choose how it joins your channels. So, I'm going to do a manual selection. and then select those channels. Um, I see a list of channels I can search. So I'm going to just, I'm going to join a channel Okay, so I'm not seeing... Um, private channels, just they appear to be only public channels?" }, { "line": 84, "speaker": "likely Person M", "quote": "I would want to add it to private channels first. Um, We have about 30,000 people in the Slack workspace." }, { "line": 103, "speaker": "likely Person M", "quote": "People are people and we sometimes say, you know, hey, this student had gotten an F on this assignment that might be in a public Slack. Yeah. discussion because someone doesn't know what they're doing." } ], "why": "Person M's reaction gives you a concrete picture of how an IT champion inside a large, risk-averse institution tests the Agent—starting in a small private peer channel rather than open public channels.", "strategy_quote": null, "reaction_inferred": null, "options": [], "people": [ "Person M", "Person D", "Person B" ], "confidence": 0.81 } ], "reading_notes": [ { "meeting_id": "NPM16-01", "substance": "In the Every Agent launch standup, likely Person H, likely Person E, likely Person I, likely Person D, and Person B reviewed the Figma landing page, discussing how to present automatic compounding without making it feel like work and whether to replace the pricing table and enterprise contact tier with a free-to-start flow.", "selection_reason": "Skipped because the landing-page copy and free-to-start pricing discussion remained open proposals for an afternoon whiteboard session rather than a settled call or a stronger moment than the four selected flags." }, { "meeting_id": "NPM16-02", "substance": "Likely Person D and Person B onboarded beta user Person K, a ghostwriter for tech founders and CEOs who runs a 15–20-person client Slack workspace, accepted a morning call-prep automation for tomorrow, and said using the Agent for client ops could save him from hiring a virtual assistant.", "selection_reason": "Skipped because the UC Davis onboarding call in NPM16-05 surfaced a sharper product constraint around private-channel selection in a large workspace, and Person K's accepted call-prep automation has not yet run." }, { "meeting_id": "NPM16-03", "substance": "Person L and Person E planned next Friday's All Access Office Hours as a non-technical Every Agent beta sneak peek, while discussing hard-to-reach teammates, Person L not finding Person I's Remotion skill by asking the Agent, and building dedicated software tools instead of burning tokens recreating repeatable visuals.", "selection_reason": "Skipped because their discussion of your meeting-attention workflow is already familiar context and their internal enthusiasm for the AI-pilling pitch is not an external customer reaction." }, { "meeting_id": "NPM16-04", "substance": "In the Every Agent product review, likely Person D and likely Person B debated the automation-focused roadmap—where two of three workspaces reject onboarding automation suggestions—alongside landing-page-to-install drop-off, making automations "skills on a schedule," and upfront engineering spec detail.", "selection_reason": "Selected as a room_texture flag because likely Person B's pushback that users do not care about skills and likely Person D's test that rising skill usage without satisfaction, revenue, or membership gains would be a failure expose a concrete product tension beyond the previously surfaced activation rate." }, { "meeting_id": "NPM16-05", "substance": "Likely Person D and Person B onboarded Person M, the Executive Director of IT for UC Davis's College of Engineering, who installed the Every Agent beta into a university Slack workspace he described as having roughly 30,000 people and wanted private-channel-first onboarding before exposing noisy public channels.", "selection_reason": "Selected as a room_texture flag because Person M's reaction to the public-only onboarding channel picker and his concrete example of a student's F grade appearing in public Slack vividly show how risk-averse institutional champions approach workspace installation." }, { "meeting_id": "SPM16-01", "substance": "Person N shared two one-prompt game videos from private model testing—a golf game with a course flyover, ball camera, and ball-stealing pelican, and a game with a sourdough bread sword and $4,200 rent—drawing excited reactions from Person O and Person P.", "selection_reason": "Skipped because the private early-access demos rely on attached videos that were not independently inspected and carry less strategic or operational weight than the four selected flags." }, { "meeting_id": "SPM16-02", "substance": "Person Q shared firsthand writing spot checks across Polaris, Quasar, Wisp, Marzipan, and Grok, documenting cold-reader context failures in an email draft and "navel-gazey" social pitches that persisted even when she supplied her homebrewed "Reader A lens" skill.", "selection_reason": "Selected as a worth_spreading