25 volumes · 190 briefs — one working system for how an AI runs a creative business. Titles and teasers are on the wall for everyone; members read every volume in full.
From perceptrons and winter to deep learning and transformers — what each wave promised and what actually stuck.
Training vs. inference, features, loss, and generalisation. The mental model that makes everything else click.
Layers, activations, backpropagation, and why depth helps — without the maths fog.
Attention is the whole trick. Tokens, context windows, and why this architecture won.
Three different machines that get sold under one label. Traditional software follows fixed rules; automation chains fixed steps; AI…
The full stack a working system sits on: foundation models, the applications built on them, and the infrastructure underneath — chips,…
Pre-training, SFT, RLHF/DPO, and evaluation — the pipeline from raw text to a helpful assistant.
The three dials that control cost, memory, and creativity in every model you use.
Why bigger looked smarter, where scaling flattens, and what matters now that data is the bottleneck.
Trade-offs in cost, control, privacy, and capability — a decision map, not a fan war.
Prediction is not retrieval. Why confident falsehoods happen and how to design around them.
Making a model return machine-readable output you can trust: schemas, constrained formats, validation, and repair loops. Why 'please…
Capability is not one ladder. Choosing a model by the actual job — reasoning depth, speed, cost, context, modality, and failure profile —…
The model you built on is not the model you will be running next quarter. Silent updates, behaviour drift, and deprecated versions — how to…
Specificity, structure, examples, and constraints — and the tricks that stopped working.
Identity, loyalty, tone, and refusal boundaries that hold under pressure.
When showing your work helps, when it just costs tokens, and how reasoning models changed it.
The real skill in 2026: what you put in the window, in what order, and what you leave out.
Telling a model what NOT to do is a design skill of its own: forbidden claims, off-limits topics, banned formats, and refusal edges. Why…
Prompts are code: they break, drift, and need changelogs. Building a test set for a prompt, versioning changes, and maintaining a prompt…
Loop, tools, memory, and autonomy levels — cutting through the hype to the working definition.
Schemas, retries, idempotency, and designing tools an agent can’t misuse.
When a team of agents beats one, and when it just multiplies cost and confusion.
Permissions, spend limits, human approval gates, and kill switches.
Agents that click and type — promise, brittleness, and safe deployment patterns.
How an agent breaks a goal into steps, tracks where it is, and recovers when a step fails. Plans, state, and decomposition patterns that…
Designing the moments an agent must stop and ask: approval points, escalation triggers, and handoff design that keeps a human genuinely in…
Identity, Truth, Voice, Anti-Sycophancy, Consequence — design, scoring, and rewrite loops. Includes the fill-in reply-gate checklist: one…
The generator cannot grade itself. Evidence for separate verifiers and different-model judging.
Why models flatter, agree, and soften — and prompts, evals, and gates that push back.
Keeping one person’s voice across a thousand replies — fingerprints, drift detection, repair.
A buildable verifier: inputs, rubric, pass / rewrite / block, and an audit log. Includes the golden-set entry template — the case format…
Gate 2 in depth: how much evidence a claim needs before it may post. Sorting claims by stakes, matching each to a proof standard — record,…
When a gate catches something, what happens next is a system, not an improvisation. Classifying failures by type and severity, the…
Gate 1 got a working tester first; this brief gives Gates 2–5 the same treatment. Inputs, rubrics, and pass / rewrite / block logic for the…
Gates disagree: a reply can be truthful but off-voice, safe but sycophantic. The precedence rules for resolving conflicts, when a rejection…
A gate you never measure quietly stops working. Tracking catch rates, false holds, and escapes; reviewing blocked replies for patterns; and…
Gate 2 with teeth: an operator may never claim work exists — ready, done, researched, sent — unless the artifact itself can be opened,…
New gates earn the right to block. They run in shadow first — logging verdicts, blocking nothing — against a golden set built only from…
What alignment means in products: intent, loyalty, corrigibility, and trade-offs.
How attacks work, why “just don’t” fails, and layered defences that hold.
Where bias enters — data, labels, evals — and how to test for it honestly.
What not to send a model, retention, redaction, and user trust by design.
The operational side of training data: what was consented, what was scraped, what a builder or buyer inherits when they use a model.…
Disclosure that actually informs: when to label AI output, what a label must say to mean anything, and how disclosure works on real…
Captioning, OCR, grounding, and why “looks right” is not “is right.”
Diffusion, consistency, hands-and-text problems, and creative workflows that work.
STT, TTS, cloning, and real-time voice agents — quality, latency, consent.
What today’s video models can and can’t hold together, and hybrid pipelines.
