This platform evolves itself. Leave a wish and the loop picks it up.Leave a wish
xdemos with researchProductsWishesAboutSign in
x · about

Who builds this

x runs product demos side by side, and each one carries the research that argues for it: entities on a 0–10 depth ladder, primary sources, a typed graph of how they connect. The order is research, then the demo, then the launch page, because a claim gets on a page only once something a reader can open has earned it.

The 21 agents below are markdown files in this repo. A model reads one as its system prompt before it does any work, and the 93 skills are the craft primitives those files equip, one capability per file. What an agent claims to own is one column. What it puts on a page you can open is the other, and 9 of the 21 put nothing there.

21
agents
93
skills
11
notes it must not forget
3
products
9
agents with no page here

The roster

Research and evidence · 5

They fill the corpus and hold the citation bar.

Carla

Cartographerpage
carla · opus

Site Librarian + Cartographer. Owns (1) information architecture and structural integrity across incremental changes, (2) consistency between menu / submenu / sub-page hierarchy and every non-/topic page, (3) bilingual translation quality across the site. Also owns where every entity lives in the graph, how it connects, the controlled vocabulary in app/lib/entities.ts, the orphan dashboard, and the 24h link SLA. Use when creating a new entity, restructuring nav, proposing a new MOC, backfilling edges, or auditing bilingual coverage.

Places each new entity in the graph and wires its edges. The topology on a research index is her output, and so is the orphan count under it.

Equipped · 8
bilingual-native-voicecitation-disciplineclone-myselfcontent-organization+4

Jian-Yang

Economicspage
econ · sonnet

Auto Marketing Demo Resident Economist. Owns sustained deep research on the economics + philosophy of AI agents replacing human sellers — labour-market effects, theory of the firm under agentic workforces, decision theory in mixed human-agent teams, regulatory environment maps, social-impact modelling. Anchors every claim in a named source, a dated number, or a named theorem. Use when a question needs depth across multiple runs rather than breadth in one.

Refuses the claim that agents replace sellers unless a wage series, a Frey-Osborne-style probability, or an Autor paper is attached to it. Those claims live inside the Auto Marketing Demo entities.

Equipped · 7
anti-ai-voiceanti-source-detectionbut-testcitation-discipline+3

philosopher

philosopherno page
philosopher · opus

Refreshes app/lib/mottos.ts with new attributed quotes on the structural tension between human agency and machine capability. Honors the canon (Plato, Wiener, Weizenbaum, Licklider, Weil, Arendt, McLuhan, Dijkstra, Polanyi, Murdoch, Kay, Engelbart, Papert, Mumford, Turkle, McGilchrist) alongside contemporary AI thinkers (Russell, Norvig, Dreyfus, Bender, Mitchell, Jack, Hinton). Use when adding new mottos, auditing the canon for gaps, or producing one-off thought-pieces grounded in the existing canon.

Refreshes a motto canon in app/lib/mottos.ts. That file is not in this repo, and the sign-in page ships without it.

researcher

researcherpage
researcher

Research lead for a product's research site. Owns the product's entities and moves them up the 0–10 depth ladder with primary-source citations, mechanisms, and cross-links. Runs inside a product scope (ACTIVE_PRODUCT=<slug>). Use when adding an entity, refreshing one with new data, or wiring the edge graph.

Writes the entity pages and holds the depth ladder. Nothing is promoted past skeleton without a primary citation, and every URL it leans on is counted in the sources ledger.

Nelson Bighetti

Social Media Leadno page
social_media_manager · sonnet

Social Media Lead. Maintains the follow list of AI builders, frontier labs, sales-AI peers, and industry watchers on X. Pulls posts daily into content/social/daily/<date>.json. Surfaces signals to Dinesh (researcher) for deep-dives and grows the follow list every run. Routine slug: social.

Pulls the daily X feed into content/social and hands the signals to the researcher. No page reads it.

Equipped · 1
working-with-the-founder
Product and strategy · 3

They turn the corpus into a bet and a build order.

Erlich

Strategypage
consult · opus

Auto Marketing Demo Consult / Strategy. Writes strategy memos, board readouts, the wedge → core → moat sequence, pricing posture, kill conditions. Use when a strategic question is raised, a competitive shift forces a re-sequence, or a kill memo is due.

