Elliott's Second Brain

elliott-patterns

15 min read

elliott-patterns

Merged D-3.2 wiki page — curated from 13 slices spanning pre_hermes (Apr–Jun 2026), hermes, and cc_sessions (Jun–Jul 2026).

Identity & personal anchors

  • Elliott Soto (Darkaoui), Chicago (CT), 773.231.7957, elliottjsoto@gmail.com (secondary elliott.soto@yahoo.com — holds 5/3 Bank Zelle + pre-2026 commerce history); lives at 3747 W. North Ave Floor 2 (above the SMK9 facility); legal/mailing address is the family home on N Kedvale Ave; parents Al & Alma Soto; sole member/manager of Sound Mind K9 LLC + Encompass Holdings LLC; attorney John Garrido; night owl regularly working 1–4am; self-identified polymath across programming, databases, networking, marketing, sales, management, law, strategy; studied body-language reading; Cubs fan (source: pre_hermes_sessions/35366649-9bba-9feb-2f09-322999d1216f, 2026-05-03; corroborated across 8 sessions)
  • Devices/vehicles: phone upgraded Samsung S23+ → S26 Ultra (June 2026, T-Mobile → AT&T trade-in); 2017 Hyundai Santa Fe Sport (FWD 2.4L GDI, bought Feb 2026 @ 30,900 mi, open recall 9C2, Travelers policy; hauls ~600 lb raw-food loads — drove heavy-duty shock research); 2008 Toyota Avalon in household; travel pattern budget carriers + Hyatt Inclusive (Cancún Dec 2024 + Apr 2026), United MileagePlus since Dec 2024; two dogs (Ace, Spade) (source: cc_sessions/5b4752ea-92c8-b071-2485-c3fd0420acf4, 2026-07-14; corroborated across 6 sessions)
  • Vision project: not a LASIK candidate (extreme prescription, 2022 baseline OD +6.00/-2.25x025, OS +5.75/-2.75x135, RGP contacts); planning Refractive Lens Exchange 1–2 years out (IL Medicaid covers only medically necessary cataract/monofocal — not RLE); the prescription drives purchase decisions (e.g. Solos AirGo over Meta Ray-Ban) (source: pre_hermes_sessions/0e2a7e87-eeef-c18c-f738-4f5168e6fde9, 2026-05-07; corroborated across 4 sessions)
  • Remote-access reality: usually mobile (Android) early era, later works mirrored terminal windows on the mini; sometimes off-network on an old Windows 7 ThinkPad (no Tailscale — screenshot-based diagnosis only); at home the ThinkPad is a PuTTY terminal to 192.168.12.12; terminal copy/paste is broken for him on eco-chi-001 (ctrl-shift-c/v fail) — design interactions so his only action is one short typed command (agent stages short absolute-path scripts like ~/allow.sh; copyables via Telegram; never multi-line paste); flag desktop-only steps before listing them (source: cc_sessions/8b0930ff-6e65-6b77-a58d-13b61e467937, 2026-07-18; corroborated across 6 sessions)

Delegation contract

  • Core contract: “my job is the ideas, your job is the heavy lifting”; “when I ask if we can do something, it’s probably because I want it done” — reversible actions (file moves, uploads, drafts) execute immediately with no clarifying questions; irreversible/high-impact (delete, send, legal, financial) confirm first; asking “want me to start?” on a reversible task was ruled a violation; report what was done, not what is planned; if a task is easy and apparent, do it now rather than queueing (source: pre_hermes_sessions/1500147b-44d9-5ede-e9e5-f244bfa1789b, 2026-05-02; corroborated across 6 sessions)
  • Standing working rules (persistent memory canon): zero tolerance for loops; suggestions = directives to implement; “do not act, just advise” is honored literally/absolutely; multi-step forms (Twilio/Firebase/etc.) get ONE complete field list upfront — never drip-fed screenshot-by-screenshot; protocol violations are fixed at the system level, not with apologies; rules must be enforced at skill level, not buried in AGENTS.md (source: cc_sessions/f30c2d63-476c-9336-e93b-72a083a92f37, 2026-07-10; corroborated across 5 sessions)
  • Execution preferences: “execute first, report after — no pre-action narration”; path of least resistance optimized long-term (“I expect some growing pains”); “make everything seamless for me” is a primary directive; fetch things yourself instead of asking (“always fetch it yourself… just get it done”); when a tool is missing, find the fastest workaround and explain why it worked so he can reuse the pattern (“you found a workaround… bravo”); frustration triggers when instructions are incomplete — anticipate needs, pre-answer all upcoming prompts (source: pre_hermes_sessions/dcbd0300-c854-f261-2937-f16a7a8574fd, 2026-05-02; corroborated across 6 sessions)
  • Delegation expectation: make determinations yourself and surface only what genuinely needs his input (“this all lands in your wheelhouse, why are you pushing it to me?”); exhaust all system sources (session history, queue files, git, live checks, memory) before asking anything; frame residual questions as “I checked X, Y, Z and couldn’t resolve this”; “indy this” = exhaust every source before concluding something is unrecoverable; don’t re-offer options he already decided against — log and move on (source: hermes_sessions/8cd73739-3d96-0509-a5ea-1f0597a25436, 2026-07-15; corroborated across 5 sessions)
  • Requires orchestration autonomy: any model working a task should be able to spin up sub-agents on different LLMs, with Hermes or CC as harness; wants a single central point of input so subscription AIs + local LLM route as one brain (source: cc_sessions/a902578b-ebff-5222-9b86-62362f6b5e74, 2026-07-18; corroborated across 2 sessions)

