chore: initial commit of vault before dashboard work

wiki/sources/2026-07-14-everything-we-knew-about-software-has-changed.md

@@ -0,0 +1,48 @@
+# Everything we knew about software has changed — Theo Browne (AIE)
+
+#source
+
+## Source Metadata
+
+- **Date:** talk at AIE (year not stated in source; references Nov–Dec 2024 as recent past)
+- **Raw path:** `raw/sources/Everything we knew about software has changed.md`
+- **Source type:** conference talk (conclusions/notes), 16:01 — https://youtu.be/xUnRQ9vLXxo
+- **Speaker:** [[theo-browne]] (t3.gg)
+- **Ingestion date:** 2026-07-14
+
+## Core Claims
+
+- Models improve faster than developers can. You can't keep pace by "getting better" — you must **[[think-wider-not-bigger|think wider]]** and pick ideas that feel embarrassingly ambitious. "If your idea doesn't feel stupid, it isn't big enough."
+- Recent releases are **qualitatively distinct eras**, not increments (see [[ai-agent-evolution]]): tool-call → long-running-task → orchestration.
+- You only feel a new model's gain if you **push scope to match it**. Work that made sense on the prior model won't feel different on the new one.
+- Long-tenured engineers carry identity baggage (tools, language, git norms, sunk cost) that now actively holds them back — the "skeuomorphic phase" of development. See [[decoupling-identity-from-profession]].
+- Every project category shifted **down exactly one tier**; a brand-new bottom tier appeared: **a markdown file as executable prose**. See [[code-as-throwaway]].
+- Compete on **breadth**, not just depth — breadth is now viable for small teams. Be the right *shape* so users extend you (Slack effect).
+
+## Key Evidence / Details
+
+- **Three model eras** (speaker's framing/names): tool-call era = "Sonnet 3.5" (first reliable tool calls in a real codebase); long-running-task era = "Opus 4.5" (multi-hour tasks, self-testing; his "AI psychosis" began Nov–Dec 2024); orchestration era = "Mythos / Fable" (understands *itself*, spawns sub-models, verifies afterward — a prompt replaces a "software factory").
+- **iOS skeuomorphism analogy:** iOS 6 imitated physical objects to convince you the phone could replace them; iOS 7 dropped the pretense and got more useful. Developers are in their iOS-6 phase.
+- **Identity symptoms:** terminal fetishism, language-as-identity ("he writes JavaScript"), ceremonial git norms (why *can't* we commit `.env`?), guilt-merging PRs. One gift of agents: no guilt throwing work away.
+- **Tier shift table:** side project → shifts down; startup (his "Ping / Zoom for streamers", YC) → becomes a side project; "too big" (full-stack cloud: Vercel + auth + DB) → becomes a startup; new "too big" is unknown (train your own model? your own OS? compete with npm?).
+- **"G-brain tier" = a markdown file.** He replaced a PR-triage service with a markdown file piped to Codex/Claude on a 9 AM cron; by 9:15 it read four repos, prioritized the day, and shipped a static HTML file to S3.
+- **Breadth vs depth:** old rule — pick a vertical, go deep (Vercel vs AWS). New reality — bolt a serviceable database/platform layer into your product in a day or two of prompting; cover enough that users can *start*. **Slack** accidentally became everyone's agent platform because its *shape* invited extension, not because it's good.
+
+## Connections
+
+- **Entity:** [[theo-browne]]
+- **Concepts:** [[think-wider-not-bigger]] · [[code-as-throwaway]] · [[decoupling-identity-from-profession]] · [[context-as-scarce-resource]] (implied by orchestration)
+- **Timeline:** [[ai-agent-evolution]] (his eras align with Konstantin's tooling timeline)
+- **Related sources:** [[2026-07-14-skills-based-on-git]] (orchestration/harness view from the model-builder side); [[2026-07-14-gap-between-ai-users-irreversible]] (markdown-as-skill echoes G-brain tier); [[2026-07-24-youre-reading-way-too-much-code]] (Theo's follow-up: the *discipline* behind code-as-disposable — [[make-more-cheap-code]])
+- **Tools mentioned:** [[claude-code]], Codex, Vercel, AWS, Slack, npm
+
+## Open Questions
+
+- What is the new "too big" ceiling now that orchestration models exist? (Speaker explicitly doesn't know.)
+- Model names "Mythos / Fable" are the speaker's framing — how do they map to shipped model IDs? (Fable 5 is a real current model; "Mythos"/"Opus 4.5" mapping unverified.) Status: tentative.
+
+## Change Impact on Wiki
+
+- Created concept pages [[think-wider-not-bigger]] and [[code-as-throwaway]]; contributed to [[decoupling-identity-from-profession]] (shared with [[2026-07-14-sebastian-eugene-interview]]).
+- Seeded the [[ai-agent-evolution]] timeline with the three model eras.
+- Added [[theo-browne]] entity.

