MCP HubMCP Hub
SKILL·BB0A86

intelligence-collection-disciplines

deanpeters
Mis à jour 6 days ago
1 vues
5,952
730
5,952
Voir sur GitHub
Autreai

À propos

Cette compétence fournit un cadre de renseignement structuré avec huit disciplines de collecte (comme l'OSINT et le SIGINT) pour mener des recherches concurrentielles multi-sources. Elle enseigne les chaînes de signal à inférence et la fusion de preuves pour transformer des données brutes en conclusions solides et défendables. Utilisez-la lorsque vous avez besoin de valider des mouvements de marché ou des stratégies concurrentielles au-delà de l'analyse à source unique.

Installation rapide

Claude Code

Recommandé
Principal
npx skills add deanpeters/Product-Manager-Skills -a claude-code
Commande PluginAlternatif
/plugin add https://github.com/deanpeters/Product-Manager-Skills
Git CloneAlternatif
git clone https://github.com/deanpeters/Product-Manager-Skills.git ~/.claude/skills/intelligence-collection-disciplines

Copiez et collez cette commande dans Claude Code pour installer cette compétence

Documentation

Intelligence Collection Disciplines

Purpose

Stop doing competitive research like a term paper. The intelligence community solved this problem decades ago: they don't collect "data," they run collection disciplines — independent channels, each with its own sources, tradecraft, and blind spots — and then they fuse them. This skill gives product managers that playbook: eight disciplines, each telling you (1) what to collect, (2) where to collect it, (3) which signal → inference chains to run, and (4) which PM artifact it feeds (TAM/SAM/SOM, ICPs, personas, battle cards, roadmap bets, win/loss, positioning, pricing). One signal is an anecdote; three correlated signals from independent disciplines is intelligence.

Input

Works best with the six instantiation variables below — fill them once and every discipline becomes engagement-specific. Leave them blank and this stays a teaching artifact.

[TARGET]      = The competitor, partner, or acquirer you're researching
[MARKET]      = The category and its NAICS/SIC/NACE codes (or nearest equivalent)
[GEOGRAPHY]   = Regions in scope, at COUNTRY level, not continent level
[BUYER]       = Who signs the check (drives review sites, job titles, conferences)
[CAPABILITY]  = The strategic move you suspect (platform play, market entry,
                pricing shift, compliance land-grab, etc.)
[DECISION]    = What this research will change (roadmap bet, positioning,
                pricing, market entry, deal defense, ICP refresh)

If [DECISION] is blank, stop. Research without a decision is a hobby.

Anything supplied with the invocation itself — text after the skill name, a pasted context dump, or an appended ARGUMENTS: line — counts as variables already filled. Use it; don't re-ask.

Arriving empty-handed? That works too. Read the disciplines as a course in tradecraft, or ask the agent to walk the instantiation block with you one variable at a time.

Example invocation: Use intelligence-collection-disciplines — [TARGET]: Meridian Freight Systems, [CAPABILITY]: suspected move into warehouse robotics, [DECISION]: whether we accelerate our own integration roadmap.

Key Concepts

The Eight Disciplines

DisciplineIntel Community NamePlain EnglishPrimary PM Artifact
1. OSINTOpen Source IntelligencePress, social, periodicals, analystsBattle cards, positioning
2. FININTFinancial IntelligenceFilings, earnings calls, procurementBattle cards, SOM capture rates
3. GEOINT/DEMOINTGeospatial & Demographic IntelligenceCensus, labor, trade, economic statisticsTAM/SAM/SOM, ICPs, personas
4. TECHINTTechnical IntelligencePatents, technographics, changelogsRoadmap bets
5. HUMINTHuman IntelligenceTalent moves, employee chatter, win/lossRoadmap bets, battle cards
6. SIGINTSignals IntelligenceWeb diffs, pricing changes, job postsBattle cards, pricing strategy
7. MASINTMeasurement & SignatureSupply chain, ops indicatorsThreat assessment
8. All-Source Fusionn/aCross-validation and confidence stackingEverything above
  • Independence is the design. The disciplines matter because they fail differently: a press release can lie, but a customs record, a patent filing, and thirty job posts are three separate bureaucracies that would all have to be lying in the same direction.
  • Signal → inference chains are the tradecraft: each observable signal maps to a bounded interpretation, labeled per the autonomous-investigation protocol (Fact / Inference / Assumption).
  • What this is NOT: espionage. Every source below is published, filed, posted, or publicly observable. See Guardrails in Common Pitfalls.

