Claude Skills by mcorbett51090

Sequence lease rollovers, recovery improvements, and capex against a quarterly NOI target so a held asset tracks (or beats) its acquisition underwriting. Reach for this once an asset is owned.

Convert a face rent to net effective by netting TI, free rent, and leasing commissions, so comps and underwriting use the rent the landlord actually earns. Reach for this on any rent comp.

Frame a cap rate as a risk premium over the 10-yr Treasury, not an absolute level, so a 'compression' is read correctly. Reach for this whenever a cap rate enters a memo.

Size the debt, schedule DSCR through the hold, and surface the refinance year and the rate at which the deal breaks. Reach for this before any levered return is quoted.

Build a CRE base case on contractual in-place income before any pro-forma step-up — separating real income from assumed growth so the return rests on something sourced. Reach for this when a deal is being sold on stabilized rents.

Curate and augment the dataset, choose and fine-tune the model (YOLO / DETR / SAM / CLIP / EfficientNet / ViT) on the task and target budget, design the loss and metric to match the cost, spend the annotation budget with active learning, handle class imbalance, and build an eval harness you can trust with per-slice metrics and drift detection. Model/checkpoint specifics verify-at-use; no PII.

Frame the computer-vision task (classification / detection / segmentation / OCR / pose / tracking / VLM) on the decision the system must make, choose the metric that mirrors the business cost, decide build-vs-fine-tune-vs-API jointly with the deployment target, and design the data & annotation strategy. Model/metric specifics verify-at-use; no PII, no image data stored.

Design streaming-video pipelines that hold the frame budget — frame sampling / keyframe strategy, ROI cropping, tracking-by-detection so the detector doesn't run every frame, and batching where latency allows — and deploy to edge/embedded targets (Jetson, mobile NPU, Coral) with the camera/sensor capture and pre-processing counted inside the budget. Device numbers verify-at-use; no PII.

Make a trained vision model fit and run on its target: budget latency on the real device, optimize in order of leverage (quantization INT8/FP16 with calibration, then pruning, then distillation), export to the runtime the target uses (ONNX / TensorRT / CoreML / TFLite / OpenVINO), and re-check accuracy against the operating point after every step. Device/runtime numbers verify-at-use; no PII.

Build and maintain a CPM schedule: activity list, durations, logic ties (FS/SS/FF), critical-path calculation, float analysis, baseline establishment, weekly updates, look-ahead schedules, delay analysis (as-planned vs. as-built, TIA), and recovery scheduling. Tool-agnostic; notes P6 and MS Project specifics.

Run a complete GC estimate and bid assembly: quantity takeoff from drawings, unit pricing (labor/material/equipment), markup vs. margin conversion, subcontractor scope review, general conditions budget, overhead and profit, contingency, and bid-letter qualification. Covers original bids and change-order re-estimates.

Manage the GC's submittal register (identify required submittals, set lead-time-driven due dates, track review status), draft and track RFIs (structured question format, response tracking, overdue escalation), and document change orders (CO pricing coordination, package assembly, CO log maintenance). The three processes are connected: RFIs clarify design, submittals confirm procurement, and COs pay for scope changes.

Design the dialog layer of a voice agent: explicit dialog-state modeling, LLM orchestration, context/window management over a long call, and mid-call function/tool calling that hides latency with natural fillers and treats tool failure as a first-class branch. Design the unhappy path (silence, misrecognition, tool failure, confusion) before the happy path. Provider/SDK specifics verify-at-use.

Engineer the ASR and TTS layer of a voice agent: choose STT/TTS providers on the channel (narrowband telephony vs wideband web), languages, streaming support, WER, and cost; stream transcription and synthesis to cut latency; tune VAD/endpointing; handle diarization, prosody/SSML, codecs & sample rates, and noise robustness; and measure WER on real audio. Provider/version specifics verify-at-use.

Integrate a voice agent with real telephony and validate it: SIP/PSTN trunking and media handling, WebRTC for web/app, DTMF capture (in-band vs RFC 2833/SIP INFO), IVR-vs-conversational per step, call routing and warm/cold human transfer with context, and a voice-agent eval harness scoring task success, WER, end-to-end latency, and interruption handling on representative recorded calls. Protocol/provider specifics verify-at-use.

Choose the voice-agent pipeline shape (cascade STT->LLM->TTS vs speech-to-speech), the channel (telephony / web / SDK), and the build-vs-platform bet (Twilio / Vapi / Retell / LiveKit / Pipecat), then allocate the end-to-end latency budget per hop and design the turn-taking / barge-in / fallback model. Model/platform/latency specifics verify-at-use.