wiki/sources/2026-07-14-gap-between-ai-users-irreversible.md

@@ -0,0 +1,50 @@
+# In 1 Year, the Gap Between AI Users and Everyone Else Will Be Irreversible — Allie Miller
+
+#source
+
+## Source Metadata
+
+- **Date:** interview (year not stated; references a "2026 Goals doc")
+- **Raw path:** `raw/sources/In 1 Year, the Gap Between AI Users and Everyone Else Will Be Irreversible.md`
+- **Source type:** interview (conclusions/notes), 59:16 — https://www.youtube.com/watch?v=YfRkj9kmQf0
+- **Guest:** [[allie-miller]] (ex-Amazon AI leader; advises OpenAI / Google / Anthropic and Fortune 500s)
+- **Ingestion date:** 2026-07-14
+
+## Core Claims
+
+- The compounding advantage is no longer *which prompt you write* — it's whether you've built a **[[personal-ai-operating-system]]**: persistent context docs, reusable [[skills-as-memory|skills]], and proactive scheduled workflows. In 12 months the gap between someone who invests one focused week and someone still using AI as a chat-box will be **irreversible**.
+- Reframe AI from *intern → first-class teammate*, from *tool I open → OS running 24/7*, from *prompt engineering → [[context-as-scarce-resource|context engineering]]*.
+- Winners vs losers don't differ in expertise — they differ in **mindset**: winners use AI to challenge/augment their thinking and keep agency; losers offload judgment.
+- Build **3 foundation documents first**, then layer skills and proactive workflows on top.
+- The meta-skill of the era is **knowing what good looks like** (taste), not execution skill.
+
+## Key Evidence / Details
+
+- **Her setup:** 36 proactive workflows, ~28 master agents, ~100 total agents on schedules; productivity 2×–10× depending on task.
+- **4 Claude surfaces:** web chat (Q&A, low action) · Cowork (agentic business work, medium) · [[claude-code]] (max control, scheduled tasks, high) · Chrome extension (drives a browser tab). Skills transfer across surfaces and to ChatGPT/Perplexity/Gemini — "just folders with markdown." See [[skills-as-memory]].
+- **Skill abstraction:** a folder with one MD file (*what to do*) + optional resources; skills **compose** (LinkedIn-voice skill calls anti-AI-language skill) and are shared between agents. Built-in skill-creator skill.
+- **3 foundation docs** (spend ~1 hour letting Claude interview you): **Personal Constitution** (values, working style — nothing time-bound), **2026 Goals doc** (annual→quarterly→monthly→weekly), **Core Business Strategy doc** (who you serve + off-website nuance). Outputs jump from generic to ~50% "Allie zone."
+- **How to prompt now:** *"Just complain"* — rambling frustration is rich context Claude turns into proposed skills. Two universal patterns: (1) "Ask me questions before doing this" (interview-then-execute); (2) push back when it refuses.
+- **Proactive workflows:** ~6 AM Morning Brief (top-3 stories ranked to impress your boss, weather+clothing, local events, per-meeting kickoff notes); Friday Email Recap (urgent unreplied emails ranked, with drafted replies). Scheduling is native to Claude Code / Cowork / Codex.
+- **4-tier model of AI work:** Microtask → Companion → Delegate → Teammate. Enterprise super-users get 3–5× and **hoard** the knowledge; treating AI as a *team* asset breaks hoarding.
+- **Trust calibration:** high inside your expertise; low outside (pair with a human); legal → AI first-pass + 15-min lawyer review (cautionary tale: founder fired a contractor on ChatGPT-advised contract reading, got sued, owes more). Ground high-stakes work with retrieval, not raw weights.
+- **12-month predictions:** self-learning models (real weight updates, not the memory-file trick) · "market of one" (every site rendered for you, e.g. Flint/Nike) · agent-to-agent communication. Side effect: **human relationships become more valuable** (echoes [[connections-as-moat]]).
+- **Teams:** Path A (cut headcount 8→2) vs Path B (keep 8, 5–10× output). Income play: AI fluency + diversified income + intelligent frugality; some take a short-term step back to pivot.
+
+## Connections
+
+- **Entity:** [[allie-miller]]
+- **Concepts:** [[personal-ai-operating-system]] · [[skills-as-memory]] · [[context-as-scarce-resource]] · [[connections-as-moat]] (human-relationship side effect)
+- **Related sources:** [[2026-07-14-skills-based-on-git]] (skills-as-memory from the engineering side — strong overlap) · [[2026-07-14-everything-we-knew-about-software-has-changed]] (markdown-as-skill ≈ G-brain markdown tier)
+- **Tools mentioned:** [[claude-code]], Claude Cowork, Codex, Flint
+
+## Open Questions
+
+- What's actually in a good Personal Constitution vs a Goals doc — a reusable template? (Webinar-relevant deliverable.)
+- Are "self-learning models with real weight updates" a near-term reality or aspiration? Status: tentative (speaker's 12-month prediction).
+
+## Change Impact on Wiki
+
+- Created [[personal-ai-operating-system]] and contributed to [[skills-as-memory]] and [[context-as-scarce-resource]] (shared with [[2026-07-14-skills-based-on-git]]).
+- Added [[allie-miller]] entity; extended [[claude-code]] with the 4-surfaces framing.
+- Contributed the human-relationship angle to [[connections-as-moat]].