Application

1. OSINT — The Journalist's Desk

What a good beat reporter knows before the press release drops.

Source TypeFreePaid
[TARGET] press & newsroomCompany newsroom pages, Google Alerts, PR Newswire feedsMeltwater, Cision
Industry periodicalsTrade publications, association newsletters, vertical Substacks for [MARKET]Analyst subscriptions
Analyst coverageGartner/Forrester press summaries, free webinar replaysGartner, Forrester, IDC full reports
Social & communityLinkedIn exec posts, Reddit, Hacker News, XBrandwatch, Sprout listening
Review sitesWhichever your [BUYER] reads: G2, Capterra, TrustRadius, app stores, TrustpilotG2 Buyer Intent data
Conference footprintSession titles, sponsor tiers, booth size, speaker rosters at [MARKET] eventsn/a
Prediction marketsPolymarket, Kalshi, Metaculus, Manifold (crowd-priced odds on regulation, approvals, tech milestones relevant to [MARKET])n/a

Signal → inference chains:

  • [TARGET] exec suddenly posting about a new problem space → positioning pivot incoming (execs test messaging on social 3-6 months before launch)
  • Sponsor tier jump at a [MARKET] conference → market entry or doubling down
  • Review complaints clustering on one feature → their roadmap pressure point = your battle card ammo
  • Analyst briefing requests (visible via analyst posts) → category creation attempt
  • Webinar topics shifting → what they're teaching the market is what they're about to sell
  • Sudden silence on a product line → sunset in progress
  • Prediction-market odds moving on a regulation or milestone that gates [MARKET] → crowd-priced expectations for scenario planning; treat as a leading indicator of consensus, not ground truth, and check liquidity before trusting the number

Feeds: battle cards (objection handling from review mining), positioning (their words vs. customer words gap), win/loss context.

2. FININT — The Forensic Accountant

Follow the money. Companies lie in press releases. They lie less in filings, because lying there is a felony.

Source TypeFreePaid
Public filingsSEC EDGAR (US), Companies House (UK), BRIS and European e-Justice company search (EU), national securities regulators for [GEOGRAPHY]AlphaSense, Sentieo
Earnings callsCompany IR pages, Seeking Alpha transcriptsAlphaSense (search across calls)
Private company signalsCrunchbase free tier, incorporation records, insolvency registers, beneficial-ownership registers where legally accessiblePitchBook, CB Insights
Government spend & procurementUSAspending.gov, SAM.gov (US), TED and national procurement portals (EU), country platforms (MENA), development-bank procurement (World Bank, EBRD, AfDB, UNGM)GovWin
Competition & state-aid casesEuropean Commission merger/antitrust/state-aid databases; national competition authoritiesn/a
State & sovereign capitalSovereign wealth fund reports, state-owned enterprise annual reports, PPP pipelinesn/a

Signal → inference chains:

  • Risk Factors section changes year-over-year in an annual filing → what genuinely scares them (they must disclose it)
  • Segment reporting restructure → strategic reprioritization; follow which segment got promoted
  • Earnings call Q&A dodges (analysts ask, execs deflect) → soft spot; probe it in your positioning
  • New entity registrations or branches in [GEOGRAPHY] → market entry before any announcement
  • Deferred revenue trends → actual sales momentum vs. stated momentum
  • Merger filings → market definitions and named competitors, straight from [TARGET]'s own lawyers
  • Sovereign fund or state-aid money backing [TARGET] → their runway math just changed; discount-pressure plays won't work
  • Prior Information Notices and expressions of interest → tenders telegraphed 3-24 months out

Feeds: battle cards (financial stress = "they'll be desperate at quarter-end" plays), SOM capture rates (their revenue ÷ claimed customer count = deal size reality check, feeding the GEOINT/DEMOINT sizing recipe), account targeting (procurement award patterns).