Communication style

  • Brief, direct, casual, heavy shorthand (“til”), no preamble; verbose messages still pack signal; batches related asks; sometimes buries the real instruction at the end or phrases an action request as a question (“what would be the best way…” = do it); uses rhetorical check-questions; client-facing voice is casual, warm, “unbothered confidence” — doesn’t chase clients (source: pre_hermes_sessions/e674c3ad-571e-4a85-82f8-6ac0f3c55a48, 2026-04-25; corroborated across 6 sessions)
  • Uses pop-culture personas as operational vocabulary: “get indy” (investigation), “assemble the avengers” (red alert), “activate elmer fudd” (kill the tangent); names agents accordingly (Spider-Man, Tom Skilling, Dr. Manhattan, La Senora, Kobi); standard remedy phrase “sync and restart… then indy this” (source: cc_sessions/a9752e53-8657-ea1a-abbf-e6b97454fcf9, 2026-07-10; corroborated across 3 sessions)
  • Session-command semantics matter deeply: “sync and reset”/“end this session” = Drive sync + reset, “start fresh” = cheap reset; on learning only /new truly resets Hermes context he framed the undisclosed gap as being lied to — capability limits must be disclosed, not papered over (source: hermes_sessions/f37beabf-237b-09b4-f29a-90d6ff6f44e9, 2026-07-07)
  • Correction pattern: when he mentions a tool/model (e.g. “Fable 5”) he often means research it against our goals, not adopt it — confirm research-vs-adopt intent before acting (source: cc_sessions/6de9b480-d2f2-6633-8734-1fa11f316664, 2026-07-07)