wiki/sources/2026-07-14-nina-interview.md

@@ -0,0 +1,48 @@
+# Nina Interview — HR Use Cases Validate the Webinar Thesis
+
+#source
+
+## Source Metadata
+
+- **Raw path:** `raw/sources/Nina interview.md`
+- **Source type:** interview conclusions/insights doc (auto-generated by Eugene's record→diarize→transcribe→summarize tool)
+- **Participants:** [[eugene]] (SPEAKER_00, interviewer/webinar author) and [[nina]] (SPEAKER_01, HR recruiter at [[virtido]])
+- **Ingestion date:** 2026-07-14
+
+## Core Claims
+
+- The webinar thesis holds: non-programmers have real, repetitive problems (job descriptions, interview write-ups, LinkedIn sourcing) that small self-built AI tools solve **today**. The old belief "software is slow and expensive" is dead.
+- **Adoption is blocked by friction, not resistance.** Nina's colleagues would use these tools "if it were simple" — the missing piece is a simple all-in-one path, not persuasion.
+- For recruiters, **the transcript matters more than the AI summary** — note-taking during a call wrecks the conversation ("I don't concentrate on the conversation"), especially on 5 a.m. cross-timezone calls.
+- AI is a **multiplier**, not a replacement; "it makes mistakes" is a weak objection when compared with human error rates.
+- The deep payoff is **lower cognitive load**, not just more output ("even if your output doesn't change… your life overall gets better").
+- [[skills-as-memory|Skills]] turn personal expertise into a transferable asset: do the task through AI, correct it, then freeze it into a skill a junior hire can run ([[solve-first-then-skillify]]).
+- Agents driving a **real browser** beat paid sourcing tools (LinkedIn Sales Navigator end-run), with anti-bot risk at team volume.
+- Local + subscription beats API/SaaS for this tool class: no per-call cost, no licensing, no login/security surface.
+
+## Key Evidence / Details
+
+- **The demo tool:** one red button → record → diarize + transcribe + analyze → structured report. ~10 min to process 1 hour of audio, locally on GPU, on a Claude subscription. (This very document is its output format.)
+- Sourcing demo: "Computer Vision companies under 200 people, most senior reachable contact, save profile link" — runs in the background.
+- Nina could go from ~5 to ~25 candidates/day; multilingual output (EN/DE/UK) is a genuine unlock for her.
+- Practice notes: one responsible agent per project (Eugene ran 7 in parallel); Karpathy-style Obsidian knowledge base to be shared as a post-webinar "gift"; consent before recording candidates.
+- Webinar: first one in English, titled *"From a chat box to your own operating system"*, aimed at non-programmers.
+
+## Connections
+
+- Entities: [[eugene]], [[nina]], [[virtido]], [[inspectron]], [[claude-code]]
+- Concepts: [[skills-as-memory]], [[solve-first-then-skillify]], [[personal-ai-operating-system]], [[code-as-throwaway]] (the "software is slow and expensive is dead" claim), [[context-as-scarce-resource]]
+- Companion interview: [[2026-07-14-yulia-interview]] (same HR team, overlapping pain points)
+
+## Open Questions
+
+- How to bridge Eugene's programmer-grade demos to a non-programmer's on-ramp — he admits simple examples "hang in a vacuum."
+- Can transcribe→summarize plug into **Manatal** (the team's ATS) and sync across recruiters (e.g., via git)?
+- Will LinkedIn flag automated browsing at HR-team volume? Threshold unknown.
+- Are skills personal IP or employer work product? Unresolved between Eugene and [[sebastian]].
+- Larysa (project management) not yet interviewed — PM use cases missing.
+
+## Change Impact on Wiki
+
+- Created [[nina]], [[inspectron]]; created concept [[solve-first-then-skillify]].
+- Updated [[eugene]] (identity evidence strengthened: webinar author, Inspectron, tool builder), [[virtido]] (HR-team vantage point), [[skills-as-memory]] (skills as handoff/de-risking), [[overview]] (HR-practitioner lens added to through-line).