3. GEOINT/DEMOINT — The Cartographer

The terrain map, not troop movement. Government statistics are free intelligence most PMs never open — and the backbone of every ICP, persona, and TAM that survives scrutiny.

Source TypeFreePaid
US market structureCensus Bureau (County Business Patterns, Economic Census, NAICS establishment counts), BEAIBISWorld, Statista, Grand View Research
US labor & buyersBLS (occupation counts, wages, industry employment), FRED (macro conditions gating budgets)TalentNeuron
EU market structureEurostat, PRODCOM manufacturing statistics, national statistical institutes, ECB datanational data resellers
EU trade flowsEurostat COMEXT, Access2Markets, TARIC (tariffs, quotas, rules of origin)Panjiva, S&P Global
MENA regionalGCC-Stat, Arab Development Portal, ESCWA, Arab Monetary Fund, SESRICn/a
MENA nationalGASTAT (Saudi), FCSC (UAE), CAPMAS (Egypt), HCP (Morocco), and peersn/a
Global cross-checkWorld Bank Data and Enterprise Surveys, IMF country reports, OECD.Stat, UN Comtrade, ITC Trade Mapn/a

Signal → inference chains:

  • Establishment counts by industry code and employee band → the denominator for bottom-up TAM
  • Regional industry concentration → where your SOM actually lives, and where field sales should live
  • Occupation growth curves for your [BUYER] roles → is the population you sell to growing or shrinking
  • Wage trends in buyer roles → willingness-to-pay ceiling shifts; pricing corridor validation
  • Firmographic distributions (size bands, legal forms, sectors) → ICP boundaries drawn from data, not vibes
  • Buyer-title prevalence by [GEOGRAPHY] → persona localization; the "VP of Product" you message in Boston is a "Head of Digital" in Frankfurt and may not exist in Riyadh
  • Trade-flow shifts in product-specific codes → market entry or supply relocation before any announcement

The TAM/SAM/SOM Recipe (this discipline's signature dish):

TAM: Establishment counts for [MARKET] (Census/NAICS, Eurostat/NACE,
     or national equivalent for [GEOGRAPHY])
     × employment/spend benchmarks (BLS, Eurostat, trade associations)
     (validate against two independent analyst reports; if they disagree by 3x, say so)

SAM: TAM filtered by your actual constraints: [GEOGRAPHY], segment, compliance
     requirements, tech prerequisites (see TECHINT technographics),
     local-content and vendor-registration eligibility where applicable

SOM: SAM × realistic capture rate derived from [TARGET] public filings via FININT
     (their revenue ÷ their claimed customer count = deal size reality check)

Feeds: TAM/SAM/SOM (the backbone — see tam-sam-som-calculator), ICP definition, personas, messaging localization, market entry prioritization, pricing corridor validation.

4. TECHINT — The Patent Examiner

R&D leaves fingerprints 12-18 months before products ship.