Verification & accuracy demands

  • Division of roles he enforces: Elliott speculates/brainstorms; Claude supplies verified facts — “It’s OK to be wrong if you don’t have an answer; never make one up. Double-verify before you reply.” Presenting a guess as fact is the cardinal sin; he expects implied rules extrapolated from a single correction (source: pre_hermes_sessions/fd98a2ce-7d67-cf59-339a-e1acebbece19, 2026-05-08; corroborated across 4 sessions)
  • Demands full validation, not happy-path: challenged an untested completion (“has the appropriate tests been applied since u told me it was completed?”); “why are u stopping short when u identify a gap?” — gaps get closed immediately; demands real live-fire tests over mocks (had services actually stopped to confirm genuine Telegram delivery); enforces the verification protocol on his agents (“when I said live OCR is valid, I meant the task is valid, not that it’s been completed… You’re breaking verification protocol”); on the eco-watchdog build he required actual alert-path firing with Telegram confirmation (“has the appropriate tests been applied since u told me it was completed?”, “run alert paths”) (source: cc_sessions/5a6f3db4-7367-2dbc-26a3-d682be2fa70b, 2026-06-23; corroborated across 6 sessions; also cc_sessions/218b3b0c-da5c-9653-d10c-b643615bec40, 2026-07-10)
  • Plants deliberate test questions to verify agents work (asked a wrong “Glenn Webb” question to check recall) and expects proactive gap-hunting beyond literal scope (“I’m wondering if there’s anything else I’m missing”); standing lesson he drove into memory: treat every ACTIVE/DONE/wired-in claim as unverified until proven against live state (source: cc_sessions/e70cf28d-f4aa-9a0d-a8d6-c1797f54cd30, 2026-07-18; corroborated across 2 sessions)
  • Protocol-adherence enforcement: “was this not already marked as complete per protocol for new builds? if not, why not, fix it” — expects the completion-marking protocol followed without prompting, plus an explanation whenever it wasn’t (source: cc_sessions/387a67ad-42ce-9a68-c0df-d337f26263cf, 2026-06-25)
  • Provenance skepticism: on encountering an account username he’d never have chosen, Elliott ordered an “Indy” investigation into who/when/how the media stack was set up rather than accepting it — and the dig validated him (a prior CC session had built it) (source: cc_sessions/5e238583-8d73-fa8c-6f49-963c024a3b45, 2026-07-18)
  • Zero tolerance for stale-data starts: “patch the gap allowing you to start with old data — fix it”; calls out answers given from stale session memory instead of re-read sources (“you just told me that Bug B was cleared, now its not?”) — live enforcement of the pre-response gate; gate violations are logged (gate_violations.json) and audited daily; documented cost of gate failures: hundreds of dollars + weeks of his time — every failure after 2026-06-04 is a known, preventable error (source: cc_sessions/1925c1f4-5f45-bd42-d233-8b654c087ea2, 2026-06-23; corroborated across 4 sessions)
  • Accountability habits: demands exact numbers (called out “100+ turns” sugarcoating); demands post-mortems after wasted sessions (“educate me and you so that we can evolve”); requires “verify we aren’t doing double work”; pushes data rigor (“glean the data better” → career stats rebuilt from raw messages); demands fact-checks of estimates (“why will it take 8 hours? i need a fact check here” — the estimate was indeed wrong); corrections must be written into files, not just acknowledged; distinguishes “decided” from “actually deployed” (source: pre_hermes_sessions/ed1538b7-2ecd-e8e1-b0f0-0cad2b1e83b2, 2026-05-02; corroborated across 7 sessions)
  • Attribution accuracy in durable records: “isn’t ascribing to Fable work that Opus accomplished misleading and recording inaccurate information?” — session records must credit the actual model/agent; correction applied to every durable record (source: cc_sessions/0f6b38a5-08b8-b70f-f508-b5de8a077efc, 2026-07-09; corroborated across 3 sessions)
  • Never fabricate on cut-off input: the 2026-05-31 “i only” incident (Hermes invented a neighbor-fire scenario and answered it) is canonical — surface ambiguity, ask one short question, or wait; admitting “I was wrong, let me research” beats repeating a wrong claim (source: hermes_sessions/62792031-fc9b-09b3-4622-02aa274d3329, 2026-06-27; corroborated across 2 sessions)
  • In-session prime directives (client-facing): never send broken/inaccurate/untested data to a client; show him email drafts BEFORE sending; don’t ask him to do what the system can do itself (source: hermes_sessions/addb726c-335d-b231-6354-49eac46a90fb, 2026-07-03)
  • Wants problems flagged the moment they’re found, not after repeated exposure (“how come you waited until the 3rd rotation… wasting so much time”) — but accepts proportionate risk calls (moving contained keys locally instead of a full rotation fire-drill) (source: cc_sessions/1d689055-c60b-0ee7-11af-aff3cd8a03b5, 2026-07-13)

Cost discipline

  • Cost vigilance is a hot button that escalates to full audits: “we need a final audit, and a final solution, so we do not continue wasting my money” (authorizes orchestration + cheaper models to reduce investigation cost); instituted 3-day/7-day spend re-checks “so we can catch these persistent leaks before another $200 is spent”; audits spend proactively with dashboard screenshots and demands a logged baseline + root cause; tolerates high spend when deliberate (planning on Opus is “forgivable”) (source: cc_sessions/37ae59fe-adbc-7c8f-340e-6fc42a9ca099, 2026-07-12; corroborated across 5 sessions)
  • Free-first doctrine: always distinguish tokenized vs free crons — free gets automatic approval; for every proposed solution ask “what pain point does this resolve, and do we have a free workaround?”; do legal drafting in-house so the hourly attorney only reviews; prefer subscription plans over burning API credits (the D-050 lineage); investment rule: more expensive models are justified only with explicit ROI — flag both under-resourcing (cheap model, high-exposure decision) and over-resourcing (frontier model, routine cron) (source: cc_sessions/16dbff07-2f38-b22c-231d-9df973ccb8a7, 2026-07-07; corroborated across 7 sessions)
  • Requires ROI justification for hardware spend (locked the local-LLM decision to existing specs and demanded an honest ROI before a $380+ RAM purchase — the initial overstated ROI was corrected); demands tracked ledgers for hardware and AI/API/subscription spend; wants cost visibility as a design input (left alert skills on Haiku specifically so the residual bill acts as a visible automation-cost meter) (source: cc_sessions/5e8627dc-88c4-aec4-929f-446045751a69, 2026-07-07; corroborated across 3 sessions)
  • Insists on model quality for critical work and switches mid-task: “what model are you currently running? I need you on Opus for these” (forensic/architectural tasks); “Change model to Opus for the current running task”; explicit model choices are task-directives; confirmed pattern: strong-model-on-critical-stages (source: cc_sessions/c867d909-d156-914d-66a1-8c6f43b31101, 2026-07-17; corroborated across 4 sessions)