wiki/sources/2026-07-14-sebastian-eugene-interview.md

@@ -0,0 +1,50 @@
+# Sebastian Interview — How AI Is Reshaping Software Engineering
+
+#source
+
+## Source Metadata
+
+- **Date:** pre-webinar conversation (date not stated)
+- **Raw path:** `raw/sources/sebastian interview - conclusions and insights.md`
+- **Source type:** interview (conclusions/insights), ~56 min. Built from a Whisper large-v3 + speaker-diarized transcript; quotes lightly cleaned.
+- **Participants:** [[sebastian]] (founder of [[virtido|Virtido]], ~11-yr outsourcing firm) and [[eugene]] (interviewer; CV/embedded developer, content creator, harness-builder)
+- **Ingestion date:** 2026-07-14
+
+## Core Claims
+
+- The debate over *whether* AI can write software is over — the whole game is **how you use it**. Cost of code → zero (see [[code-as-throwaway]]).
+- **Teams shrink and roles merge:** the 8-person scrum team collapses to 2–3 who share coordination and agent-wrangling. See [[seniority-and-the-junior-squeeze]].
+- **[[harness|"Bring your own harness"]] hits an enterprise wall** — compliance/liability force a company-managed resource. *That gap is the business opportunity.* See [[enterprise-ai-reality]].
+- Pure coding skill is **leveled** by AI (20-yr veteran ≈ fresh grad output); judgment, [[product-ownership|ownership]], and **human [[connections-as-moat|connections]]** rise in value.
+- **[[connections-as-moat|Relationships are the last non-commoditized asset]]** — the standout insight of the interview.
+- Learn to **[[decoupling-identity-from-profession|decouple identity from profession]]** or "we will feel worthless in a couple of years."
+
+## Key Evidence / Details
+
+- **The question changed:** "It's not a question *if* AI can write software anymore — it's just a question of *how* you use it."
+- **Team collapse:** no scrum master + PM + requirements engineer + big dev team — instead one coordination/ownership role + one or two agent-managers sharing roles.
+- **Harness at enterprise scale:** Eugene demoed his own (Telegram-like UI, one agent per project, inter-agent messaging, per-agent memory, "done thinking" signal; tools: **Conductor** for git-worktree isolation + auto PRs/merges). Sebastian's counter: "bring-your-own-harness will not be the way forward… it has to be a company-managed resource." Enterprise reality: engineers on centrally-managed VMs, zero self-install (Roche SAP transformation ~1,200 engineers for years; banks moving from banned → cautious adoption "because it's just so good").
+- **Seniors up, juniors out:** seniors know *where things go wrong* ("AI does the same mistakes humans do because it's trained on our mistakes"); the junior "yes… yes… allow for all future" habit is how "API keys are leaked, databases get dumped." "Give a junior… access to this almighty Claude and… the codebase — they will [wreck] it in two days."
+- **Ownership story:** an engineer implemented "change your photo," ticked every acceptance criterion, but shipped it ugly (visible in the corner) because they never looked at the result. Reframe: stop thinking "what needs to be done" (tickets); think "what problem needs to be solved." "No one ever needed a programmer… people have problems that you are solving."
+- **Planning debate:** colleague **Daniel** floated returning to waterfall; Sebastian disagrees — coordination overhead now exceeds the work. On a 2-person, ~1-day/week project he's *faster alone*. **Printer anecdote:** unfamiliar direct-printing protocol in Java (never written), 2-day deadline — went on-site, Claude Code solved it in ~30 min; his edge was *knowing how to instruct and verify*.
+- **Connections (the moat):** in-person 2–4 days/week (lunches, events, conferences) wins business; sales agencies / cold calling / email / LinkedIn campaigns / content / SEO = "Big zero." A connection forms on the *second* meeting in *different* circumstances → recognition → trust → referrals. Be memorable in your **humanness** (renovating a house, two kids, a cat) — because everyone's AI output looks identical (riff: Virtido's `humans.verti.com` "human badge").
+- **Other points:** open source grows as code becomes ~free (Eugene's cynical read: OSS is largely marketing); legacy niches persist (COBOL in banks — no training data); outsourcing economics — low-cost expectation selects for bad code ($10k/mo → excellent people; $800/mo won't beat a good hire).
+
+## Connections
+
+- **Entities:** [[sebastian]] · [[eugene]] · [[virtido]] · [[claude-code]] · [[hermes]] (Eugene references it among his tools)
+- **Concepts:** [[harness]] · [[enterprise-ai-reality]] · [[seniority-and-the-junior-squeeze]] · [[product-ownership]] · [[connections-as-moat]] · [[decoupling-identity-from-profession]] · [[code-as-throwaway]]
+- **Related sources:** [[2026-07-14-skills-based-on-git]] (harness definition + BYO-harness from the practitioner side) · [[2026-07-14-everything-we-knew-about-software-has-changed]] (identity baggage, code-as-throwaway)
+- **Raw reference (not ingested):** `raw/sources/Ideas for webinar.md` echoes many of these (harness, connections, "describe problems not waterfalls," Daniel, HR search demo).
+
+## Open Questions
+
+- How does AI transform *huge* (~1,200-engineer, multi-year) enterprise programs? (Sebastian: doesn't know.)
+- How do you unify wildly different personal workflows into one company process? (Eugene + collaborator spent a month, failed.)
+- Concretely, how does an individual build a connections network from a standing start? (Principles offered; step-by-step unresolved — Eugene's ~6-month blocker.)
+
+## Change Impact on Wiki
+
+- Created [[harness]], [[enterprise-ai-reality]], [[seniority-and-the-junior-squeeze]], [[product-ownership]], [[connections-as-moat]]; contributed to [[code-as-throwaway]] and [[decoupling-identity-from-profession]].
+- Added entities [[sebastian]], [[eugene]], [[virtido]].
+- [[eugene]] noted as the likely vault owner / webinar author (Status: tentative).