State the deal thesis — why this target, why now, how value is created — in one sentence before any model is built. Reach for this at the very start of any acquisition, before valuation.

Turn synergies into an owner/date/cost-to-achieve register and a 100-day plan, and price integration cost back into the valuation — so the deal isn't underwritten on flawless integration. Reach for this before signing, not after close.

Build a confirm-or-kill diligence plan where every workstream tests a named thesis assumption and every finding is mapped to thesis or price impact — not a data-room checklist. Reach for this once a thesis exists and a deal is live.

Cross DCF, trading comparables, and precedent transactions into a valuation range where the divergence between methods is itself the finding. Reach for this whenever a price needs defending — one method is an opinion.

Build the recurring-revenue engine: design club/membership tiers on member lifetime value (shipment value/frequency, benefits), read and reduce churn by cohort, and manage DTC e-commerce. A club member is worth more than a case sold — but churn quietly eats the club.

Cost a craft-beverage unit and plan production against real capacity: decompose COGS per unit (raw material, yield loss, packaging, overhead absorption), read tank/barrel/time capacity and turns, and plan batches against the demand-by-channel plan. You can't price what you can't cost.

Read and lift the tasting-room funnel: visits -> tasting -> purchase -> club sign-up. Find where conversion leaks, tie it to the experience and the club offer rather than raw foot traffic, and weigh tasting room vs e-commerce vs events on contribution. DTC keeps the margin wholesale gives away.

Model the go-to-market structure: self-distribution (margin kept, sales/logistics cost, eligibility limits) vs a distributor (reach and depletion, margin given away, franchise-law lock-in), plus the channel margin math (DTC net vs wholesale net after distributor/retailer take). Every TTB/state-licensing/excise specific is jurisdiction-specific — flag and route it.

Decide which revenue lines a creator should run and in what order — grounded in audience size × engagement × buying-intent, weighted toward recurring revenue and away from single-platform/single-sponsor concentration. Reach for this at the START of monetizing an audience, or when income is volatile/over-concentrated. Driven by creator-business-strategist.

Plan a creator's content strategy, sustainable cadence, platform growth, and owned-audience funnel: content pillars, a capacity-matched schedule, repurposing one core piece into platform-native cuts, and converting rented reach into an email list/community — measured by engagement + owned conversion, not vanity reach. Driven by content-and-audience-manager.

Value a sponsorship and build a creator rate card on VALUE, not a flat CPM: audience value + deliverable + usage/exclusivity rights as separate line items + a walk-away number, with rate benchmarks dated + verify-at-use and clear disclosure. Reach for this when a brand makes an offer, or to set standard rates. Driven by creator-business-strategist.

Design the signal set and mart model for an account expansion-readiness view alongside the churn-risk tier — identifying which signals predict expansion, how to surface them without confusing the CS team, and what mart additions are needed. Reach for this skill when a CS team wants both a churn-risk view and an expansion-opportunity view from the same health mart.

Back-test a candidate churn signal or the full tier rule set against historical renewal outcomes to validate predictive strength before the signal enters the production health tier. Reach for this skill when a new signal is proposed, when the tier is misfiring, or after the first full renewal cycle to tune thresholds.

Audit the metrics published on a CS-health dashboard against the mart layer to identify inconsistencies, stale definitions, missing signals, and metrics that bypass the mart. Reach for this skill before a CS analytics rebuild, during a QBR preparation review, or when a CS leader reports that the numbers don't match what they expect.

Design a transparent, explainable rule-based customer-success health tier (Green/Yellow/Red) from multi-source signals: signal selection, weighting, thresholds, per-signal evidence display, and tuning against actual past churn. Reach for this skill when designing a new health tier, refreshing one that has stopped predicting outcomes, or deciding how a Red account should explain itself. Domain-neutral — generalized from the EdTech partner-health-scoring pattern; no vertical assumptions.

Design the renewal-risk workflow and save-play triggers from a health tier plus renewal proximity: the risk = proximity × engagement rule, the renewal watchlist surface, the trigger-to-play mapping, the expand/maintain/recover decision, and the who-do-I-call-today actionability bar. Reach for this skill when designing how at-risk renewals surface and what fires when, or when a renewal that should be safe shows no movement. Domain-neutral — generalized from the EdTech renewal-play-design patte...

