M Made Scientific MADE OS · 02 · Skill Library

Skill Library

A skill is one job, written down once, in a form both a human in a chat window and an agent on a cron can execute identically. 62 are active. They are the unit the whole OS composes from: an agent is a skill plus a trigger, and a fleet is a set of agents sharing a library.

Counted 2026-08-04 from Transfer os_skills — 62 status=active.
Active skills
62
Transfer os_skills
Domain groups
4
commercial · shared · identity · platform
Wired to an agent
33
via os_skill_agent_map
Human-only so far
29
no agent uses them yet
Median maturity
60
self-assessed 0–100
P0 skills
3
identity matching core
The definition, because it does real work

A skill is not a prompt, and it is not documentation

A prompt is disposable and lives in someone's chat history. Documentation describes what happened. A skill is an executable instruction set with a name, a trigger vocabulary, declared data sources, and a maturity score — so that the fourth time someone needs to classify an account, they run the same thing the agent runs, and both get better together.

Practically, this is what makes the fleet compound. When account_matching_ladder improves, every agent that uses it improves, and so does the analyst doing it by hand. Twenty-nine of the 62 skills are currently human-only — that isn't a gap, it's the pipeline: skills get written for humans first, then wrapped in an agent once the shape is stable.

The load-bearing three

Identity resolution is where the rigour lives

Three skills carry a p0 priority, and they are all about the same problem: deciding whether two records are the same real-world thing. Everything downstream — account metrics, campaign attribution, engagement tiers, a rep's view of a relationship — is a function of getting this right. They are also the best-engineered skills in the library, and worth reading as the house standard.

P0

confidence_scoring

activematurity 70resolver-cron
What it doesLog-odds signal fusion with a three-axis disposition — evidence tier × cardinality × score. Replaces a flat CONF_BY_METHOD lookup table.
Signals fused15
Unit tests31

Why it's the standardIt has a negative prior and a corroboration-only ceiling — meaning it can conclude "probably not a match" and it refuses to get confident from a single weak signal. Most scoring code can do neither.

Answers"Why 0.87?" — with a breakdown, not a number.

P0

account_matching_ladder

activematurity 65
What it doesOrdered rungs — negative check → alias → rule → domain → name → trigram → LLM → mint — producing a candidate set, not a pre-picked winner.

The two rules that matterMatch before mint (never create a new ID before exhausting the ladder) and never silently discard (an unmatched candidate becomes a proposal, not a deletion).

Why a set, not a winner
setA human can adjudicate between three candidates.
winnerA human cannot audit a decision they never saw alternatives for.

FeedsMADE-ID CUSTODIAN, RESOLVER.

P0

contact_matching_ladder

activematurity 60
What it doesID-chain-first ladder with a role-address filter and a multi-person safety net.
Documents its own bug. The skill records a rung-inversion defect: 94 conflicting assignments turned out to be duplicate MADE-IDs, not genuine ambiguity — a different problem needing a different fix. Writing the bug into the skill is why the next person doesn't misdiagnose it.

StatusP0/P1 hardening documented in-skill.

A fourth, one rung down but worth naming: contact_account_rollup (p1) — "the 31% gap." Roughly a third of CRM contacts carry a MADE-ID but resolve to no company, which understates every account-level engagement metric by about a third. It gates campaign attribution. This is a known, quantified, scoped gap — Tier 2, not Tier 1, because nobody is pretending it's solved.
By domain

Commercial — 34 skills

The revenue-facing half of the library, split by function: demand (find and engage), delivery (propose and produce), operations (run the machine), knowledge (how we talk).

Demand 9
SkillWhat it doesMaturityAgents
outreach_message_composerCompose A+ outreach messages against the 6-dimension MADE rubric.750
social_listeningMonitor and interpret social signal.702
competitive_battlecardPer-competitor battle cards for BD conversations.654
campaign_trackerTrack campaign state and performance.601
sfdc_opportunity_briefOne-page exec brief for an SFDC opportunity.651
inside_salesInside-sales motion and cadence.550
pipeline_executionPipeline execution discipline.551
brand_authorityBrand authority and media positioning.450
campaign_crosswalkMember-set Jaccard across SFDC / HubSpot / bc_outreach. Records itself as currently cosmetic — 99 groups × 1 system each, zero cross-system merges ever. Blocked on contact_account_rollup.450
That last row is the library working as intended. campaign_crosswalk declares in its own description that it has never produced a cross-system merge and names what it's blocked on. A skill that documents its own uselessness is more valuable than one that quietly returns empty results.
Delivery 6
SkillWhat it doesMaturityAgents
proposal_rfp_reviewReview a Made proposal against a client RFP for scope, technical accuracy and pricing.752
content_studioContent studio and outreach generation.702
made_report_template_designReport template design.650
made_presentation_builderBranded MADE pptx from a markdown outline.601
proposal_excellenceProposal quality standard.551
campaign_message_preview_modalRead-only preview modals for campaign messages in React + Tailwind.500
Operations 18
SkillWhat it doesMaturityAgents
elt_weekly_reportThe ELT Weekly Report for the Wednesday LT call.801
crm_field_sync_governanceWhich system owns which field, and what may overwrite what.751
news_intelligenceNews monitoring and relevance scoring.704
priority_inbox_triageTriage an inbox using email_insights priority scores.701
buyer_segment_classificationAssign a buyer segment.652
kpi_reviewKPI review discipline.602
meeting_intelligenceExtract decisions and actions from meetings.601
portfolio_dashboardPortfolio-level reporting.601
database_metricsStandard DB metric definitions.551
demand_planningDemand and capacity planning.552
pricing_frameworkPricing and costing framework.551
revenue_forecastingRevenue forecasting method.55
crm_integrationCRM integration patterns.550
monday_integrationMonday.com board integration.550
data_hygieneData hygiene standards.552
account_classificationClassify an account.502
territory_managementTerritory definition and assignment.
scorecard_reviewScorecard review cadence.
Knowledge 1
SkillWhat it doesMaturityAgents
made_commercial_opsMADE commercial vocabulary, voice registers and anti-patterns. The reason two agents describe the same deal the same way.751
By domain