Source TypeFreePaid
Patentspatents.google.com, USPTO Patent Center, EPO Espacenet, WIPO PatentScopeClarivate, LexisNexis PatentSight+
TechnographicsBuiltWith free lookups, WappalyzerHG Insights, BuiltWith Pro, 6sense
Product telemetryPublic changelogs, API docs diffs, status pages, GitHub org activityn/a
Standards bodiesWhichever govern [MARKET]: IETF, W3C, ISO committees, CEN/CENELEC/ETSI work programs, industry consortian/a
Funded researchCORDIS and Horizon Europe project databases, university project repositoriesn/a
Academic & preprintsarXiv, Google Scholar, Semantic Scholar, SSRN, the conferences that matter to [MARKET]Dimensions, Scopus
Trademark filingsUSPTO TESS, EUIPO, WIPO Global Brand DatabaseCorsearch

Signal → inference chains:

  • Patent clusters (5+ filings in one classification in 12 months) → committed bet on [CAPABILITY], not exploration
  • Inventor names repeating across filings → the actual product team behind the initiative; track their conference talks and LinkedIn
  • Trademark filing for a product-sounding name → launch inside 6-12 months (trademarks are cheap; companies file close to launch)
  • [TARGET] repeatedly appearing in funded consortia → their long-range bet, 12-48 months of lead time
  • Research pilot sites → likely launch customers, named in public deliverables
  • [TARGET] chairing a standards committee → they intend to shape the rules of [MARKET], not just play by them
  • [TARGET]-affiliated authors publishing preprints → R&D direction 6-24 months before patents; a paper cluster plus a hiring surge in the same specialty is one of the strongest fusion pairs available
  • Author affiliations shifting from university to [TARGET] on successive papers → they hired the lab, not just the idea
  • API docs adding endpoints for an unreleased capability → beta program running now
  • Your prospects' tech stacks (technographics) → your SAM refinement: who can actually buy you

Feeds: roadmap bets (where to accelerate vs. concede), SAM refinement, battle cards (feature-gap countdown clocks), build/buy/partner decisions.

5. HUMINT — The Sports Scout

Organizations announce strategy through job boards long before press releases. People are the tell.

Source TypeFreePaid
Job postingsLinkedIn Jobs, [TARGET] careers pages, IndeedJobsPikr, TalentNeuron, Revelio Labs
Employee sentimentGlassdoor, Blind, Reddit communities for [MARKET]n/a
Leadership movesLinkedIn announcements, pressBoardEx, The Org
Win/LossYour own sales team debriefs, churned-customer interviewsClozd, DoubleCheck
Conference hallwayYour field team's ears at [MARKET] trade showsn/a

Signal → inference chains:

  • Hiring surge in one specialty (30+ postings in a quarter) → building [CAPABILITY], not a feature
  • Regional specialist roles appearing for a [GEOGRAPHY] you haven't seen them in → expansion pre-announcement
  • Job posts naming specific technologies → confirmed stack choices → integration roadmap intel
  • Senior product/tech leader exits within 6 months of a strategy announcement → the strategy is in trouble
  • Your own alumni landing at [TARGET] → assume they know your playbook
  • Employee reviews mentioning "pivot," "reorg," "leadership churn" → 2 quarters of internal distraction = your window
  • Win/Loss interviews: the only source that tells you why deals actually close (everything else is inference)

Feeds: roadmap bets (their build signals), battle cards (org instability plays), win/loss program (ground truth for everything).

6. SIGINT — The Wiretap You're Allowed to Have

Companies broadcast constantly through what they change on the public internet. Most competitors never listen.

Source TypeFreePaid
Website diffsWayback Machine, Visualping free tierVisualping, Klue, Crayon (auto-monitoring)
Pricing pagesManual snapshots + WaybackKlue, Crayon, Kompyte
SEO/SEM movesGoogle "site:" queries, free Semrush lookupsSemrush, Ahrefs, SpyFu
App store metadataVersion notes, screenshot changes, keyword shiftsSensor Tower, data.ai
DNS/infrastructurecrt.sh (new SSL certs reveal new subdomains), DNS recordsn/a
Webinar/event cadence[TARGET] events pages, registration platformsn/a

Signal → inference chains:

  • New subdomain SSL cert (e.g., [capability].[target].com) → product launch staging, often weeks ahead
  • Pricing page removes a tier → packaging overhaul, usually toward enterprise
  • Sudden SEM bidding on your brand terms → they consider you the threat now (congratulations)
  • Case study page pattern shifts (new vertical or [GEOGRAPHY] appearing) → segment push
  • Messaging A/B visible via Wayback diffs → they're unsure of positioning; hit the wound

Feeds: battle cards (the freshest layer — this is what keeps cards from going stale), pricing strategy, positioning counter-moves.