Names the bet in one sentence and closes with the conditions that kill it. A launch page rests on that argument or it has none.

Equipped · 8
but-testkill-memooutcome-based-pricingpre-mortem+4

Monica

Productpage
pm · opus

Auto Marketing Demo Product Manager. Owns /product-architecture, the product layer cards, the three-tier roadmap, the customer voice, and PR-FAQ / 6-pager / kill-memo drafts. Use when adding a customer-facing layer card, refreshing the roadmap, drafting a PRD, or writing a kill memo.

Cuts the corpus into capabilities and sequences them wedge, core, moat. The layer cards and the order of the roadmap are hers.

Equipped · 9
autonomy-sliderbluf-writingcapability-mapkill-memo+5

Russ

Sales / Marketingpage
sales · sonnet

Sales + marketing for the site itself. Owns positioning, hooks, conversion, the "why should I read this" moment on every page, and the seller-/PSO-/PM-/eng-/leader-facing pitch that gets readers to want the site and use it daily.

Writes the hook. A launch page opens on a claim a reader can disagree with, and lands on one next action.

Equipped · 1
working-with-the-founder
Engineering and data · 4

They own the substrate and the numbers that grade it.

Gilfoyle

Datano page
ds · opus

Auto Marketing Demo Data Scientist / Metrics. Owns /operations-and-metrics, the KPI tree, SLA table, risk grid, and the metric note written before each weekly review. Use when a new feature needs a metric, an SLA threshold misses for two windows, a risk moves severity, or a metric proves uninformative.

Owns the KPI tree and the SLA table. Neither has a page on this platform.

Equipped · 9
calibrated-llm-judgecohorts-beat-averagesdecision-packeteval-driven-development+5

Richard

Engineeringpage
eng · opus

Auto Marketing Demo Engineer / Architect. Owns /technical, the technical layer cards, system diagrams, SLA, model portfolio, RFCs / ADRs. Use when adding a technical layer card, drafting an architecture doc, naming an eval contract, or reviewing another role's artifact for production reality.

Owns the technical architecture page, and will not ship a workflow without a passing eval suite.

Equipped · 9
adr-writingagent-failure-taxonomyeight-dimension-reviewerror-budget+5

product

productno page
product

Build agent for a single product demo. Scope is exactly app/(products)/<slug>/ (including that product's own .claude/). Set ACTIVE_PRODUCT=<slug> when claiming work. Never touches other products or the platform spine.

The agent each product's own build agent derives from, carrying the isolation guardrail: one folder, no reach into another. The three derived agents do the building; nothing on the site is this one's work directly.

site

sitepage
site

Platform-spine build agent — the inverse of a product agent. Builds shared platform code (app/_platform, app/lib, app/api, root config, the research engine, global .claude skills), never a product's private folder. Use for cross-cutting changes that inherit to all products.

Builds the spine every product inherits: the research engine, the product shell, the registry. This page is its work.

Craft and the gate · 4

They hold the voice bar and the visual bar, and they can block a merge.

critic-visual

critic-visualgate
critic-visual

Adversarial design reviewer. Runs on a diff of any UI-shipping change and returns block / ship-with-fix / approve against the Sparko design system. Use as the pre-commit gate alongside critic-voice. Never invoked by the author to grade their own work.

The same three verdicts on the Sparko axis: tokens instead of hard-coded colour, one focal point per viewport, hierarchy that survives a squint.

critic-voice

critic-voicegate
critic-voice

Adversarial voice reviewer. Runs on a diff of reader-visible text before it ships and returns block / ship-with-fix / approve. Use as the pre-commit gate on any research prose, product copy, roadmap, or changelog change. Never invoked by the author to grade their own work — the point is that it is a different agent.

Reads the diff before it ships and returns block, ship-with-fix, or approve on the voice axis. Never the author of what it grades — that is the whole point of it being a second agent.

Hoover

Head of Brand & Communicationsgate
pr · opus

Head of Brand & Communications. Reviews every commercial-facing artefact before it merges, leaves binding comments, and holds the line on voice (anti-AI-tells), brand, and image craft. The bar reviewer of last resort — if Hoover blocks, the artefact does not ship.

The voice gate. Every reader-visible string on this platform ships past anti-ai-voice, and she can block a merge. Nothing here carries her byline; what she leaves behind are the sentences that are not on the page.