Shared — 28 skills

Cross-cutting capability: how output gets made to look right, how identity gets resolved, and how engineering work gets done safely. These are the skills any agent or any person can pick up regardless of which fleet they sit in.

Design & delivery 10
SkillWhat it doesMaturityAgents
made_design_systemApply @made-sci/design-system to any Lovable/Vite app.650
made_brand_guidelinesOfficial MADE brand identity for presentations and visual output.602
figma_design_intelligenceRead and reason about Figma design intent.550
excalidraw_diagramsDiagram creation from natural language.550
image_generationOn-brand images, palette and imagery direction pulled from the brand pack.503
video_generationVeo clips, HeyGen talking-head, assembled short video.502
landing_page_generationOn-brand landing pages driven by Made v2 tokens.500
brand_pack_scaffolderTurn a brand bible into a working brand pack.500
ui_scaffoldingReact/Tailwind UI that is on-brand by construction.500
social_carouselMulti-slide LI/IG carousels with a consistent Made look.451
Nine of these ten have zero or one agent attached. Design skills are currently a human capability with agent assistance, not an automated pipeline — which is the honest state and arguably the right one. Worth revisiting only if volume makes it a bottleneck.
Identity & data 7
SkillWhat it doesMaturityP
confidence_scoringLog-odds fusion, 15 signals, 31 tests. See above.70p0
account_matching_ladderEight-rung ladder producing a candidate set.65p0
contact_matching_ladderID-chain-first with role-address filter.60p0
contact_account_rollupCloses the ~31% of contacts with a MADE-ID but no company. Gates campaign attribution.55p1
account_matchingAccount matching and deduplication (the general skill).60
data_hygieneHygiene standards, shared with commercial ops.55
coordinateHub-and-spoke multi-agent prompting discipline for complex tasks.60
Engineering discipline 11
SkillWhat it doesMaturity
commitStructured commit messages that function as long-term memory for the next session.80
health_checkHealth check across all Supabase databases; updates SYSTEM_STATUS.75
primeSession priming — catch an agent up on codebase state.75
session_closeEnd-of-session protocol: update docs, commit, log outcomes.75
plan_with_riskRisk-classified planning for HIGH-risk changes (the ULTRAPLAN equivalent).70
verifyAdversarial self-verification — catch issues before marking work done.70
supabase_auditSecurity + performance audit against a Supabase project.70
lovable_rls_policy_checkPre-flight RLS check before adding any new Supabase write in a Lovable app.70
switch_dbSwitch context to a specific Made database for MCP queries.70
claim_project_numberReserve a PROJ-NNN via the registry before creating a new project.65
Why the boring ones matter most

The highest-maturity skill in the library is commit

Not a model call. Not a classifier. A convention for writing commit messages — scored 80, the top of the library. verify, prime, session_close and plan_with_risk sit just behind it.

That ordering is not an accident and it is not a gap. Those five skills are what make everything else repeatable: they encode how work starts, how it is checked, and how it is handed to the next session. An agent fleet without them produces output nobody can audit six weeks later — which is exactly the condition the four red findings describe.

Gaps, both tiers

What the library is missing

Split the same way as everywhere else on this site: things we believe are working that aren't, versus things we know we haven't built.

AMBER

Two skills declare themselves non-functional — and that's the good outcome

Believed working, self-corrected

Whichcampaign_crosswalk ("currently cosmetic — 99 groups × 1 system each, zero cross-system merges ever") and contact_account_rollup ("the 31% gap").

Why amber and not redBecause they say so themselves, in the registry, in the description field. Anyone querying the library gets the caveat with the record. That is the mechanism working — a skill honest about its own limits cannot mislead.

Worth copying as a convention: if a skill's output is not yet trustworthy, say it in the description, not in a side doc. The description travels with the data; the side doc doesn't.

ActionNone urgent. Both are correctly labelled and correctly blocked.

FUTURE

Three skills exist for agents that don't

Registry-first roadmap

Whichclinical_trials_monitoringtrialwatch (p1) · vertical_signal_monitoringsentinel (p1) · quote_win_loss_analysisVERDICT (p2).

Why this is deliberateSkill first, agent second. Writing the skill forces the thinking — what are the inputs, what is a good output, what would make this wrong — before anyone picks a harness. An agent built before its skill is a prompt with a schedule.

ActionNothing. This is the pipeline in its correct state.

The one structural gap worth naming: there is no telemetry from real runs back into the library. maturity is self-assessed, and agent_count counts wiring, not usage. So nobody can currently answer "which skill actually gets invoked most, and which one fails most often." That is a future item — but it's the one that would make every other number on this page earn its place.