7. MASINT — The Satellite Photo

Measure the physical and operational exhaust. Abnormal resource allocation never lies.

Source TypeFreePaid
Supply chainImport/export records via ImportYeti (free tier), Eurostat COMEXT and customs codes, UN ComtradeS&P Global Supply Chain Intelligence, Panjiva, ImportGenius
Facilities & projectsCommercial real estate news, local business journals, permits in [GEOGRAPHY]; industrial-zone tenant announcements, environmental permits, EPC contract awards; satellite imageryCoStar
Ops capacitySupport response time sampling, status page incident frequencyn/a
Certifications & safetyWhichever gate [MARKET]: ISO, SOC 2, FedRAMP, CE marks; notified-body designations and safety recalls (EU); sector registriesn/a

Signal → inference chains:

  • 20%+ volume change in critical inputs → pre-launch or demand collapse (check which via FININT)
  • New supplier geographies or country-of-origin shifts → market entry, tariff hedging, or resilience play
  • Compliance certification "in process" listings → 12-36 month runway into a regulated segment, visible to anyone who checks the registry
  • Product recalls or repeated safety-alert patterns → quality strain; battle card ammunition with a public citation
  • Land allocation, power/water capacity reservations, or engineering-design contracts preceding equipment procurement → facility buildout 6-36 months before any launch announcement
  • Support response times stretching + hiring freeze in support roles → cash constraint or overwhelmed by growth (disambiguate via employee sentiment)
  • Office consolidations → cost compression; expect pricing aggression to follow

Note: supply chain and facility signals are strongest for hardware and industrial players. The software equivalent is ops capacity plus infrastructure-scale language in job postings.

Feeds: threat assessment, launch prediction and capacity estimates, battle cards (capacity-stretch objections: "ask them about their support SLAs lately").

8. All-Source Fusion — The Situation Room

One signal is an anecdote. Three correlated signals from independent disciplines is intelligence.

Apply the Confidence Stacking Rule from the autonomous-investigation protocol: 1 discipline = watch item; 2 = working hypothesis; 3+ = actionable intelligence; conflicting disciplines = someone is bluffing — dig. And treat announcements as intent until funding, procurement, land, permits, hiring, or contracts corroborate them: ambition is OSINT; commitment shows up in FININT, MASINT, and HUMINT.

Fusion template — detecting [TARGET]'s [CAPABILITY] play:

DisciplineSignal (fill in what you found)
MASINTResource/input anomaly: ____
TECHINTPatent cluster or repo activity: ____
HUMINTHiring pattern: ____
SIGINTInfrastructure or web change: ____
FININTFiling language, procurement award, or earnings dodge: ____
GEOINT/DEMOINTTerrain check: does the market they'd enter actually exist at the size the move implies? ____

Fusion verdict: ____ disciplines, one story → confidence level → recommended response.

The fusion cadence:

  • Weekly: SIGINT sweep (site diffs, pricing, job posts). 30 minutes.
  • Monthly: OSINT + HUMINT digest. Review mining, employee sentiment, conference intel.
  • Quarterly: FININT + TECHINT deep pass. Filings, patents, procurement awards.
  • Annual + every TAM refresh: GEOINT/DEMOINT pass. Statistics releases lag; sizing rot is slow but real.
  • Event-driven: MASINT alerts, material filings, leadership exits. React within 48 hours.