flag because Person Q's concrete three-item roundup of what a cold reader cannot decode and her test of the "Reader A lens" skill give specific, quotable evidence on model writing judgment." }, { "meeting_id": "SPM16-03", "substance": "Following a legal question on maximum data retention, Person R and Engineer A audited retention across Anthropic CMA, R2 backups, OpenRouter/Gemini, inference.net, Intercom, and Browserbase profiles, and Person B decided workspace-deletion retention should be 90 days rather than 14 while the privacy policy is reworked.", "selection_reason": "Skipped because Person B resolved the 90-day versus 14-day retention question in the thread and the remaining subprocessor and Browserbase cleanup items already have assigned owners." }, { "meeting_id": "SPM16-04", "substance": "Person D reported a brief Every Agent dashboard and API outage and initially suspected usergroup-materialization timeouts on a new ~30k–36k-member workspace install before Person J clarified that his environment-variable update for the 90-day deletion period had restarted the server and worker.", "selection_reason": "Skipped because the outage recovered within minutes and Person J identified his environment-variable restart as the likely cause, while Person D is already investigating the large-workspace onboarding timeouts." }, { "meeting_id": "SPM16-05", "substance": "Person J announced that Every is now in Slack's native agent mode with a continuous transcript, suggested prompts, and a hover Stop button, and later clarified for Person I and Person B that the Stop button appears in DMs while typing "stop" cancels runs in channels.", "selection_reason": "Skipped because you already reacted "Sick!!" to the launch in the thread and the follow-up exchange only clarified how Stop works in DMs versus channels." }, { "meeting_id": "SPM16-06", "substance": "Person S created a Checks benchmark from your task dump, explained MCP token setup and Members-page ownership transfer to Person T, and Person T opened Checks PR #30 after you reported a site render issue and asked about reviewing ratings and adding checks.", "selection_reason": "Skipped because you participated directly in the thread and the setup and render follow-ups are routine operational handoffs with PR #30 already opened for review." }, { "meeting_id": "SPM16-07", "substance": "Person T shared his plain-English definition of evals as replaying past attempts, elaborated on your SAT-versus-reference-check analogy as noting red and green flags without reducing hiring to a count, and Person U added his 30/70 reference-check framing before you and Person Q aligned on wanting audience-relative checks.", "selection_reason": "Skipped because you drove both the SAT-versus-reference-check framing and the audience-relative checks request in the channel, making the thread largely familiar context." }, { "meeting_id": "SPM16-08", "substance": "Person V and Person W announced that Studio apps will integrate Every's Kintsugi-backed sales-tax API instead of Stripe Tax, and Person X gave Person P guidance on SaaS tax codes and state registrations before correcting his earlier note to clarify that only main Every history is in Kintsugi so far while app exports remain pending with Person V.", "selection_reason": "Skipped because this is routine finance and tax-integration coordination with clear owners and explicit no-backfill guidance for app teams." }, { "meeting_id": "SPM16-09", "substance": "Person L, Person I, Person E, and Person B coordinated keeping the September 25 All Access Office Hours as an Every Agent beta sneak peek ahead of the reported October 6 launch and planning an October 9 camp, with Person I pushing to include product makers Person B and Person D so top-tier subscribers speak directly to the builders.", "selection_reason": "Skipped because the prelaunch All Access demo concept was already covered in a prior card and Person B agreed to join the September 25 session." }, { "meeting_id": "SPM16-10", "substance": "Person N and Person H reviewed button and card styling for the Every Agent dashboard, agreed on wanting a separate staging environment to test UI changes in context, and settled on sharp corners for both cards and buttons.", "selection_reason": "Skipped because Person H and Person N resolved the corner-styling choice in the thread and the staging-environment idea was already surfaced in a prior card." }, { "meeting_id": "SPM16-11", "substance": "Person G asked whether the "Rebuilding Compound Engineering" story slated for Monday would hold, and Person Q replied that it needs more time given the current Vibe Check load and changes since the initial pitch, leaving Person G's follow-up on bumping it by one week or two unanswered at cutoff.", "selection_reason": "Skipped because whether to bump the editorial slot by one week or two is a routine scheduling call between Person G, Person Q, and Person P rather than a tiebreaker needing your intervention." }, { "meeting_id": "SPM16-12", "substance": "After