How the Brief and Counsel Are Built

Daily Brief — Free

A personalized selection of 7 high-signal articles, generated on the first page load of each day and frozen until midnight in your timezone. The goal is a clean, 5-minute morning read — not another feed to scroll.

1. Inputs from your profile

The brief is personalized entirely from what you've told us:

• Role — CTO, ML Engineer, PM, Researcher, etc.
• Topics — the taxonomy nodes you've added, with OR/AND logic
• Filters — proficiency, abstraction level, media type, intent, audiences, fields, content types
• Sources and authors — preferred, excluded, or discovered
• Timezone — controls the "midnight" boundary for brief refresh

2. Candidate pool — progressive time windows

We query for candidates in three windows, stopping at the first non-empty result. This guarantees your brief is never empty, even if a given day was quiet:

• Last 48 hours — fresh articles (standard)
• Last 7 days — fallback if the 48h window is empty
• No cutoff — cold-start, best available

If all three windows return zero matches for your exact filters, we fall back to articles for your stated role only — so a new user who hasn't tuned anything still gets a useful brief.

3. Importance score — the 30/25/20/25 formula

Every candidate article gets a 0–1 importance score computed at query time:

• 30% — Source quality. Tier A (1.0) / B (0.75) / C (0.5) / D (0.25), from our curated source registry
• 25% — Intent. Research (1.0) / Analysis (0.85) / Tutorial (0.7) / News (0.6) / Opinion (0.5)
• 20% — Length. Extended (1.0) / Long (0.85) / Medium (0.7) / Short (0.55) / Quick (0.4)
• 25% — Recency. ≤1 day (1.0) / ≤7 days (0.85) / ≤30 days (0.65) / older (0.45)

A tier-A Research paper from arXiv published yesterday lands around 0.94. An Opinion piece from a mid-tier blog last week sits around 0.56.

4. Composite ranking — freshness first, quality breaks ties

The top candidates are re-ranked with a continuous composite that keeps breaking news prominent while letting quality win ties:

• 55% — recency (exponential decay, 1.0 → 0 over a few days)
• 45% — importance score from step 3

The top 7 (configurable 5 / 7 / 10) become your brief.

5. Signal meter and velocity headline

• Signal meter — the percentage of today's picks with importance score ≥ 0.65. Tells you at a glance whether today is a "big day" or a quiet one.
• Velocity headline — if one of your followed topics has ≥3 articles this week and ≥50% growth over last week, you see a line like "3 topics you track surged this week: RAG (+340%), GPT-4o (+120%)."

6. Frozen for the day

Once generated, today's brief is stored and served unchanged until midnight. Same 7 articles every time you open the app — so "Mark today's brief complete" actually means something, and your streak reflects the habit, not the refresh count.

Counsel — Pro

Counsel turns the same source infrastructure into a relevance-ranked synthesis, calibrated to the AI decisions you're actually making. Everything the Daily Brief doesn't know about you — the 2–3 calls on your plate right now — is what makes Counsel different.

1. Conversational context builder

Setup is a 5-minute guided conversation, not a blank textarea. You pick a decision domain (or skip and type directly), then a Gemini-backed conversation asks 2–3 targeted questions to sharpen the context. Seven starting domains cover the common decision surfaces:

• Dev productivity & staffing — coding-assistant strategy, headcount, team structure
• Models & infrastructure — LLM provider selection, cloud vs self-host, vendor lock-in
• AI in our product — RAG production readiness, agents, evaluation pipelines
• Data & training — pipelines, fine-tuning, knowledge bases
• AI in our operations — process automation, customer support, pilots
• Risk & compliance — GDPR, prompt injection, EU AI Act, exit paths
• Business impact & strategy — EBITDA impact, board case, 18-month strategy

2. Structured context profile

Your freetext gets parsed by Gemini into a structured profile stored on your account. Each active decision and capability track also gets a 384-dimensional embedding (BAAI/bge-small-en-v1.5), which is what powers semantic matching against articles:

• Active decisions — description, type, deadline, signal keywords, embedding
• Operating constraints — regulatory, budget, team size, compliance
• Capability tracks — strategic areas you're tracking alongside your active decisions, with interest level
• Current stack — inference, vector store, frameworks, cloud, what you're evaluating
• Already-reads — sources you already follow, so we don't re-surface them

3. Contextual reranking — the top-50 pool

We fetch up to 500 candidate articles from the last 90 days using OR logic across your context, then narrow them to a working pool of about 50 using three signals:

• Embedding similarity — cosine similarity between the article and each active decision, capability track, and the verbatim Q&A answers you gave during setup
• Keyword overlap — matches on auto-extracted signal keywords pulled from your decisions
• Feedback bias — per-user: articles semantically close to ones you rated Useful get a small lift; ones near articles you rated Irrelevant get a small headwind (six independent learning layers — see step 7)

That top-50 is then stratified before the LLM judge sees it: each of your active decisions, each Q&A axis you answered, and a tier-A floor each get a guaranteed slice of the 15 spots that go to the judge. Without that, a dominant decision or a chunk-lucky low-tier blog can monopolize the window and crowd out the rest.

4. The relevance judge — two axes, per article

Cosine retrieval rewards topical overlap, not actionability. So before any prose gets written, a Gemini-flash judge reads each of the 15 candidates against your context and grades them on two independent axes:

• Substantive relevance (0–3) — does this article carry a developed answer or transferable lesson for one of your decisions? 3 = direct, 2 = strong adjacent lesson, 1 = topical-only (vocabulary overlap without substance), 0 = tangential. Anything below 2 is dropped before slot selection.
• Decision impact (0–3) — independent of relevance. A perfectly relevant article can still be a well-developed restatement of consensus advice ("standardize on one platform", "test before you ship") that wouldn't change how you'd act tomorrow. 3 = specific numbers/thresholds/named warnings you'd adjust around; 2 = concrete but known terrain; 1 = "nice to know" but no adoption lever; 0 = trivial.