Mapping disciplines to PM artifacts

ArtifactPrimary DisciplinesRefresh Cadence
TAM/SAM/SOMGEOINT/DEMOINT + FININT (capture rates) + TECHINT (technographics)Annual, plus event-driven
ICPs & PersonasGEOINT/DEMOINT + HUMINT (win/loss ground truth)Semi-annual
Messaging & localizationGEOINT/DEMOINT + OSINTSemi-annual
Battle cardsSIGINT + OSINT + HUMINT (win/loss)Weekly SIGINT layer, monthly rebuild
Roadmap betsTECHINT + HUMINTQuarterly
PositioningOSINT + FININT (earnings language)Semi-annual
Pricing strategySIGINT + FININT + GEOINT/DEMOINT (wage/WTP corridors)Event-driven
Threat assessmentAll-Source FusionQuarterly brief + event-driven

Regional note

The disciplines do not change across markets; the sources do, and so does the evidentiary burden. Two lessons that generalize: in rich-disclosure systems (EU), the challenge is finding the right record in a fragmented landscape — no single portal holds it all. In announcement-heavy markets (MENA), the challenge is separating national ambition from funded, procured, permitted commitment — and an announced budget, an approved budget, committed financing, a tender value, and an awarded contract are five different numbers wearing the same headline. Search in local languages; never assume the English portal contains every notice.

A copy/paste collection plan — instantiation variables, discipline-selection table, cadence, and fusion table — lives in template.md.

Examples

Illustrative fusion fill (all names fictional): you suspect [TARGET] = Meridian Freight Systems is building a warehouse-robotics platform. The sweep finds: +20% specialized component orders in customs records (MASINT, Fact), 15 new filings in one robotics patent class (TECHINT, Fact), 30+ platform engineers hired in a quarter (HUMINT, Fact), a new robotics.meridianfreight.com SSL cert (SIGINT, Fact), the CFO dodging an analyst's capex question (FININT, Inference), and establishment counts confirming the addressable segment supports the investment math (GEOINT/DEMOINT, Fact). Six disciplines, one story: high-confidence platform threat. Response: accelerate your own platform roadmap and arm sales with a maturity battle card before their launch, not after.

Single-discipline quick win: your battle card claims a competitor is "financially strong." One FININT pass finds deferred revenue declining three quarters straight while claimed customer count grew — Inference: deal sizes are shrinking. The card gains a quarter-end discount-pressure play, with the filing URL attached.

See examples/sample.md for a complete worked collection plan (fictional) — four disciplines chosen and three consciously skipped, a fusion table that escalates to actionable, and a June signal that correctly stayed a watch item until other channels corroborated. examples/sample-industrial.md is the plan where MASINT stars — customs, registries, permits — and an absence enters the fusion table as a Fact.

Common Pitfalls

  • Feature-matching theater. The goal is never "they shipped X so we ship X." Signals inform decisions about outcomes — "this changes which customer problem wins the next investment."
  • Single-discipline conviction. A hiring surge alone is a watch item, not a strategy brief. Escalate confidence only as independent disciplines corroborate — that's the stacking rule.
  • Announcement inflation. Re-planning your roadmap around a press release. Ambition is OSINT; wait for the money, the permits, or the hires.
  • Unsourced claims. A battle card assertion without a source and date is an opinion wearing a badge. Document provenance on every claim; archive source documents at collection time, because portals restructure and files get replaced.
  • Crossing the line. Pretexting, soliciting NDA-protected information, hiring someone specifically to extract their former employer's secrets, scraping against terms you agreed to — all off-limits. Everything in this skill is published, filed, posted, or publicly observable. The stage test: if you'd be uncomfortable explaining your method at [TARGET]'s user conference, don't use it.
  • Running every discipline on every question. The artifact-mapping table exists so you run the two or three disciplines that feed your [DECISION], on the cadence that matches how fast that evidence actually changes.