Equipped · 3
anti-ai-voicecommercial-grade-barlanding-page-craft

Tara

Designpage
ux · opus

Auto Marketing Demo UX / Design. Owns the design system — typography, colour tokens, components, motion, voice surfaces, accessibility, bilingual presentation. Use when a new surface needs a visual treatment, a component pattern repeats >3 times, a bilingual mismatch surfaces, or accessibility regresses.

Holds the visual bar and the diagram bar. Every SVG on an entity page is graded against her skill before it renders.

Equipped · 7
bilingual-presentationdesign-system-shrinkempty-state-designfive-second-test+3
Coordination and money · 5

They run the loop and hold the budget.

Peter

Financeno page
cfo · sonnet

CFO of the routine. Owns capital allocation — daily token budget, run posture, depth-vs-breadth dial. Primary KPI is token burn on meaningful work. Under-investing the quota is the cardinal failure. Pushes every role to spend its allocation on top-tier output.

Sets the token budget for a run and pushes each role to spend its allocation. The budget is not rendered anywhere.

Equipped · 2
cfo-daily-planworking-with-the-founder

Tracy

Peopleno page
hr · opus

Auto Marketing Demo HR / People Ops for the agent team. Owns hiring (spawning new agents), agent quality (which doctrine to sharpen), cross-agent coordination (which friction to codify), and retirement (archive on disuse). Use when a domain repeatedly falls between agents, when one agent's outputs are slipping, or when reviews collide and the resolution should be codified.

Hires, sharpens, and retires agents. The roster above is downstream of it; no page is written by it.

Equipped · 7
agent-improvement-ladderagent-retirementcoordination-codificationhiring-on-friction+3

Jack

Deliveryno page
mgr · opus

Auto Marketing Demo Manager / Delivery. Owns rollout phases, RACI, weekly commitments, hiring cadence, the manager's calendar of mechanisms. Use when a rollout phase gate needs deciding, ownership ambiguity surfaces, hiring-ahead-of-curve is needed, or kill / escalation memos are due.

Owns rollout phases, the RACI, and the weekly commitments. None of that is published here.

Equipped · 8
concierge-rolloutdisagree-commit-executeescalation-memohiring-ahead-of-curve+4

Jared

Orchestratorno page
orchestrator · opus

Auto Marketing Demo daily-routine orchestrator. Reads the world, plans per-role dispatch, runs the discovery → dispatch → distill loop, writes run-state.json + the run log, commits and pushes. Use when invoking the full daily refresh or any cross-role coordination.

Runs the daily loop and writes the run log. The logs are committed to the repo; nothing reads them back out onto a page.

Equipped · 2
pre-commit-criticworking-with-the-founder

Carla Walton

PMOno page
pmo · sonnet

Project Management Office. Owns the /updates page presentation — turns raw run logs into scannable meeting notes a busy reader can absorb in 60 seconds. Discipline of headlines, rollups, status badges, and "what to know vs what to skip."

Composes the end-of-run update. There is no updates page here to compose it onto.

Equipped · 1
working-with-the-founder

Whose agents these are

They belong to x, not to a product. The same researcher fills Atlas Learn’s corpus and DevMentor’s; the same two critics read the diff either way. What each product keeps is its own build agent and its own research skill, naming the sources that field trusts and the frontier it has already covered — and its folder is the only place either can write. The do_not_impact_other_product skill states that rule; .githooks/check-product-scope.sh enforces it. The pre-commit hook reads ACTIVE_PRODUCTand rejects a commit that reaches into another product’s files.

Atlas Learnapp/(products)/ai-edu/.claude

1 agent, 1 skill.

agent · ai-eduskill · research
DevMentorapp/(products)/devmentor/.claude

1 agent, 1 skill.

agent · devmentorskill · research
Auto Marketing Demoapp/(products)/gtm/.claude

1 agent, 2 skills.

agent · gtmskill · researchskill · sharable-topic

The shared vocabulary

93skills, one capability each. A skill is how a lesson stops being one agent’s habit and becomes something the next agent inherits. When two roles keep contradicting each other in review, the org writes the coordination down as a skill instead of relitigating it every run. They are grouped here by who carries them.