Design a statistically meaningful QA sampling program and a tier/escalation model. Reach for this on a quality or routing question.

Model self-service/KB deflection and the cost avoided before sizing headcount. Reach for this first on a cost-to-serve question.

Read SLA and backlog as arrivals against resolution capacity — project the days-to-clear. Reach for this on a backlog question.

Read CSAT/NPS segmented by channel/tier/issue-type and tie it to FCR — never a blended score. Reach for this on a satisfaction question.

Size agents from workload and a target occupancy band — not a fixed agent:ticket ratio. Reach for this on a staffing question.

Set the control-testing cadence, build evidence collection and continuous control monitoring, decide Type I vs Type II readiness, run a gap assessment and manage the auditor PBC list, and own third-party risk — vendor tiering, SIG/CAIQ, shared-responsibility, and ongoing monitoring — so evidence is a system and the audit holds no surprises.

Choose the right security-compliance framework for the org's size/risk/customer demand, scope the audit boundary, crosswalk controls across SOC 2 TSC / ISO 27001 Annex A / NIST CSF 2.0 / 800-53 so one evidenced control attests many, and author a Statement of Applicability whose every exclusion is justified against the risk register.

Build a risk register (assets, threats, likelihood x impact scoring), drive control selection from risk rather than from a framework checklist, choose a treatment per risk (mitigate / accept / transfer / avoid), and track residual risk with a named owner — so every control traces to a risk and every top risk has a control.

Design and operate data access governance: define role-based access tiers from the classification scheme, implement column-level masking and row-level security, build an access request and review workflow, and produce an access audit report that proves least-privilege is enforced.

Build a maintained data catalog: automated sensitive-data/PII discovery and column-level tagging, end-to-end lineage captured from pipeline/dbt metadata, a business glossary tying terms to tables, and surfaced access for governance.

Design and apply a usable data classification scheme: a small set of levels (public/internal/confidential/restricted) plus a PII/sensitive flag, handling rules per level, and a mapping to enforceable controls — governing the highest-risk data first.

Engineer privacy: build executable data-subject-rights pipelines (access/erasure/portability) that locate data via the catalog, track lawful basis + granular revocable consent, minimize collection, automate retention/deletion, and distinguish pseudonymization from anonymization.

Design and operate a data retention and deletion programme: classify data by retention period, build automated purge pipelines, verify deletion propagates to all copies including derived and replicated data, and produce evidence for audit — so data does not persist beyond its legal or business justification.

Pick the right data-pipeline orchestrator for a described workload by traversing the orchestrator-selection decision tree (workload shape → latency → asset-centric vs task-centric → team/ops capacity → cloud lock-in → engine), then return the recommended engine, its scheduling model, its executor/runtime, the trade-offs, and the conditions that would flip the choice. Reach for this when the user asks "Airflow vs Dagster vs Prefect?", "self-host or managed (MWAA/Composer/ADF)?", or "what shoul...

Design a correct, minimal DAG or software-defined-asset graph for a pipeline — derive the real upstream→downstream edges, pick the partition grain, choose the scheduling/triggering model (cron / sensor-deferrable / data-aware-asset), and decide catchup behavior — then capture it in the DAG design doc. Reach for this when the user says "design the DAG/assets for <pipeline>", "what should depend on what?", or "how should these jobs trigger?". Used by `pipeline-orchestration-engineer` (primary).

Make pipeline tasks safe to re-run and plan backfills that don't corrupt state — prove idempotency (deterministic partition keys, overwrite-by-partition), then add bounded retries with exponential backoff + jitter, and run a controlled backfill (partition strategy, catchup config, concurrency caps, monitoring, rollback). Reach for this when the user says "our pipeline isn't safe to re-run", "add retries", or "we need to backfill <range>". Used by `pipeline-orchestration-engineer` (primary).

Author a custom Airbyte connector (low-code manifest.yaml vs Python CDK) — monotonic server-set cursor, checkpointed state, bounded-window backfill, Retry-After backoff, idempotent destination write, maintenance posture at design time. Used by connector-developer. NOT for configuring a vendor-shipped Airbyte/Fivetran connector (that's connector-configuration).

Compare cloud databases (Supabase, Neon, RDS, Azure SQL, Fabric, DuckDB, MotherDuck, Snowflake, Databricks, Turso) for an SMB consulting engagement — pricing tables with retrieval dates, setup-complexity matrix, when-to-pick guidance. Used by `database-setup-guide` to recommend a DB choice for a new engagement.