The impact axis is informed by the article's extractive insights — the essence, principles, method, and practical steps that our ingestion pipeline pulls out at index time. A rich practical list and named-entity principles are the fingerprint of a high-leverage piece; empty or near-empty practical with generic principles is the fingerprint of a consensus restatement. The judge sees both, plus the two matched passages and its previous verdicts on articles you've rated — so it calibrates "direct" to your personal threshold, not an abstract definition.

5. The cross-pool impact ranker

The per-article judge is honest but can't see overlap. Two articles can both score (3, impact 3) yet say the same thing — and reading both wastes a slot. A second Gemini pass takes the qualified survivors and orders them by which one, if you read it now, would most change how you'd act on the decisions you haven't already resolved through earlier items. Hard rules in that prompt: down-rank confirmation of things widely known, down-rank generic principles without concrete adoption levers, down-rank insights that overlap an already-higher-ranked article.

6. Two-pass architecture: deterministic selection, LLM prose

Selection and writing stay separated so Counsel is stable and debuggable:

• Selection — picks are deterministic. Survivors are sorted by (substantive relevance, impact, source tier, cosine), the cross-pool ranker re-permutes the top of that list, and then structural caps — tier-A floor, max per Q&A axis, max constraint-only items, low-tier ceiling — apply to fill the 5 surfaced slots and the 5-item "More in your areas" reserve. Same inputs → same output.
• Prose — a separate Gemini call (temperature 0.3) writes the title and body for each already-selected item. The selection never changes mid-write. Every claim has to substring-match a passage from the cited article or it gets dropped from the body (grounding verification).

7. What the order means

Counsel is a single ranked list, not a tabbed feed. The order is the verdict:

• Top items — relevance and impact lined up: the article's concrete steps or named thresholds map onto a decision you're actively working on, framed directly against your context (e.g. "Given you're weighing build-vs-buy on your agent stack, this vendor pricing change affects your Q2 calculus.")
• Down the list — adjacent or supporting context that still cleared both the relevance and impact bars: papers and analysis worth 20 minutes from the right person on your team
• QUIET — decisions and tracks where the corpus is silent. We scan every embedded article in the 90-day public corpus against each of your area embeddings; if the maximum cosine similarity stays below 0.45, the area is named in QUIET. The section reports the number of articles scanned so the claim is measurable, not rhetorical. For a leader weighing five decisions, knowing three are quiet this week changes where you spend attention.

8. Feedback loop — six learning layers

Each Counsel item has four quick ratings (right_to_the_point / useful / already_knew / irrelevant), plus a click signal when you open the article. Ratings and clicks feed six independent learning layers — each tuned conservatively so a single rating never warps your future ranking:

• Layer 0 — query offset: shifts your decision embeddings toward articles you accepted, away from articles you rejected (±0.05 cosine).
• Layer 1 — per-URL semantic bias: articles semantically close to ones you rated get the rating's bonus or penalty (±0.08, decision-scoped, time-decayed with 21-day half-life).
• Layer 2 — concept aggregates: your taste profile per topic / source / signal-type. Future articles from domains you favoured get small boosts (±0.015 each).
• Layer 3 — cross-user (source × topic) prior: veteran-user taste seeds new users — articles from source-topic pairs other users consistently like get a small lift, even before you've rated anything.
• Layer 4 — article CTR: heavily-shown-never-clicked articles get a small penalty across all users — guards against repeat-show bias.
• Layer 5 — judge calibration: your accepted/rejected examples become few-shot calibration inside the relevance-judge prompt, so the judge calibrates "direct vs adjacent" to your personal threshold.

All six layers are wired and active by default. The Counsel page surfaces a small "calibrating from N ratings · M clicks" subhead once you have enough feedback for personalization to matter — your way of seeing the loop you fuel.

9. Cadence — on-demand plus a conditional morning regen

Counsel is primarily user-pull: you open it when you need it, and a "N new matches — Refresh" badge fires when fresh evidence has accumulated since your last generation. On top of that, a scheduled job at 05:30 UTC re-evaluates each active user's Counsel: if at least three new high-relevance articles have landed since the prior generation, we regenerate and email you the new edition; otherwise we send a short "quiet today" note. This honors "every morning" without burning a Gemini synthesis run on essentially the same Counsel. You can mute the morning email from your Newsletter settings.

10. Ask Counsel — one-shot grounded Q&A

A second Counsel surface for questions that don't fit a decision frame. You type a free-text question (e.g. "What does the evidence say about open-source vs proprietary LLMs for a 30-person team?"); we embed it, retrieve up to 12 top article candidates within a 180-day window, pull their highest-similarity chunks, and ask Gemini to produce a grounded prose answer with verbatim passage citations and a confidence band (high / medium / low). Different from your decision-grounded Counsel: no stored context, no ranking-against-decisions, single one-shot answer. Quota: 15 questions per month on Solo, unlimited on Team.

11. Refine, don't restart

When a decision closes or a new one opens, you don't redo onboarding. Click Refine, add a freetext update, and the conversation picks up with your existing context already hydrated.

At a glance

Related

• How we select and curate the 500+ sources — source-level editorial framework
• Pricing — Free vs Pro vs Team
• Back to the homepage