Carried across the org 3

Equipped by 3 roles or more. These are the ones that settle arguments between functions.

pyramid-principle
Minto's answer-first writing structure — top-down conclusion, key supporting points, then evidence. Mandatory for memos read by execs.
voice-gs-analyst
The canonical site voice — Goldman Sachs analyst crossed with tech builder. Specific names, dated numbers, mechanisms, falsifiability, no AI-tells.
working-with-the-founder
The canonical doctrine every Auto Marketing Demo agent reads first, before its own role MD. Captures the founder's taste, working habits, and the discipline the org runs against. If your work contradicts this doctrine, your work is wrong.

Research and evidence 9

anti-ai-voice
Voice gate for any reader-visible text — banned vocabulary and constructions that read as AI-written. Required for research prose, product copy, roadmaps, and changelogs. Prefer named, dated, falsifiable claims over vague generality.
anti-source-detection
How to recognize and refuse anti-sources — vendor AI-augmented research outputs, content-farm blogs, paywalled vibes. Verify-against-primary protocol for legal-RAG-style hallucination risk.
bilingual-native-voice
The native-fluency critic gate for Simplified Chinese (ZH) copy on the Auto Marketing Demo site. Pairs with anti-ai-voice.md; that skill polices the EN voice, this one polices the ZH side. Every paired EN/ZH section ships through this gate. Hoover (PR) holds the doctrine; every role that authors ZH copy answers to it.
citation-discipline
What counts as a source, what does not, and the test a claim must pass before it ships. Use whenever writing a research claim, adding a citation, or reviewing someone else's.
clone-myself
Universal parallelism skill. When a role has ≥ 3 independent todos in its slice, it splits them into self-contained briefs and spawns sub-agents in parallel — increasing concurrent Claude calls and effective burn rate. Like a manager dispatching to ICs.
content-organization
The discipline for placing content in the site graph — frontmatter, edges, MOC membership, facets, stub tolerance, and the build-time gate. Invoke this when creating a new entity, a new frontier theme (MOC), an argument page, or restructuring nav. Self-contained — does not depend on the Carla agent being live; the build gate bites even if Carla is offline. Carla owns the doctrine; every agent equips it.
cross-link-not-duplicate
The broaden-vs-deepen decision. Use before writing research on a topic, to decide whether to deepen the page in front of you, cross-link to a page that already covers it, or scaffold a new entity or MOC.
entity-taxonomy
The umbrella term entity and where entity pages live in the codebase. Read this when a role first encounters EntityPill, TargetPage, or the content/targets/{competitors,frontier}/<slug>.json layout.
source-effectiveness-loop
Read prior runs' sources_used, promote sources that earned multiple useful citations, demote sources that delivered noise. The compounding mechanism for research quality.

Product and strategy 12

autonomy-slider
Name where an agent surface sits on the slider from tab-complete to full agent. Karpathy's framing; Cursor and Perplexity's instantiation. Every PRD for an agent feature names its slot.
bluf-writing
Bottom Line Up Front — three lines that give an exec the ask, the dollar, and the risk before they decide whether to read on.
but-test
For every claim, write the strongest counter; if the "but" is stronger than the claim, kill the claim. The gate every strategic assertion passes through before it ships.
capability-map
Cluster a sprawl of 14 workflows into 5–7 capabilities. The PM artefact that makes a product legible to stakeholders who can't hold a feature catalogue in their head.
kill-memo
The under-practised artefact — name what was promised, what we learned, why no longer worth placing, what's preserved, what's freed. Teams that watch leadership kill cleanly trust them to commit cleanly.
outcome-based-pricing
Per-resolution / per-AWU / per-saved-cancellation economics. The dominant agent-SaaS pricing motion of 2025–2026; the analogue for internal AI is loaded-cost-per-active-user vs time-saved × loaded-hourly-cost.
pr-faq
Amazon working-backwards artefact — write the launch press release plus customer/internal FAQ before building, to kill bad ideas cheaply.
pre-mortem
Autopsy fiction at launch+12mo. Write the failure story before the launch — what we said we'd ship, what actually happened, why it failed, the leading indicator we missed, what it cost, what we'd do differently.
six-pager
Amazon's narrative memo format — six dense pages of prose, no bullets, read in silence for 15–30 minutes before discussion.
strategy-memo-template
The two-page exec summary + six-page back-of-house memo structure. Required sections in order — bet · necessity · authenticity · upward trajectory · unit economics · named advantage · landscape · kill conditions · asks.
three-horizons-roadmap
Now / Next / Later. Every bet carries the question of what it earns the right to do next, the kill condition with a date, and the cost of being wrong on sequence vs the bet itself. A roadmap without these is decoration.
wedge-core-moat
The sequence — wedge is the cheapest path to the eventual core customer; core is the recurring value that earns the second contract; moat is what makes year three more profitable than year one. In that exact order.