you noted Quasar triggering goals on simple tasks, Person F remarked that pausing a goal was confusing because task agents kept working, and you noted that goals made more sense when agents stopped early.", "selection_reason": "Skipped because you initiated and concluded the exchange yourself, so a card would mostly repeat a conversation you just had." }, { "meeting_id": "SPM16-13", "substance": "Person U asked whether Thesis ticketing and registration included the sponsor data-sharing opt-in, and Person I replied that it is live for the afterparty while he waits on sponsor names from Person U before launching the rest.", "selection_reason": "Skipped because this is a routine operational handoff between Person I and Person U with a clear next step." }, { "meeting_id": "SPM16-14", "substance": "Person T reported that Person AG at a16z told him he wants to test Every's agent and asked the team for the beta enrollment process.", "selection_reason": "Skipped because it is a brief, unanswered operational question on how to enroll a beta tester rather than a committed experiment or decision waiting on you." }, { "meeting_id": "SPM16-15", "substance": "Following up on his Vercel-inspired idea for department subagents, Person I suggested letting the installer optionally point the Every Agent to department leads during onboarding, and Person D replied that workspace self-introductions could infer contacts from Slack job titles by default with an optional installer override.", "selection_reason": "Skipped because the exchange is an early, agreeable product riff on future onboarding rather than an active dispute or measured learning." }, { "meeting_id": "GPM16-01", "substance": "In EveryInc/checks, Person S closed Person T's PR #19 importing your benchmark via a codebase task—stating benchmarks should be added through database CRUD via MCP—and merged PR #20 so new accounts mint an unbound token and have their AI call create_benchmark in conversation instead of using a web form.", "selection_reason": "Skipped because you were already in the companion Slack thread (SPM16-06) where Person S and Person T walked through MCP benchmark creation and ownership setup for your account." }, { "meeting_id": "GPM16-02", "substance": "Person D's open EveryInc/plus-one PR #1733—approved by Engineer A this afternoon—moves unasked caveats into a Slack context: footer while keeping important unasked data facts in the body, documenting that five descriptive prompt edits scored 0/15 on footer use until a worked example and the per-turn output-policy note were added to both the prompt and the eval harness.", "selection_reason": "Skipped because the core prompt-and-harness measurements in PR #1733 date from September 14–15 with only an lgtm approval added this afternoon, whereas PR #1755 (GPM16-03) carries fresh afternoon commit evidence." }, { "meeting_id": "GPM16-03", "substance": "Engineer A's open EveryInc/plus-one PR #1755 and its afternoon commit 1d42ffd add a one-time silent retry for degraded HTTP 200 external jobs (widening the opening gate from "fails" to "fails or comes back degraded") and document that banning words like "counts" or "tallies" primed the model to drop ordinary numbers in unrelated replies—scoring 0/3 even with an explicit disclaimer.", "selection_reason": "Selected as a worth_spreading flag because the measured finding that banning a word class primes the model to withhold numbers everywhere—and that adding a disclaimer scored worst (0/3)—is a counterintuitive, reusable prompt-engineering lesson backed by today's commit." }, { "meeting_id": "GPM16-04", "substance": "Person P merged EveryInc/baby-agent PR #204 (P64b), aligning external AI employee profiles with his external-agent-employee-profile skill by interviewing other Slack bots in an open thread across eight fields and keeping observed_behavior visibly separate from the agent's self-reported claims.", "selection_reason": "Skipped only due to the four-card batch cap, after prioritizing two live meeting texture moments, Person Q's writing spot checks, and Engineer A's measured prompt-priming regression." }, { "meeting_id": "GPM16-05", "substance": "Person P merged EveryInc/baby-agent PR #201 (P64), allowing Baby Agent to list other Slack bots as external AI employees, hand briefs to them via employees_handoff, park the turn on an ApprovalRequest, and resume the same turn when the message handler captures the other bot's thread reply.", "selection_reason": "Skipped due to the four-card batch cap, though its pattern of parking a turn on an external bot's thread reply and never @mentioning bots outside active briefs is a clean architectural companion to PR #204." } ], "notes": "Slack messages and GitHub PRs have verified account labels (with Claude-assisted PR text noted in metadata), while the five Notion meeting transcripts (NPM16-01 through NPM16-05) are unlabelled multi-speaker transcripts where speakers are attributed as likely from direct-address handoffs, attendee rosters, and role continuity." }