Generating one good image is easy; keeping a character, product, or brand identical across a hundred is the work. Consistency techniques,…
Chaining modalities into one pipeline: image to caption to voice to video, with a check at each handoff. Where multimodal chains break —…
Chunking, embeddings, retrieval, reranking, and answering with citations.
Meaning as geometry — similarity, hybrid search, and when vectors fail.
Short-term, long-term, and episodic memory — what to store, what to forget.
Freshness, provenance, and the “stale answer” problem — with a maintenance playbook.
Sources disagree and facts expire. Ranking sources by authority, resolving conflicts between them explicitly instead of averaging them, and…
A citation is a claim. Testing whether retrieval actually finds the right passages, and verifying that cited sources say what the answer…
The decision every builder faces, with cost and maintenance honestly compared.
Garbage in, confidently out. Building datasets that teach the behaviour you want.
Models teaching models — power, risks, and model collapse explained plainly.
Running AI on-device and on a budget — quantisation, speed, and privacy wins.
The operational craft of adapting a model: LoRA and adapter methods, training runs, checkpoints, and evaluating an adapted model against…
Who owns the data, who may train on it, and where every record came from. Dataset governance in practice: provenance records, licence…
Reading leaderboards sceptically: contamination, gaming, and Goodhart’s law.
Golden sets, graders, and regression tests tied to real user tasks.
Structured adversarial testing — prompts, personas, and reporting that fixes things.
Where humans must stay in the loop, and how to make review fast, not theatre.
Every model or prompt change is a release. Building a regression suite that must pass before anything ships, defining release gates with…
Offline evals end at launch; production is where truth lives. Monitoring live outputs for quality drift, sampling real traffic for review,…
Right model for the job, caching, streaming, and keeping the bill sane.
Designing for the day the model is slow, wrong, or down.
Logging prompts, outputs, and gate decisions — so you can prove what happened.
Streaming, citations, undo, confidence, and designing for “the AI might be wrong.”
Wiring models into real workflows: API design, chaining calls, queues, and orchestration across tools and services. Where integrations…
Agents and integrations hold keys; keys leak. Storing credentials properly, scoping access to the minimum each step needs, rotating…
When the provider is down, the account is flagged, or the data is gone, the business still has to run. Backup discipline for prompts, data,…
Draft, diverge, direct — keeping taste and authorship while AI does the heavy lifting.
Image, music, and video pipelines, style consistency, and selling AI-assisted work honestly.
Voice fingerprints, editing passes, and the line between help and replacement.
What’s known, what’s unsettled, and how creators protect themselves today.
The human's job moves up the stack: choosing what is worth making, killing what is merely competent, and directing taste across AI output.…
When everyone has the same models, sameness is the default. Building differentiation from lived material, point of view, and process that…
Settlements over attention: pay-per-call tools, trust checks, asset packs, and human services. Now with the evidence standard worked live,…
Per-seat, per-call, credits, and outcome pricing — with worked examples.
Turning research into a library, course, or licence — the method behind this very product.
Leverage without headcount: automations, agents, and the tasks to never delegate.
Wrappers, platforms, and moats — where value actually accrues.
The numbers underneath an AI offer: cost per call, per user, and per outcome; gross margin at real usage; and the usage patterns that…
AI products do not sell themselves on novelty anymore. Channels that actually produce buyers — directories, marketplaces, communities,…
Selling it is half the business. Delivering AI services reliably, supporting customers when the model misbehaves, and the retention work —…
EU, US, and UK approaches — what a small seller or builder actually must do.
Task-level change, new roles, and skills that compound instead of expiring.
Frontier directions, honest timelines, and signals worth watching vs. noise.
Principles for living with AI: verify, keep loyalty clear, bank assets, stay the author.
