Cognitive operating system: think-before-act, memory hygiene, depth-adaptive execution. The meta-skill that sits above the rest.

Write a software design spec before code.

Break an approved design into a minimum-change plan.

Implement features with read-before-edit discipline.

Review diffs and PRs for correctness, security, performance.

Reproduce, isolate, hypothesize, fix, regress.

Unit, integration, e2e, and LLM evals.

CI/CD, IaC, progressive delivery, rollbacks.

Stripe, webhooks, idempotency, PCI awareness.

Reproducible pipelines: pinned deps, fixed seeds.

Production handoffs between Computer and Replit Agent.

Tool loops, memory, recovery, 3-tier evals.

Production system prompts, JSON schemas, few-shot.

Chunking, embeddings, hybrid search, reranking, citation grounding.

Pick the right model for cost/latency/quality.

Behavior, evals, staged rollout, kill switch.

PII redaction, injection defense, audit logs.

Dataset curation, labeling, DVC, dataset cards.

When to fine-tune vs RAG vs prompt; LoRA, QLoRA, DPO.

The meta-skill for delegating well.

Token economics, prompt caching, $/request budgets.

Model registry, canary, drift, rollback.

Streaming, citations, confidence, guardrails.

Decide when Kimi K3 wins vs. Claude Fable 5 / GPT-5.6 Sol / Opus 4.8 / GLM-5.2. Benchmark-cited, honest, refuses to recommend K3 when fit is worse.

Plan and execute a self-hosted K3 deployment on vLLM / SGLang / TokenSpeed with MXFP4 weights, preserved-thinking, and a license gate.

Apply factored CoVe as the final layer before delivery: draft → independent verification questions → fresh-context answers → revise. 40–60% documented hallucination reduction. Never labels a claim verified without independent-context confirmation.

Run a small language model (1B–14B) on hardware you own. Hardware-ceiling and KV-cache VRAM math, GGUF quantization tradeoffs, runtime selection across Ollama / LM Studio / llama.cpp / MLX / vLLM, air-gapped install. Honest about the capability ceiling: private and free, not frontier.

Build a research or study assistant on a small local model you can trust. Local RAG with local embeddings, a refusal contract tested before anything else, task decomposition that names which steps must route to a bigger model, span-level citation, verification before display. Private, offline, and honest about its range.

Run OpenClaw and Maximus on a local model so the agent actually sees its own prompt. Current config schema, the 64K context floor, setting context where it takes effect, and a check that proves nothing is being silently truncated.

Operate Ollama as a service, not a demo. Modelfiles, context and KV-cache sizing under concurrency, server settings, GGUF import with template checks, native vs OpenAI-compatible API, structured outputs.

Make small models call tools and emit JSON reliably. Native templates, grammar-constrained decoding, a smaller tool surface, one tool per step, and a measured failure rate instead of a lucky run.

Evaluate local models and quants on your own tasks and hardware: promptfoo side by side, lm-eval-harness for quant degradation, llama-bench at real context depth, and a decision memo.

Fine-tune on hardware you own and ship it to Ollama: LoRA/QLoRA with Unsloth or MLX, eval against the base model, GGUF export, and the chat template kept intact. Earns the fine-tune first.

Long-form articles: thought leadership + technical builds.

Multi-source synthesis and competitive intelligence.

Inference-driven research: hypothesis-first, adversarial verification, confidence ledger.

Research the way a working journalist or intelligence analyst does it: source tiers (primary/secondary/adversarial), timeline construction, contradiction hunt, corroboration pass. Produces a story with a spine, not a topic summary.

Systematic literature review with PRISMA-style flow, inclusion/exclusion criteria, quality appraisal (RoB 2 / NOS / ROBINS-I / field-adapted), and GRADE confidence per finding.

Deep 7-step agent for investors, journalists, partners, board members, experts.

Deep 8-step recruiter agent for candidate sourcing.

Finance-grade counterparty discovery: SEC filings, court dockets, source-by-source provenance, compliance gate before outreach.

Find the most likely professional business email from a LinkedIn URL, with pattern discovery, verification, and confidence scoring. Never labels a guess as verified.

Turn a dense deal folder — emails, term sheets, closing docs, wires — into a two-page executive transaction memo. Reconstructs chronology, reconciles numbers, separates facts from allegations, and names what is still unresolved. Never fills factual gaps with outside knowledge.

The strategy layer.

Answer Engine Optimization for LLM answers.

Generative Engine Optimization for Perplexity, SGE, Copilot.

Core Web Vitals, JSON-LD, crawlability, JS SEO.

On-page, internal linking, E-E-A-T, refresh cadence.

Umbrella audit that ties the pack together.

Measure citation share across Perplexity, ChatGPT, Claude, Gemini, AI Overviews.

Run Claude Code inside Cursor with fewer approval interruptions — launch flags, a continuous-execution rule that pre-authorizes routine reversible in-repo work, interrupted-session recovery from git state, and approval-prompt triage across the two permission layers. Honest about the limit: fewer confirmations, not fewer consequences.

Console-first Fountain screenplay authoring; industry-standard PDF (Courier Prime 12pt, 1.5" margin) and Final Draft .fdx export; runtime estimation.

Turn a locked Fountain script into scene breakdown, stripboard, one-liner, day-out-of-days. Console-first, a Movie Magic alternative.

AICP/AMPTP account-code budget templating (1000–4700 series), ATL/BTL, contingency, completion bond, fringes. CSV/XLSX Movie Magic round-trip.

Dailies workflow, offline/online conform, EDL/AAF/XML round-trip, ACES vs Rec.709, DCP prep, per-buyer deliverables matrix.

Prepare, validate, and troubleshoot feature-film delivery masters. FFprobe-first, remux-don't-re-encode, attached-thumbnail cleanup, WSP/OBA compliance flag.

Given a company, find its shareholders and investors — firms, funds, individuals — from public funding data, enrich their contacts through a connected provider, log them to a store you choose, and queue personalized outreach for approval. Optional 3-day follow-up and 7-day re-discovery flows. Maximus drafts and schedules; the rider approves every send.