wiki/sources/2026-07-14-skills-based-on-git.md

@@ -0,0 +1,53 @@
+# Git-Based Skills — the New Memory of AI Agents (Konstantin, Sber)
+
+#source
+
+> Russian-language source. Summary in English; key terms preserved.
+
+## Source Metadata
+
+- **Date:** talk (year not stated; positions "2026 (now)" and forecasts 2027)
+- **Raw path:** `raw/sources/Скиллы на базе git — новая память AI-агентов.md`
+- **Source type:** conference talk (conclusions), 53:09 — https://www.youtube.com/watch?v=a-NIeMB-Hj8
+- **Speaker:** [[konstantin]] (Sber, GigaChat dev team, agent R&D)
+- **Ingestion date:** 2026-07-14
+
+## Core Claims
+
+- **Skills + user data + auto-improvement = the new memory of AI agents**, the architecture all agentic systems will converge on. Tools (how to call), data (what you work on), and usage history live together in one **git repo**; the [[harness]] decides what to load into context.
+- **Simplicity beats complexity:** a universal agent with ~10 file-ops functions outperforms elaborate multi-agent graphs. The era of **universal agents = [[harness|harnesses]]**.
+- Agent tooling evolved **tools → MCP → skills** (see [[evolution-of-agent-tooling]]); skills' two-stage loading lets you keep *hundreds* of skills without blowing context.
+- The classic skill is "**von Neumann without data**" — add data + usage history into the same git repo.
+- **[[agentic-loops|Agent loops]]** are the next paradigm (forecast: dominant in 2027): inner (ReAct) → outer (Ralph) → meta.
+
+## Key Evidence / Details
+
+- **Harness = harness/upness metaphor:** LLM = the force, tools = the harness, data = the field; the LLM "pulls the tools across the data space." Different tasks, *same* processing method. Standard toolset ~30–40 max (Read/Edit/Write, Bash, Grep/Glob/WebSearch, TodoWrite, subagent spawn); >100 tools = even Fable 5 gets confused. See [[harness]].
+- **Human role evolution:** prompt-engineer → context-engineer → harness-builder → **loop-engineer** (CI/CD, back-pressure, meta-loops). See [[context-as-scarce-resource]].
+- **Hackathon proof:** 19 of 20 top teams at BitGen built on harnesses; his own harness placed 7th (Interpress-Ex, run in an infinite loop over a weekend — *read zero tasks by eye*) and 3rd overall / 1st on technical metrics (Snowbase, leaked Claude Code as backend).
+- **tools → MCP → skills:** Tools (2022–24) each consume context; MCP (Anthropic, late-2024→25) adds ready tools via protocol but many servers = dozens of tools = context blowout + network-load security risk (he doesn't hate MCP; issues largely solved). **Skills (2025→)** = directory + gentleman's agreement: `SKILL.md` + tools + (his proposal) **data**. Two-stage loading: short description always in system prompt; long description loaded only when the agent enters the dir. Hermes grew 56 → 100 skills and still works. Cost: MCP needs a developer; a skill needs someone who can write text (any language). Claude Code turns MCP into skills, erasing most MCP downsides. See [[skills-as-memory]] and [[evolution-of-agent-tooling]].
+- **Why git:** skill-first (many agents on one skill — Claude Code, Hermes via Telegram, OpenClaude anywhere) · CI/CD integration · team sharing · merge-conflict resolution now done by models (parallel actors on bare git, no transactional DB — "not for payments, but for much else, yes"). **Mandatory AGENTS.md rules:** `git pull` at start, `push`+`commit` at end, **CI as back-pressure** (returns a drifting agent to spec).
+- **His daily skills (with data):** DNA/medicine (~100 GB sequencing for the whole family; Claude analyzed an undescribed mutation, found the gene, called **AlphaFold**, rendered the protein, concluded the fold is fine — own full genome for ~$1000 and an evening vs $3B/13 yrs historically) · recommendations (~1000 ratings) · trip planning (photo a ticket in Telegram → agent git-pulls, files the PDF, commits; "hotel not booked, 2-hr layover — don't even try") · HR/résumé scoring (self-corrected over 10–20 feedback iterations to ~10% error) · finance · team agent-R&D.
+- **Auto-improvement (Hermes):** >5 tool calls on a task → candidate skill. Curator: prune unused at 30 days (deactivate) / 90 days (archive); consolidate pairs weekly so short-descriptions don't eat context. Memory is now skills, not RAG-extracted facts.
+- **Agent loops:** inner = ReAct. **Outer/Ralph loop** (Geoffrey Huntley): `while true; do claude -p "solve" || true; done` — runs for days, stays in the **"smart zone"** (first ~⅓ of context) avoiding summarization decay. **Meta loop** (his own): periodically wipe all agent-created data to an unseen archive and restart, dodging Karpathy's **collapse** (same solution reworded) — good for research tasks.
+- **Auto-agent result:** weak GigaChat went from **1/89 → 11/89** benchmark tasks over a weekend of self-improvement (11× gain, verified no overfit). Idea from Karpathy's "autoagent."
+- **Missing piece:** perception / qualia (the "walk vs drive to the car wash" example); maybe world-models — open question.
+
+## Connections
+
+- **Entity:** [[konstantin]] · tools: [[claude-code]], [[hermes]], Codex CLI, OpenClaude, Cursor, Conductor, AlphaFold
+- **Concepts:** [[skills-as-memory]] · [[harness]] · [[evolution-of-agent-tooling]] · [[agentic-loops]] · [[context-as-scarce-resource]]
+- **Timeline:** [[ai-agent-evolution]]
+- **Related sources:** [[2026-07-14-gap-between-ai-users-irreversible]] (skills = markdown folders; strong overlap) · [[2026-07-14-sebastian-eugene-interview]] (harness, BYO-harness) · [[2026-07-14-everything-we-knew-about-software-has-changed]] (orchestration era, markdown-as-service)
+
+## Open Questions
+
+- What data belongs in a skill, and is there an upper limit? (No standards yet; he loads 200 GB in one skill, 100 KB in another — "haven't found the ceiling.")
+- How do you build world-models / give agents perception? (Open.)
+- Are the pruning windows (30/90 days) and the >5-tool-call threshold generalizable, or Hermes-specific heuristics? Status: tentative.
+
+## Change Impact on Wiki
+
+- Created [[skills-as-memory]], [[agentic-loops]], [[evolution-of-agent-tooling]]; contributed heavily to [[harness]] and [[context-as-scarce-resource]].
+- Added entities [[konstantin]] and [[hermes]]; extended [[claude-code]].
+- Extended the [[ai-agent-evolution]] timeline with the tools→MCP→skills and loop generations.