References

  • autonomous-investigation (Workflow) — the protocol that labels and stacks what these disciplines collect
  • tam-sam-som-calculator (Component) — consumes the GEOINT/DEMOINT sizing recipe
  • company-research and company-intel — single-company deep dives built largely on OSINT + FININT + HUMINT
  • Investigation skills that run these disciplines: market-landscape-scan, competitive-research-snapshot, competitive-intel-watch, battle-card-builder
  • SCIP Code of Ethics — the competitive-intelligence profession's reference standard
  • Regional Source Overlays (EU/MENA companion; planned as a future component skill)
  • Adapted from Dean Peters' "Competitive Research on Steroids" compendium and the Competitive Analysis Playbook (Productside).

Dépôt GitHub

deanpeters/Product-Manager-Skills
Chemin: skills/intelligence-collection-disciplines
0
ai-agentsai-product-managementclaude-skillspm-frameworksproduct-management
FAQ

Questions fréquentes

Qu’est-ce que le Skill intelligence-collection-disciplines ?

intelligence-collection-disciplines est un Skill Claude créé par deanpeters. Un Skill regroupe des instructions et des ressources que Claude charge à la demande pour effectuer des tâches liées à intelligence-collection-disciplines sans consigne supplémentaire.

Comment installer intelligence-collection-disciplines ?

Utilisez les commandes d’installation de cette page : ajoutez intelligence-collection-disciplines à Claude Code comme plugin ou clonez son dépôt dans votre dossier skills, puis redémarrez Claude pour charger le Skill.

À quelle catégorie appartient intelligence-collection-disciplines ?

intelligence-collection-disciplines appartient à la catégorie Autre.

intelligence-collection-disciplines est-il gratuit ?

Oui. intelligence-collection-disciplines est référencé sur AIMCP et son installation est gratuite.

Compétences associées

llamaguard
Autre

LlamaGuard est le modèle de Meta, doté de 7 à 8 milliards de paramètres, conçu pour modérer les entrées et sorties des LLM selon six catégories de sécurité comme la violence et les discours haineux. Il offre une précision de 94 à 95 % et peut être déployé avec vLLM, Hugging Face ou Amazon SageMaker. Utilisez cette compétence pour intégrer facilement le filtrage de contenu et des garde-fous de sécurité dans vos applications d'IA.

Voir la compétence
cost-optimization
Autre

Cette compétence de Claude aide les développeurs à optimiser les coûts du cloud grâce au redimensionnement des ressources, aux stratégies d'étiquetage et à l'analyse des dépenses. Elle fournit un cadre pour réduire les dépenses cloud et mettre en œuvre une gouvernance des coûts sur AWS, Azure et GCP. Utilisez-la lorsque vous devez analyser les coûts d'infrastructure, redimensionner les ressources ou respecter des contraintes budgétaires.

Voir la compétence
sports-betting-analyzer
Autre

Cette compétence Claude analyse les marchés des paris sportifs, incluant les spreads, les over/under et les paris spéciaux, en examinant les tendances historiques et les statistiques situationnelles pour identifier les paris à valeur ajoutée. Elle fournit une sortie en markdown structuré avec des recommandations actionnables à des fins éducatives. Les développeurs doivent l'utiliser pour des outils d'analyse de paris sportifs tout en notant qu'elle est conçue uniquement pour le divertissement et l'éducation.

Voir la compétence
quantizing-models-bitsandbytes
Autre

Cette compétence quantifie les LLMs en précision 8 bits ou 4 bits à l'aide de bitsandbytes, permettant une réduction de 50 à 75 % de la mémoire utilisée avec une perte de précision minime. Elle est idéale pour exécuter des modèles plus volumineux sur une mémoire GPU limitée ou pour accélérer l'inférence, prenant en charge des formats comme INT8, NF4 et FP4. La compétence s'intègre à HuggingFace Transformers et permet l'entraînement QLoRA ainsi que l'utilisation d'optimiseurs en 8 bits.

Voir la compétence