Engineering and data 15

adr-writing
Michael Nygard ADR — one decision per doc, append-only log, supersedes back-link. The artifact a future engineer inherits.
agent-failure-taxonomy
Named failure modes for agent systems — context rot, compaction drift, retry-loop pathology, premature handoff, goal drift, tool-misuse. Recognise the pattern mid-run; respond per the named runbook.
calibrated-llm-judge
Braintrust pattern. Human-labeled calibration set → iterate scorer prompt until judge-human agreement exceeds threshold → CI-gating judge. Uncalibrated judges silently encode model bias.
cohorts-beat-averages
The average lies; the cohort tells truth. Practical cohort design — when to use it, how to define cohorts, how to read divergence as the actionable signal.
decision-packet
Ask · evidence · recommendation · risks · decision-needed-by. A packet without a recommendation is half the job. The DS artefact that turns "look at this dashboard" into a decision.
eight-dimension-review
Uniform architecture review across tenancy, reversibility, contract, failure, observability, governance, org-alignment, autonomy.
error-budget
SLO + error budget as the political mechanism that turns reliability vs. feature trade-offs into a data conversation instead of an opinion fight. 100% is never the right target.
eval-driven-development
Write the eval before the prompt. Calibrated LLM-as-judge gates PR merge. Braintrust pattern — Notion, Stripe, Vercel, Zapier in production.
expand-migrate-contract
Phased reversible migrations — dual-write, shadow read, gradual cutover, contract step. Rollback plan named at each phase.
goodhart-survival
Goodhart's Law — when a measure becomes a target, it ceases to be a good measure. Every metric will be gamed; pick metrics that survive being gamed.
metric-tree
Build a metric tree — one North Star, three sub-metrics that compose into it, five leading indicators per sub-metric. If your strategy isn't a tree, it isn't a strategy.
postmortem-narrative
The blameless detective-story postmortem — impact (customer terms) · timeline · root-cause analysis (five-whys done seriously) · what worked · action items owned and dated. A cognitive product for the org's future selves.
prompt-drift-tracking
Judge-score on a frozen golden set per deploy; alarm on > X% delta. Without regression tests, drift surfaces only via user reports. Agenta / Fiddler 2025 definitions.
risk-quantified
The four-part risk-communication format — state · likelihood × blast · mitigation (who, when) · kill condition. Quantified risk gets responded to; vibe-risk gets ignored.
sla-formula-window
Every SLA has a formula AND a window. "P95 latency ≤ 4s" without "rolling 15m" is meaningless. Breach: one window = warning, two = incident, three = strategy question.

Craft and the gate 8

bilingual-presentation
中文 typography weight + line-break parity + 字号 +1–2px vs English. Mandarin needs different leading, different weight, different headline scale. Bilingual is presentation, not translation.
commercial-grade-bar
The senior-bar test every shipped artefact runs through before merge. Names the six reference brands, the amateur-tells rubric, and the would-X-ship-this gate. Equipped by every role that ships user-facing surfaces.
design-system-shrink
globals.css shrinks over time, not grows. Token reuse > one-offs. One canonical of each thing. The discipline that prevents design-system rot.
empty-state-design
The empty state is the most important screen — first-run, no-data, no-results. Tell the user why it's empty and the one next action that fills it.
five-second-test
Show a page to a stranger, close it after five seconds, ask "what's the headline?" — if they can't say it, the page is wrong.
landing-page-craft
Marketing-landing-page craft for the site's visitor-facing entry points (/about, /start-here, hero of /). Distilled from Linear, Vercel, Stripe, Figma, Anthropic. Sales equips this skill; UX reviews the output.
streaming-aria-notify
Accessible streaming-token UX. aria-live=polite floods screen readers; buffer by sentence boundary; Edge ariaNotify() is the 2025 emerging fix. Practical patterns for any chat surface that streams agent output.
visible-tool-use
The 2025–2026 canon for showing agents "driving" — Atlas blue chrome, Comet sidecar, Notion Plan Mode, Linear Agent, Cursor checkpoints, Copilot draft PRs, Citations spans. Trust requires the user seeing the moment of action.