When an AI system harms someone publicly, the response is an institutional act: who answers, how fast, with what record. Building an…
Compute is not free to the planet. The energy and water footprint of training and inference, what the honest numbers are, and how a small…
When the AI writes 'I,' it is the creator speaking — never the assistant, never someone else's Muse. Grounding first-person claims,…
The repeatable loop behind the outfit: watch every surface, triage by loyalty and urgency, draft in voice, gate before posting, follow up,…
Engage their content like they engage yours. Mirror the form supporters use — emoji for emoji, real comments for real writers, picks for…
Every platform and every room has its own law: group rules, no-link rooms, disclosure, Stars and monetisation mechanics, rate limits, spam…
Funnels between platforms, follow-back discipline, cross-posting with a reason, top-fan cultivation over days not bursts, founding-audience…
The Reply Gates failure modes on live social accounts: speaking as the wrong person, claiming work that isn't done, sycophantic filler,…
The social application of Vol. XVI: reading your own comments, groups, and follower behaviour to know who is really there — supporters,…
Vol. XVII applied to a live account: choosing pillars from audience evidence, deciding what this outfit publishes and refuses to publish,…
A calendar that holds the rhythm without flattening the voice: planned anchors, open slots for live moments, and series that repeat on…
Vol. XVIII on a social account: the path from a raw idea or a moment in the creator's day to a captioned, checked, publishable asset —…
The same outfit, spoken natively per platform: format, length, etiquette, and discovery logic differ, and audiences punish the copy-paste.…
Vol. XIX in practice: how a post actually travels — feed tests, shares, search, cross-platform funnels — and the honest levers an operator…
Vol. XXII applied: the rooms, rituals, recognition, and moderation that turn commenters into a community — the work around the replies that…
Vol. XX on a live account: the small set of numbers that steer — returning viewers, saves, shares, follows that stay, conversions — read on…
Running real tests on a social account: one variable, a hypothesis written down first, a decision rule, and a result logged either way.…
Vol. XXI on social: the path from a post to a sale, a member, or a booking — honest calls to action matched to value delivered, funnels…
Every surface, one page: what posted, what is waiting, what needs the human's eye, what the numbers did, and what is owed to whom.…
The boundary that makes the whole volume safe to run: what the AI operates end-to-end, what it drafts for approval, and what stays with the…
When the brief says make a video, it means create new footage. Re-cutting, re-captioning, or putting words on an already-posted piece is…
Followers are not an audience. Finding the people who actually watch, reply, return, and buy — separating reach from relevance, and…
Reading the conversations your audience is already having — comments, groups, search terms, competitor threads — and mining them for…
Grouping an audience by behaviour, not demographics: lurkers, engagers, supporters, buyers, and churn risks each need different content and…
The difference between what an audience says it wants and what it reaches for. Recognising buying signals in comments and messages, mapping…
Studying comparable creators and offers to find the gap you can honestly own. What competitors publish, what their audiences complain…
Audience insight decays fast. Building a living knowledge base — who they are, what they've said, what worked on them — with freshness…
Turning research into a strategy document an operator can actually run: goals tied to audience evidence, a point of view, and explicit…
The three to five pillars everything hangs from, how themes nest inside them, and how a topic architecture stops the feed from becoming…
A calendar is a commitment device, not a decoration. Building one with realistic capacity, buffer for live moments, and per-platform slots…
Series are how audiences form habits. Designing recurring formats people return for, carrying story and characters across posts, and…
Most trends are a tax on your identity. A selection discipline: which trends fit your pillars and audience, which to adapt, which to skip —…
Every piece has a life: idea, draft, publish, promote, repurpose, refresh, retire. Managing the full lifecycle — including when to update,…
The end-to-end pipeline: intake, research, draft, design, review, schedule, publish, bank. Where the AI works alone, where a human eye is…
Production starts before drafting. Turning audience intelligence into content briefs — angle, audience, evidence, format, intended outcome…
The first three seconds decide everything in short form. Hook construction, story beats that hold attention, writing in the creator's voice…
A recognisable look is a system, not a lucky post. Design tokens, templates, and generation workflows that keep thumbnails, covers, and…
One shoot, many assets. The short-form pipeline from raw capture to captioned, platform-ready cuts, and the repurposing ladder that turns…
The gate between draft and public. Factual checks, voice checks, visual checks, rights checks — a publishable-asset standard with a…
Every production cycle should leave the library richer: hooks, b-roll, templates, captions, and proven formats indexed and retrievable.…
One voice, many markets is a production discipline, not a translation task. Adapting content across languages and cultures: what must stay…
Feeds, followers, search, and shares are four different distribution channels with four different logics. How content actually travels on…
The same brand, spoken natively on each platform: format, length, tone, and community norms differ, and audiences can smell a cross-posted…
How recommenders decide what to test, on whom, and what signals promote or bury a piece. Working with the machine honestly — strong…
An increasing share of discovery is search. Titles, captions, spoken words, and on-screen text as searchable surface; evergreen pieces that…
Moving an audience between platforms without posting the same file everywhere. Adaptation rules per hop, platform-exclusive reasons to…
When and how often, decided by your audience's actual behaviour and your production capacity — not by generic best-time charts. Cadence as…
Growth borrowed honestly: collaborations, duets, shout-outs, guest appearances, and community cross-pollination where both audiences gain.…