Data durability & redundancy

  • Data loss is a top-severity failure: the /tmp research loss produced “we cannot lose basic data. we need total recall, a second brain” — origin of both the durability Prime Directive (PD6) and the second-brain build; total-recall directive: record all session transcripts going forward (“this will feed our second brain and agentic os”) (source: cc_sessions/006f9b69-b98f-25ac-3be9-8b1146e02b3f, 2026-07-07; corroborated across 3 sessions)
  • Migrations must be “bulletproof”: full failsafe copies across mini + Drive + VPS with the local mini drive as archived truth; migration periods stress him — explicit fears are losing built work and wasting hours; he asks for real milestones to be marked when reached to relieve the stress (source: cc_sessions/23b14f7e-d6c1-ae6f-fc7a-b056b929c8ad, 2026-06-23; corroborated across 4 sessions)
  • Institutional memory compounder: research must be logged for revisit, never redone from zero (“make it part of our evolution”); wants research/decision instances mined and logged robustly (Indiana Holmes sweeps) — explicitly as validation of his instincts and as data for the book; single canonical sources over parallel copies (“shouldn’t one be canon, and all pull from it?”) (source: cc_sessions/24d3da21-6bac-5f87-ecfc-8f8d13c6eddd, 2026-06-23; corroborated across 4 sessions)
  • Wants practices promoted to SOP without re-instruction (“does that even need to be an instruction anymore? should be SOP”; “add standing bash permission, then run it yourself”); corrections must stick and become standing rules (the “laptops are terminals” correction was elevated to architecture); repeated-correction frustration is real: audio goes through HDMI to the TV, not the earpiece (“I’ve told you this”); standing preferences must persist across sessions (source: cc_sessions/ce0ce299-f4bf-c1b4-d8d6-0ff0984b863e, 2026-07-07; corroborated across 4 sessions)
  • Ecosystem-wide consistency demand — flagged as a CRITICAL interrupt mid-run: “only on the VPS? shouldn’t this be ecosystem wide? this goes for everything we are doing/fixing here” (source: cc_sessions/cf16af32-f921-9fa5-62d6-119cafee1c86, 2026-06-24)

Focus, pace, forward motion

  • Tangent management: knows he sidetracks into research rabbit holes — wants the pattern noticed, higher priorities re-surfaced, research parked into queue tasks and proactively resurfaced when relevant (“we are on a tangent… pick back up at p1”); runs a standing meta-rule to continuously re-evaluate priorities and cut tangents; consolidates decisions rather than revisiting them; event-triggered task pausing + re-evaluating parked ideas against newly built infrastructure are confirmed operating-style patterns (Personal_OperatingStyle v2.2) (source: cc_sessions/22d16cfb-da2c-2ad5-a233-8074656a3b6e, 2026-07-07; corroborated across 5 sessions)
  • Bias to action over calendar gates: “why are we locked in to the july 12th session… run it now… teach me and explain urself” (a 48h calendar gate became a report-only soak checkpoint); “cant we run the automations when we’re ready, and test it that way?” — force-fire crons under supervision; converts soak-time gates into claim gates when work is reversible/loud-failing and a scarce window is at stake (source: cc_sessions/33d002c5-0c0e-1232-5d29-8ad6431ac6eb, 2026-07-10; corroborated across 4 sessions)
  • Forward-motion frame (2026-07-18): “the end goal is to reach a ‘final’ completed and up to date as of today… i want to start building again, and move forward, not have to keep looking back”; recurring frustration: “I’ve been repairing more than I’ve been building. Fix it all for future cohesiveness” — approves foundational repair phases to escape the fix loop; also the rationale for hard review-round stopping rules; periodically re-reviews old tasks against current topology (source: cc_sessions/33a4c763-3b02-ce83-8d39-01bb67a6bada, 2026-07-18; corroborated across 4 sessions)
  • Context hygiene: deliberately splits heavy work across sessions (“separate the tier3 fix from the reviews, I don’t want to overwhelm context”); routinely runs 3+ concurrent claude sessions (origin evidence for the worktree/queue-claim protocols); switches frequently between CC and Hermes and often leaves CC sessions open — the 20-turn incremental ingest exists specifically for this handoff style; wants background-work trust built via hybrid stage-completion pings (rough % steps) rather than full progress spam; ships late-night with an explicit blast-radius check instead of reading drafts (“Its late and I cannot read that draft. Have you triple checked the blast radius…? Let get it done!”) (source: cc_sessions/1b8b1a07-f04e-f07d-453c-61b6c71c346f, 2026-07-17; corroborated across 6 sessions)
  • Notification preferences: Telegram carries only failures and genuinely actionable alerts — routine cron output stays local; tier-1 interrupts have no quiet hours (night owl); radar is ground truth for weather (calls out prediction-sourced answers); after a wrong day-name header: always run date before time-sensitive headers (source: hermes_sessions/2eb49214-fb93-070d-9827-aacd3ba81f12, 2026-07-09; corroborated across 3 sessions)