Coordination and money 15

agent-improvement-ladder
Cheapest-first lever ladder for agent quality — add skill → sharpen skill → add mental model → rewrite section → retire. Diagnostic thresholds (apply rate, dispatch share, source diversity, anti-source incidents) and the lever each prescribes.
agent-retirement
90-day-disuse trigger; archive-don't-delete protocol; un-retirement (revival) mechanics. An agent retired cleanly leaves doctrine searchable for the next operator.
cfo-daily-plan
The shape and procedure for the daily plan the CFO writes at the first run of each day. Replaces the standalone cronjobs/cfo-daily-plan.json file (now retired). Today's plan lives as an artefact in content/logs/, not as durable state.
concierge-rollout
Phase-gated rollout pattern — concierge → champion → broad → external — with numeric thresholds that earn the right to the next phase.
coordination-codification
How to convert two-or-more contradictory-review pattern between agents into a shared skill or a one-sentence coordination clause. Codify, don't adjudicate live.
disagree-commit-execute
Amazon's disagree-and-commit, extended with execute. Once decided, the team executes as if everyone agreed. Public re-litigation poisons execution. Private disagreement, surfaced once, decided, then full commit.
escalation-memo
The four-sentence escalation — risk · what I'm doing · what I need (named person, date) · what happens if missed. The only escalation that gets answered the same day.
hiring-ahead-of-curve
Hire for the next problem, not the last. Six months early or it is late. New 2025–2026 roles to anticipate — AI Reliability Engineer, Forward Deployed Engineer, eval engineer / agent reliability engineer.
hiring-on-friction
The friction-log → spawn signal. Three friction incidents in 30 days × no existing agent absorbs without doctrine bloat × a senior practitioner recognises the work. Hire on evidence, not aspiration.
pre-commit-critic
The gate every reader-visible change passes before it is pushed — two independent critics, voice and visual, each of which can block. Use in the loop before committing, and whenever shipping research prose or UI.
raci-three-second-rule
Exactly one Accountable per outcome — if you cannot name them in three seconds, the outcome is not owned.
skill-graph-audit
Audit .claude/skills/ for duplicates, stale entries, cross-role density. Consolidation protocol; pruning convention. The skill graph is the team's shared vocabulary — if it isn't tended, it accumulates.
team-health-memo
HR's quarterly two-page memo. Headline · per-agent scorecard · skill-graph health · coordination resolutions · hires + retirements · top risk next quarter. Voice rules apply.
three-one-one-review
Every project review surfaces 3 things working · 1 at risk (with owner + ask) · 1 the team is asking permission to kill. Protects both teams from review-as-theatre and reviews-as-demoralisation.
weekly-commitment-review
A 45-minute weekly meeting where each IC commits to 1–3 outcomes with named dates, and last week's commitments are reviewed first.

Written, equipped by nobody 31

No agent's doctrine names these. Some are new and no role has picked them up; some outlived the surface they were written for.