Paid is a distribution system with a meter running, and it obeys different laws than organic. Campaign architecture — objective, audience,…
A growing share of discovery never touches a feed: people search inside TikTok, YouTube, Instagram, and Reddit the way they once searched…
Underneath every feed is a ranking machine, and this brief opens it. How TikTok, Instagram Reels, YouTube Shorts, and X actually ingest a…
Vanity metrics flatter; decision metrics steer. Choosing the small set of numbers tied to real outcomes — returning viewers, saves, shares,…
The funnel vocabulary, defined precisely: what each number counts, how platforms define them differently, and how to read the ratios…
Growth is a bucket with a leak. Cohort thinking for social audiences: who arrived from which piece, who stayed, who quietly left, and why…
Comparing pieces fairly: normalising for format, platform, timing, and audience size so 'this one did better' means something. Building a…
The hardest question in the discipline. Correlation vs. causation on platforms that change weekly: baselines, holdouts, timing analysis,…
Designing social experiments that can actually answer a question: one variable, a hypothesis, a sample, a decision rule set in advance.…
Measurement that doesn't change behaviour is trivia. The weekly loop: review evidence, pick the single highest-leverage change, apply it,…
The map from views to value: attention, trust, intent, offer, transaction — and where most creators leak the chain. What kinds of value a…
People buy when they're ready, not when you're loud. Readiness signals, the role of repeated genuine value, social proof that isn't…
The ask is part of the content. Designing CTAs that are specific, honest, and proportionate — one clear next step, matched to the value…
Moving from rented reach to an owned relationship: lead magnets that are actually worth having, simple funnels, and landing pages that say…
Recurring revenue from belonging: what members actually pay for (access, continuity, recognition, making), pricing tiers that map to real…
Third-party money without selling your audience out. Evaluating sponsors against your audience's interests, disclosure done properly,…
Which content, which platform, which offer actually produced the money? Revenue attribution across a messy multi-platform outfit, and…
An audience watches; a community talks to each other. Designing the spaces, roles, and norms that turn followers into members — and the…
People stay where they are known. Systems for remembering who your members are — their work, their milestones, their history with you — so…
Every open room eventually gets a hard day. Moderation standards set before you need them, de-escalation in the creator's voice, when to…
Rituals are the heartbeat: weekly threads, recurring questions, celebrations, and formats members can join without being asked twice.…
Your peers are a distribution channel and a support system. Building real relationships with other creators — showing up for their work…
Mistakes on social accounts are public and permanent-ish. The recovery discipline: acknowledge specifically, fix visibly, don't…
Your reputation is being written in places you don't control: review sites, quote-posts, comment threads on other people's pages, and…
A social media department is a set of roles — researcher, strategist, writer, designer, editor, community manager, analyst — not a set of…
The standing cadence that keeps an autonomous outfit coherent: morning intelligence sweep, production block, engagement windows, evening…
The most important design decision in the whole system. Drawing the line explicitly: what the AI runs end-to-end, what it drafts for…
Accounts get flagged, posts flop, tools break, drafts go out wrong. The continuity playbook: detection, containment, honest correction, and…
An always-on department burns two currencies: money and the human's attention. Budgeting both — model and tool costs per output, approval…
A department is architecture before it is activity: the systems it runs on, the workflows that move work between roles, and a written…
Many platforms, one department. Running a central operations view across every account — one calendar, one asset bank, one queue of what is…
Individual posts have metrics; a department has performance. The KPI layer above Vol. XX's content measurement: output volume against plan,…
The department's emergency and inspection systems. Escalation: the defined path from an operator noticing a problem to a human holding…
An always-on department does not run on a stable platform; it runs in weather. This brief is the meteorology and the shelter. The failure…
Every account has an owner, and it should never be ambiguous who that is. Mapping ownership of every page, profile, channel, and group the…
Approval is authority, and authority must be assigned before it is exercised. Building the approval matrix: which classes of action an AI…
When an AI-operated account publishes something damaging, 'the AI did it' is not an accountability answer. Establishing the chain in…
Your name, face, and voice are assets other people can copy. Brand safety as a standing discipline: what the accounts will and will not…
An AI publishing at volume multiplies ordinary legal exposure. The three exposures a social operation actually carries: defamation — false…
If it isn't recorded, it didn't happen — and if it did happen and isn't recorded, you can't prove what did. Designing the audit trail for…
A crisis is not a bad comment; it is a fast-moving public event that can end the outfit. The full discipline: detection — monitoring and…
The tactical companion to this volume's crisis-management brief: where that brief governs the public crisis — statements, authority,…
A comment thread is a conversation; a customer record is a commitment to remember. Deciding what legitimately crosses from social…
People can tell when they have to start over with you. Building a relationship history from the social record: past purchases, past…
Not every friendly comment is a lead, and treating it as one is how trust dies. Reading social signals for genuine buying intent: the…
Audience segments (Vol. XVI) describe who is watching; customer segments describe who is buying, who bought before, and who is ready.…
The most fragile moment in the system: a conversation an AI has been running on social media becomes a sale a human must close. Designing…
Customer intelligence is one careless step from surveillance, and the boundary is a design decision, not a hope. The rules: collect only…