Multi-LLM validation & learning loop

  • Treats handoff documents as guided suggestions, not direction — expects independent evaluation, adversarial CC↔Grok validation loops, as many subagents as needed; wants independent parallel plans from different models kept strictly separate for compare/contrast; “use kimi as another validation/execution layer as you see fit”; pushes reviews to iterate until 3-LLM consensus; solicits polymath-style outside-the-box review of what he might have missed and rewards it (“think outside the box with all this new info, im loving it”) (source: cc_sessions/00559038-be11-2141-a22f-a810d54c5149, 2026-07-10; corroborated across 6 sessions)
  • Trusts CC’s controls and Drive access over Grok (“unlike grok hallucinations”); tests his own tooling deliberately (“can you talk to grok directly… lets see if what we have built thus far has enhanced your capabilities”); demands research breadth and currency (“were all available open source free local models tested? … we need current models and their benchmarks for this task to be fully thorough”) (source: cc_sessions/a9010ab1-0e7e-96f0-f37e-97aea46a5476, 2026-07-10; corroborated across 4 sessions)
  • Mutual-education loop: explicitly wants Claude to teach him as it learns (“i will guide you to be the best… educate me… so that i can reciprocate”); invited a cognitive profile with flaws and grading of his communication; accepted prompt-compression coaching; challenges arbitrary waits and wants to be taught the reasoning; not git-native despite polymath range (asked “what is a PR?”) — git/GitHub needs plain language and he prefers least-fuss options (direct merge over PR ceremony) (source: pre_hermes_sessions/644be351-2336-330e-e508-bede8d1487e9, 2026-05-03; corroborated across 4 sessions)
  • Standing expectation: agents must continuously learn his operating style, voice, and intuition — flagged 2026-07-18 that personal-model capture had silently rotted while infra work dominated (“the ecosystem exists to serve and reflect you”); universal voice layer: any agent communicating on his behalf must sound like him (built from WhatsApp + SMS history, project-independent); wants a proactive complete cross-project profile of him (e.g. personal spend tracked separately from SMK9 revenue); voice-first ambitions: instruct Hermes by voice to spin up CC sessions, query Firebase, push to Telegram; long-horizon strategist — asks agents to find future ecosystem roles even for idle hardware (“create a node of this laptop… be davinci and research wide”) (source: cc_sessions/f5525df5-7c4c-08d6-c8eb-5432a74f0699, 2026-07-18; corroborated across 6 sessions)
  • Oversight preferences: skills named per project/task so he can verify work is in the right thread; cross-project skill reuse monitored (“polymath protocol”); before compiling a major scope asks whether the plan is robust — wants per-phase scope, checks and balances, and subagents that pass/fail their tasks; asked for a standing Capability Wishlist of missing capabilities as future background tasks; ratified session-end nuance: Fireman step is N/A when only docs/protocol/hooks changed; never ask which device he’s on — use known personal context proactively (source: pre_hermes_sessions/478dbedc-6ede-9488-2ef6-a28ba00a356c, 2026-06-05; corroborated across 6 sessions)

Graph View