wiki/sources/2026-07-14-yulia-interview.md

@@ -0,0 +1,48 @@
+# Yulia Interview — Levels of AI Usage & the Candidate Knowledge Base
+
+#source
+
+## Source Metadata
+
+- **Raw path:** `raw/sources/Yulia interview.md`
+- **Source type:** interview conclusions/insights doc (auto-generated summary; speakers unnamed in text)
+- **Participants:** [[yulia]] (SPEAKER_00, HR/recruiting lead — name inferred from filename, Status: tentative) and [[eugene]] (SPEAKER_01, senior programmer ~20 years, works on [[inspectron]], $200 Claude max plan)
+- **Ingestion date:** 2026-07-14
+
+## Core Claims
+
+- **AI usage has distinct levels, and most people are stuck at level one:** web chatbot → built-in memory → Claude Code/Cowork with local files → CLAUDE.md → skills → Obsidian knowledge base → RAG. The practical ceiling for non-programmers is **CLAUDE.md + skills** ([[levels-of-ai-usage]]).
+- **Solve first, skill-ify after** — the recurring mistake is building the skill up front; instead reach the final solution once with Claude, then say "now create a skill from this" ([[solve-first-then-skillify]]). Turn any correction loop longer than ~3 messages into a skill.
+- **Both recruiting problems are the same problem:** a candidate database plus search over it — record interviews, auto-transcribe with diarization, generate standardized profiles, then query by skill.
+- **AI automates all information work; the human keeps the human parts** — "the human's role in this process is just to be human — introductions, communication." (Converges with [[connections-as-moat]].)
+- **AI prices will rise, not crash** — "what I now buy for 200 will cost about 1,000"; the bubble pops *upward*, and leverage shifts to knowing which models to use. Status: tentative (prediction).
+- **Renting GPU beats owning** for this workload: ~$0.10/candidate on RunPod vs a $1,500–2,000 machine — though a home GPU transcribes overnight for free.
+
+## Key Evidence / Details
+
+- Live demos: RunPod transcription; Claude driving a browser to pull LinkedIn HR contacts into a Markdown list. LinkedIn bans bots — only shallow passes (~20 contacts) are safe.
+- Immediate no-tooling practice change: **start recording candidate interviews now**, one file per candidate, named by candidate (consent implied per [[2026-07-14-nina-interview]]).
+- Stack recommendation: Claude (Cowork for non-programmers); RAG only matters at corporate scale; cheaper models (e.g. Qwen) do the same tasks "just worse."
+- Prompting principle: your job is to "narrow the variability of interpretation" — vague asks invite "fantasy."
+- Reusable framing: "Split yourself into two people — one slightly smarter (you, the manager), one slightly dumber (the AI, your subordinate)."
+- Deal context: webinar plan to be drafted over a weekend, **no deadline**; a paid build of the HR system is on the table (rough "$5 to $100" per-project range needs scoping), possibly a shared monthly service on Eugene's subscription.
+- Company-scale idea: a lightweight AI "assistant" chatting with employees to form an information core — automating much of the PM/status-reporting role.
+
+## Connections
+
+- Entities: [[yulia]], [[eugene]], [[nina]] (referenced), [[inspectron]], [[claude-code]], [[virtido]] (Yulia's team affiliation tentative)
+- Concepts: [[levels-of-ai-usage]], [[solve-first-then-skillify]], [[skills-as-memory]], [[connections-as-moat]], [[personal-ai-operating-system]], [[context-as-scarce-resource]]
+- Companion interview: [[2026-07-14-nina-interview]] (same team's pain points from the recruiter's seat)
+
+## Open Questions
+
+- Exact scope, deliverables, and price of the HR system build.
+- Webinar date/title (later fixed as *"From a chat box to your own operating system"* per [[2026-07-14-nina-interview]] — this interview likely predates it).
+- Rent GPU vs. buy a GPU machine — folded into Yulia's next hardware purchase decision?
+- Larisa's Claude memory complaint — fixable via built-in memory/CLAUDE.md, but her actual BA/PM task was never detailed.
+- Is SPEAKER_00 really named Yulia, and is she at [[virtido]]? Inferred from filename + references to Nina/Sergiy. Status: tentative.
+
+## Change Impact on Wiki
+
+- Created [[yulia]]; created concepts [[levels-of-ai-usage]] and [[solve-first-then-skillify]].
+- Updated [[eugene]] (Inspectron, $200 plan, paid-build offer), [[connections-as-moat]] (Eugene independently lands on the human-connection residual), [[skills-as-memory]] (~3-message correction-loop heuristic), [[overview]].

