// 40 DEEP DIVES · FOUR READING LANES
Find the method for the problem in front of you.
Start with the job: operate, verify, ship, or explore. Every path returns to evidence and an explicit next move.
01 / OPERATE
Direct the work
Tools, settings, voice, control surfaces, and the contracts that keep agent work bounded.
Running AI Agents in Production: An Operations ManualScoping, the authority line, evidence ladders, honest monitors, durable memory, and fleet coordination — one connected manual.
The four-agent toolbenchClaude, Codex, Droid and Cursor — what each is best at, how to route work between them, and the one mirrored context file that keeps them in sync.
Tuning the knobs: models & settingsReasoning effort, model tier, temperature, tool-use — and separate saved profiles for coding speed versus agentic work, with copy-usable configs.
Voice-driven developmentDriving a fleet of agents by talking — handling transcription artifacts, and the shorthand that expands one spoken phrase into a whole workflow.
Hermes Agent: Telegram & OpenRouter setup guideWhat Hermes is, how to install it, connect a private Telegram bot, switch models, add skills and scheduled work, and keep the assistant safe.
Designing skills and loops for certain and uncertain outcomesHow to package agent skills as explicit contracts, use deterministic loops for exact predicates, and bound non-deterministic work with rubrics, budgets, reviewers, and human gates.
The quiet monitorA monitoring architecture that preserves raw observations, distinguishes unknown from unchanged, and interrupts only when a decision threshold is crossed.
Codex in my pocketHow I start and steer Codex work from my phone, keep a Mac available, and return to verified results.
How I run AI agents on a Hetzner VPSCreate and secure the server, install the runtime, connect Telegram, isolate multiple agents, and operate the system with evidence.
How to scope an AI agent taskA five-part method for turning an ambiguous request into a bounded task with a clear outcome, authority, evidence, and stop conditions.
Why my new AI agent was slow before it answeredA measured AI-agent prompt-bloat teardown of fixed context, skill indexes, tool schemas, and the profile rules that separate prompt weight from end-to-end latency.
The agent never touches the buttonAn automated research fleet that proposes, tests, and verifies everything — governed by one file and adversarial review — while a human carries every irreversible action the last step.
02 / VERIFY
Protect the truth
Source-of-truth checks, evidence loops, privacy boundaries, and failures that looked successful.
Policy as a testEncode a direction-level decision as an automated check so later sessions cannot quietly undo it—and name where the technique stops working.
When the monitor’s subject diesAssert that the subject still exists before a vacuous pass can be mistaken for a healthy system.
The page you deleted is still being servedClose the origin, feed, index, and external-verification loop after retiring a public page.
How I built a fail-closed vulnerability watcherFailed feeds, uncertain product matches, and unknown versions remain evidence gaps—not reassuring conclusions.
The verification loopA practical evidence ladder for turning ‘looks right’ into a result another person or agent can trust.
Context without oversharingHow to give agents enough context to work well while keeping private systems, personal details, credentials, and non-public facts out of public artifacts.
Before the AI edits: find the live truthA source-of-truth protocol for checking production, deploy state, repositories and generated artifacts before an agent changes the wrong version.
When ‘success’ is still brokenWhy a 200 response, green check, or finished-looking agent answer can still hide a failed user outcome—and what evidence closes the gap.
The stale worktree that passed every checkA postmortem pattern for the dangerous case where tests are green because they ran against the wrong source of truth.
Provenance before conclusionsHow to turn scattered public data into a reproducible evidence archive where every file, limitation, and claim can be traced.
Diagnose the observer firstA layered incident method for separating device, network, routing, remote-system, and user-outcome failures before changing anything.
AI is collapsing the CVE response windowWhy vulnerability discovery, disclosure, and weaponisation now demand a faster intelligence loop with human-owned decisions.
03 / SHIP
Make the work compound
Recipes, publishing, handoffs, coordination, and memory that improve the next run.
The branch that predates the decisionSweep a whole long-lived branch for retired assumptions before a tidy merge silently restores them.
The cookbookSeven copy-usable recipes — cache-bust blast radius, CSP hash recompute, adversarial verify, config recovery from live headers — each with the reasoning.
From rough brief to published pageHow a loose idea becomes a structured, visual, connected page—and leaves the seed for what to publish next.
Handoff packets that do not lose the plotA copy-usable transfer format for current state, decisions, evidence, stop conditions, and the smallest safe continuation.
Parallel agents without branch collisionsA practical ownership and integration method for running agents in parallel without colliding on shared files, generated artifacts, cache versions, or release state.
Decision ledgers for long-running agent workA practical way to preserve why an agent rule exists, append evidence that confirms or changes it, and name the trigger for its next review.
04 / EXPLORE
Test the frontier carefully
Local models, open weights, and risk ideas that remain useful when the tools change.
Serving LLMs on an RTX 5090: The Working RecipesFour validated profiles from 84.5 to 139.7 tok/s, the three context numbers that matter, and the traps that cost decode.
Local LLMs on Apple Silicon: The Field GuideWhich Mac runs what, 4-bit or 8-bit, the measured speculation gains, KV budgets, and when the GPU box wins.
The other 4-bit decision4-bit weights were nearly free; a 4-bit KV cache halves reasoning scores. The per-token math, the mechanism, and the budget a 32 GB card forces.
Same model, two machinesThe 27B on both a 128 GB M4 Max and a 32 GB RTX 5090: 46× apart in prefill, 5× in decode, 3× with speculation — and what is genuinely not comparable.
The rule was 250,000 tokens agoA 40-check graded suite at four depths of a 262k window — tool calls clean everywhere, one trailing period wasn’t, and the grader failed before the model did.
Speculative decoding, measured on my M4 MaxNative MTP +24%, a DFlash 2 drafter up to +183%, the acceptance rates behind them, and the throughput metric that nearly reverted a working config.
How I Choose a Quantization for Local InferenceThe formats that matter, the 32 GB memory math, where quality actually drops, and the nine-point test that decides Q4, Q8 or NVFP4.
I Set Up My RTX 5090 AI Workstation Without Touching ItQwen 3.8 27B across vLLM, SGLang, context and KV-cache tests—then a private off-site inference proof through Tailscale.
Building My Personal AI Lab: RTX 5090, Qwen3.6 and Local InferenceA dedicated local-compute plan for MTP inference, agentic workflows and reproducible open-weight model benchmarks.
Local & uncensored modelsRunning open-weight models on your own machine — the stack, why privacy and cost make it worth it, and a plain, responsible explanation of what abliteration is.
Local model field notesA living test contract for local models — runner, quantization, configured context, latency, failure modes, and honest cloud comparisons.
Position sizing for agent autonomyWhat derivatives, portfolio construction, and surviving volatility teach about permissions, stop conditions, and bounded agent action.
Building a paper-first DFM trading systemHow an automated DFM and ADX research system handles liquidity, two-tier allocation, cash-aware sizing, manual execution, and backtests that are allowed to fail.
Qwen 3.8 moved the frontier onto my MacThe ELI5 Opus 4.6 comparison, useful local jobs, and a practical guide to choosing 4-bit or 8-bit for your Mac's RAM.
// CUSTOM SKILLS
Turn a solved problem into callable work.
Custom skills package a method as either a repeatable function or a bounded loop—with named inputs, evidence, limits, output, and an exit condition.
/seo-cluster-audit/release-verification/handoff-packet