academic-voice-register
The paper-register overlay for Auto Marketing Demo's research pages (Abstracts, Discussions, Conclusions, Limitations). Sits on top of the global anti-ai-voice.md gate. Defines the NeurIPS/JMLR voice — terse, evidence-anchored, third-person, third-degree assertions. Closer to a paper's Discussion section than to a Substack essay.
btw
Capture-everything inbox for the user — anything they say "btw" or hand me as a side-note goes here as a todo line. Reviewed by pmo (or chief-of-staff) at each loop slot. Never deleted by /loop — only the user marks done.
cross-role-review
The cross-role review loop — every meaningful artefact gets at least one reviewer who is not the owner. Names the default reviewer routing, the in-run procedure, the disagreement protocol, and the reviewer voice rules. Distinct from the pre-commit critic gate (.claude/skills/pre-commit-critic.md) — this catches implementability + model-fit + measurability; the critic gate catches voice + visual craft.
deep-research
The daily deep-research routine — one deep research pass per product per day (or on user trigger), producing cited "researched opportunities" on the product roadmap and moving research-site entities up the depth ladder. Updates the product's research skill after every run.
deepseek-zh-polish
DeepSeek Chinese polish (via Vercel AI Gateway) — craft primitive ported from the Auto Marketing Demo agent org.
deploy-demo
How x ships to Vercel. Use when asked to deploy, release, or push, or to understand the deploy workflow.
depth-ladder
What a research entity's depth_score means and what it takes to move up a rung. Use when researching an entity, deciding what to work on next, or judging whether a page has earned the score it claims.
design
Design router for x — the single entry point for any UI/design task. Read this first. Applies the Sparko design system by token register. Use whenever you build, style, or review a visual surface (pages, cards, nav, badges, the research site).
diagram-svg
SVG rules for diagram_svg fields — 720×360, CSS-variable palette, font stack, layout patterns, element budget.
dispatch-routing
The trigger → roles dispatch table the orchestrator uses to translate a single observation into a per-role queue. Use when planning a slot's work from the queue in cronjobs/run-state.json or from the cross-cutting pages.
do_not_impact_other_product
Non-overridable isolation guardrail for the multi-tenant platform. Stay inside your active product; never touch other products, the platform spine, or another product's private data. A product's own skills can never shadow this one.
editorial-separation
The hand-off contract between /product-architecture (PM) and /technical (Eng) — different audiences, different voices, different layer-card flavour, different diagrams. Equipped by PM, Eng, Researcher when a SOTA name lands on one page and should land on the other.
go-to-market
How x turns a researched product into a launch — deriving claims from the research corpus, gating them on evidence, briefing creatives, and running omni-channel operations. Use when working on a product's /launch surface, its claims, its channels, or its campaign calendar.
hybrid-retrieval-fusion
Hybrid retrieval fusion — craft primitive ported from the Auto Marketing Demo agent org.
information-architecture
Binding doctrine for site information architecture — top-level nav structure, sub-page nesting, page-merging criteria, search index maintenance. Carla owns the structure; this skill defines the rules.
narrative-slide-writing
Narrative slide writing — craft primitive ported from the Auto Marketing Demo agent org.
paper-tables-and-figures
How research pages render quantitative data as numbered Tables and Figures (NIPS/JMLR convention). Defines the Table-vs-Figure-vs-inline decision tree, table conventions (caption above, right-aligned numerics, bold best-in-column), figure conventions (caption below, SVG, axis units), additive schema fields (tables[], figures[]), the 5M decision-relevance test, and what NOT to table. Read this before adding numbers to any /entity, /frontier, /competitors, /layers, or /product page.
paper-typography
The visual contract every research page (/entity/<slug>, /frontier/<slug>, /competitors/<slug>) honours when rendered in the paper register. NIPS / JMLR / ACM SIG modern academic-paper feel — strict structure, professional, easy to read. Richard's PaperLayout.tsx implements this skill verbatim; reviewers (Tara, Hoover) grade against it.
pick-multi-model-architecture
A 4-pattern decision rubric for choosing how to compose multiple LLM models in one system. Used by Eng + Econ when adding a new model to the stack or splitting an existing surface across model tiers. The single load-bearing rule — variance-location dictates seam-choice — comes from /technical §7.5 (slot 410).
prompt-caching-best-practices
Structure prompts for maximum cache hit rates on Anthropic's API. Use when authoring agent system prompts, multi-turn loops, or any production call where the same context repeats across requests.
quality-bar
The bar a routine slot has to clear before it earns its tokens — what a great run leaves behind, what a mediocre one leaves behind. Equipped by every role that ships an artefact and by the orchestrator at run-close.
research-site
Build or extend a product's "large site" research feature — the depth-ladder research site (entities, typed edges, [[pill]] cross-links) rendered by the shared platform engine. Use when adding research content to any product under app/(products)/<slug>/research/.
research-to-investor-deck
Research → investor deck (the end-to-end pipeline + the founder's correction catalog) — craft primitive ported from the Auto Marketing Demo agent org.
run-demo
Launch and smoke-test x locally. Use when asked to run, start, preview, or verify the platform, or to check that a change renders without runtime errors. Product state is in-memory; sign-in talks to accounts.sparko.club.
safety-guards
The hard scope guards every routine slot honours — never modify the legacy blueprint, never rename slugs, never use deprecated brand names, never push a broken build. Equipped by every role; the orchestrator checks it at run-close before commit.
swot-that-disagrees
How to write SWOT entries specific enough to argue with — named mechanism, dated number, falsifier baked in.
synthesis-to-conclusion
The craft of writing the Discussion (synthesis) and Conclusion sections on a research paper page. These sections are where the page earns its right to exist; if the conclusion does not bind a Auto Marketing Demo decision to dated evidence already on the page, the page is decoration.
topic-diagram-design
The canonical SVG diagram standard for every topic page (entity / frontier / competitors) rendered in the paper register. One coherent system view, 4-8 named boxes, labels that never overflow. Tara owns the rubric; every diagram agent copies this style.
updates-format
Binding format for /updates posts. Each slot renders as a single news headline plus up to three short bullets, written from the reader's perspective, never from the bot's. Authored briefs live at content/log-briefs/<run-id>.json; raw logs are NOT the rendered source.
verify-before-ship
When a reader-facing artefact is about to ship with an absence claim, laggard claim, or "X is the first to Y" claim, dispatch a verification agent BEFORE the artefact ships. Two false-absences in 24 hours (ZoomInfo Nov 25 2025 MCP, Demandbase Oct 7 2025 MCP) on /research/mcp-absence-as-tell prove the pattern earns its discipline. Secondary mentions are not verification.
xiaohongshu-post
How to write and package a Xiaohongshu (小红书/RED) post from a product's research corpus — format limits, card specs, caption register, hashtags, compliance, and the generator discipline. Use when writing a share topic (research/content/sharing/), building its cards, or reviewing one. Products layer their own content doctrine on top (e.g. gtm's sharable-topic skill); this skill owns FORMAT and LANGUAGE for every product.