wiki/sources/2026-07-21-larysa-interview.md

@@ -0,0 +1,51 @@
+# Larysa Interview — Memory Loss, Phantom Integrations, and the Case for Skills
+
+#source
+
+## Source Metadata
+
+- **Raw path:** `raw/sources/Larysa interview.md`
+- **Source type:** interview conclusions/insights doc (auto-generated by Eugene's record→diarize→transcribe→summarize tool)
+- **Participants:** [[eugene]] (SPEAKER_00, webinar host) and [[larysa]] (SPEAKER_01, technical BA/PM, ex-mobile developer)
+- **Ingestion date:** 2026-07-21
+- **Note:** closes the "Larysa not yet interviewed / PM use cases missing" gap opened in [[2026-07-14-nina-interview]] and [[2026-07-14-yulia-interview]].
+
+## Core Claims
+
+- **The friction with agentic AI is structural, not accuracy.** Larysa's blockers are not hallucination but (1) no durable memory across sessions, (2) integrations that look available and aren't, (3) over-eager drift on loosely specified tasks.
+- **Agent memory, as shipped, is a net negative.** Eugene is blunt: "Memory is the worst thing agents have — it gives no benefit and confuses users to hell… it'd be better if it didn't [exist]." The replacement is [[skills-as-memory|skills]], not a better memory feature.
+- **Claude starts work against capabilities it doesn't have.** It accepts a task, then discovers mid-way that a connector isn't implemented or isn't available for a *private* (vs. corporate) account — after half an hour of the user's time and tokens are spent. See [[integration-dead-ends]].
+- **Consolidate, don't tool-hop.** Eugene's central pitch: one personal "operating system" where all project agents live together and can talk to each other beats spinning up Claude in separate folders — "my life just split into before and after."
+- **Leave less room for imagination.** Under-specified prompts invite drift, and the drift causes collateral damage you won't notice. Standardized procedures (skills) are the constraint. See [[leave-less-room-for-imagination]].
+- **AI-assisted coding is production-ready as of ~Claude 4.6, with carve-outs.** "Code isn't something elite anymore… though authorization and payments I still wouldn't trust to Claude."
+- **Diarization + role inference is the differentiator** in interview automation over plain transcribe-then-summarize.
+
+## Key Evidence / Details
+
+- **Larysa's actual usage** (unusually deep for a non-engineer role): opens pull requests, wires Claude to **ClickUp** (Jira analog) and **Figma**, builds clickable prototypes, writes user stories and tests.
+- **The dead-end example:** a Teams bot to watch a work chat and auto-create ClickUp tasks. Verdict — Teams is effectively a dead end (Microsoft exposes no usable external API); Slack has first-class Claude support from Anthropic but needs a paid seat she isn't given. Figma is blocked the same way: no personal paid account → no full editable access.
+- **Her quote on the failure mode:** "That's not hallucination exactly… it just first goes looking for what you don't have. I don't know how to fight this." — left explicitly unresolved.
+- **Model choice:** Eugene runs **Claude 4.7** in production, finding **4.8 "too proactive"** ("without the flights of fancy 4.8 has").
+- **Testing note:** prefer browser/web testing over emulator testing with Claude — the emulator got badly stuck.
+- **Tool convergence:** both participants independently landed on recording and auto-processing interviews. Eugene's system adds speaker diarization plus *role* inference from conversation context ("Speaker 0, HR Equity Lead; Speaker 2, Senior Programmer") — ~10–11 min for a one-hour interview. This document is its output.
+- **Webinar commitment:** [[skills-as-memory|skills]] are the one topic Eugene commits to covering for Larysa.
+
+## Connections
+
+- Entities: [[eugene]], [[larysa]], [[claude-code]], [[inspectron]], [[virtido]] (Larysa's affiliation tentative)
+- Concepts: [[skills-as-memory]], [[integration-dead-ends]] (new), [[leave-less-room-for-imagination]] (new), [[harness]], [[levels-of-ai-usage]], [[solve-first-then-skillify]], [[code-as-throwaway]], [[personal-ai-operating-system]]
+- Companion interviews (same webinar-scoping series): [[2026-07-14-nina-interview]], [[2026-07-14-yulia-interview]]
+
+## Open Questions
+
+- How to reliably pre-empt integrations Claude recommends that aren't available for the user's account type — stated as unsolved by both participants.
+- Is a Teams-triggered auto-task bot feasible at all given Microsoft's API surface? (Asserted dead end, not demonstrated.)
+- How to get full editable Figma access without a personal paid account.
+- Is 4.8's "over-proactivity" a model property or a prompt-specification problem? (It reads as the same drift [[leave-less-room-for-imagination]] blames on loose specs — Eugene treats it as the model's fault.)
+- Larysa's employer is never stated; [[virtido]] is inferred from the interview series. Status: tentative.
+- Webinar structure and scheduling — deferred to a follow-up with [[yulia]].
+
+## Change Impact on Wiki
+
+- Created [[larysa]] (entity); created concepts [[integration-dead-ends]] and [[leave-less-room-for-imagination]].
+- Updated [[skills-as-memory]] (memory-as-anti-feature: the negative case for why skills exist), [[harness]] (consolidation-over-tool-hopping claim), [[code-as-throwaway]] (the auth/payments trust carve-out), [[levels-of-ai-usage]] (Larysa as a high-rung user still missing the skills rung), [[solve-first-then-skillify]], [[claude-code]] (4.7 vs 4.8; memory and connector limits), [[eugene]], [[yulia]], [[virtido]], [[overview]].