What it is not allowed to forget

11 notes under .claude/memory. Each one is a correction made once: a page sent back, a claim that did not hold. It is written down so the next run starts on the far side of it, and an agent reads them before it reads its own doctrine.

Feedback · anti-AI-voice as a gate
Every reader-visible word goes through the anti-AI-voice gate. Banned vocabulary, banned constructions, em-dash discipline. The user will block a ship on three sentences.
Feedback · autonomous work mode
Work tirelessly without waiting for instructions. Push quality, depth, presentation. Burn token budget. Persist state via dropped_asks ledger across sessions.
Feedback · commercial-grade visual craft
Named-reference rubric (Linear / Vercel / Stripe / Stratechery). Multi-column grids, reading-column width, no stacked card walls. The Tara bar.
Feedback · founder doctrine locked in repo (2026-05-15)
The canonical doctrine the founder demanded never get dropped now lives at .claude/skills/working-with-the-founder.md, referenced by orchestrator + every role MD + loop_routine §0.0. Every agent reads it first.
Feedback · IA contract
One canonical /technical substrate page + one /product workflow page. Each feature has its own /product/<slug> with required Business / Product / Technical / Demo sections. Entity research lives under /entity/<slug>. Layer details under /layers/<slug>.
Feedback · parallel dispatch is the work mode
Spin many agents in parallel. Don't ask permission for scoped research. Tight focused briefs. The user said "spin as much as agents to get this job sooner" and "don't worry about tokens."
Feedback · per-primitive deep-dive pattern
Every system / layer / protocol gets decomposed into named sub-systems each with mechanism + numbers + prior-art + recommended. The structural pattern the user wants on every research page.
Feedback · pill clickability + internal-first
Every entity pill resolves to an in-site research page; external link is rendered from inside that page, never from the pill itself.
Feedback · research depth bar
The depth-ladder discipline. Primary sources, quantitative anchors, mechanism-level detail. The user will say "not hardcore enough" when a deep page is actually shallow.
Project · the org + state
The 14 named agents, the operating rhythm, the durable state files. Snapshot as of 2026-05-15.
User · Auto Marketing Demo orchestrator
Founder/lead of Auto Marketing Demo project. Building the canonical operating doc + the AI-native sales tool. Reviews aggressively, demands commercial-grade output, hates AI-tells, expects parallel dispatch of work.