wiki/sources/2026-07-22-ai-is-stupid.md

@@ -0,0 +1,47 @@
+# ИИ глупый! (AI Is Stupid!)
+
+#source
+
+## Source Metadata
+
+- **Date:** 2026-07-22 (ingestion date; publication date unknown)
+- **Raw path:** `raw/sources/ИИ глупый!.md`
+- **Source type:** Conclusions doc for a Russian-language YouTube Short (1:28) — https://www.youtube.com/shorts/P4eWd2jvz4k — titled «ИИ глупый! #ии #ai #бизнес»
+- **Author:** unknown (business-facing Russian-speaking creator). Status: tentative.
+- **Ingested:** 2026-07-22
+
+## Core Claims
+
+1. AI looks "stupid" not because the model is weak, but because it is systematically starved of two things: **context** (memory of *your* specific business) and a **harness** (the rules the model must reason by). Without them, even the strongest model answers like an outside expert, not like your employee.
+2. **Intelligence without context loses to context without intelligence.** Analogy: ask "how will our sales month close?" — ten Nobel laureates can only cite industry averages ("~5% up/down across Russia"), while a rank-and-file employee of your company answers better, because they see your funnel, clients, seasonality, and deals.
+3. The formula: **strong model + your business context + harness = employee-level answer.** Remove any component and you get "smart but generic," "specific but undisciplined," or "stupid AI."
+4. Giving the model your company's context improves answer quality "by orders of magnitude" (author's hyperbole: "tens of times, maybe a million").
+5. Investment advice: before swapping to a "smarter" model, invest in **context infrastructure** (data, memory, integrations — the video names RAG, long-term assistant memory, CRM/ERP/document ingestion) and in the **harness** (rules, checks, tooling). That is where the ×10…×1000 gains are, not in model version bumps.
+6. Maturity metric for an AI rollout: *can the model answer a question about your business more accurately than an outside consultant?* If not, context or harness is missing.
+
+## Key Evidence / Details
+
+- The harness is defined here as more than a prompt: an "engineering wrapper" — what the model must verify, which tools to trust, how to shape the answer, what is forbidden — turning the LLM from "an encyclopedia of hospital averages" into a **procedural agent**.
+- Target audience per the video: business owners and product managers disappointed by generic LLM answers; engineers building corporate assistants; anyone choosing between "get a bigger model" and "give the model the right data and rules."
+
+## Connections
+
+- [[harness]] — the video's second ingredient is exactly the vault's harness concept, restated for a business audience; independent convergence with [[konstantin]] and [[eugene]].
+- [[context-as-scarce-resource]] — complements it from the *supply* side: the vault's page says context is the binding constraint inside the window; this source says the default failure is not providing business context at all.
+- [[personal-ai-operating-system]] — the Nobel-vs-employee analogy is the business version of Allie's "feed the system who you are" (foundation docs).
+- [[skills-as-memory]] — tension: this video names **RAG** as a practical context mechanism, while [[2026-07-14-skills-based-on-git]] argues skills-as-memory beats RAG (see Open Questions).
+- [[levels-of-ai-usage]] — "stupid AI" is what the bottom rungs of the ladder feel like; the formula names what the upper rungs add.
+
+## Open Questions
+
+- Who is the author, and is the short connected to anyone already in the corpus (its context+harness framing matches the corpus suspiciously well)? Status: tentative.
+- The video treats RAG and long-term memory as go-to context mechanisms; Konstantin's source argues skills beat RAG (harness loads on activation instead of pre-injecting). Is the difference audience-driven (business data vs procedures), or a real disagreement?
+- The ×10…×1000 improvement claims are rhetorical, not measured.
+
+## Change Impact on Wiki
+
+- Created this source page.
+- [[harness]] — added the business-facing definition ("engineering wrapper"), the three-part formula, and this source as independent convergent evidence.
+- [[context-as-scarce-resource]] — added the supply-side facet ("intelligence without context loses to context without intelligence"), the Nobel-vs-employee analogy, and the RAG tension.
+- [[overview]] — source count 7 → 8; convergence note (no change to the spine — this source *restates* it).
+- `index.md`, `log.md` updated.

wiki/sources/2026-07-24-youre-reading-way-too-much-code.md

@@ -0,0 +1,48 @@
+# You're reading way too much code — Theo Browne
+
+#source
+
+## Source Metadata
+
+- **Date:** YouTube video (publication date not stated in source), 24:11 — https://www.youtube.com/watch?v=434cG4g5KLE
+- **Raw path:** `raw/sources/You're reading way too much code.md`
+- **Source type:** video essay (conclusions/notes)
+- **Speaker:** [[theo-browne]] (t3.gg / T3 / Lakebed)
+- **Ingestion date:** 2026-07-24
+
+## Core Claims
+
+- The problem isn't reading too much code in absolute terms — it's reading too high a *percentage*, because engineers aren't **generating enough code that doesn't matter**. "I'm not telling you to make your code cheaper. I'm telling you to make more cheap code." See [[make-more-cheap-code]].
+- **The importance-of-code spectrum** runs from "slop website with one viewer" to "pacemaker firmware." The AI-code debate stalls because people (a) overestimate where they sit on it and (b) project every claim they hear onto their own tier.
+- **Four tiers of code** (A slop / B "I'd like this to work" / C "trouble if it breaks" / D death tier), each with its own reading discipline — from "never read it" (A) to "read every line" (D). Nobody works 100% in D.
+- **Steelman granted, then inverted:** if your production code could kill someone, that's a reason to read every shipped line *and* to generate 100–10,000× more throwaway code to verify it. "Every line of code that goes in should have 100 lines of slop verifying it."
+- **There is always another layer:** if the core is too important for AI, abstract up — custom debuggers, runtimes, lint rules, load rigs. If the verification tools are too important, build tools that introspect *them*.
+- What changed: writing is nearly free, **reading still costs attention** — so the generated-to-read ratio should skyrocket while hand-verification of shipped code stays the same or increases.
+- Explicit anti-endorsements: merging unreviewed slop, reducing verification of shipped code, "move faster" as an end. "I hate them too. We're on the same side" (re: vibe-coders shipping slop).
+
+## Key Evidence / Details
+
+- **Ratio shift (his own numbers):** pre-AI ~1,000 read / 200 written / 100 merged per day; now ~1,000 read / 2,000+ written / ~500 merged — the unmerged bulk exists to test ideas and is never reviewed. Reframing chart for a tier-D engineer: read goes *up* in absolute terms (200→400), slop goes 0→800+, merged stays ~100.
+- **Shao's ratio (cited approvingly):** ~80% of AI-generated code goes into test harnesses and guardrails, with harness results fed back into the LLM's context — fewer mistakes overall.
+- **Concrete slop patterns:** on-demand custom lint rules for a found bug pattern; one-off debuggers; slop-porting a Go service to Rust just to benchmark the test suite; **10 dumb-model agents building on a fresh API — their failures are UX bugs in the API**; agent with AWS access spinning up throwaway load rigs; testing 3 theories of an ambiguous PR in parallel; bespoke one-off test suites.
+- **Tier-C reading tricks (via a Dax post):** after a big diff, ask the agent for a per-file summary instead of reading the diff — "anything weird will stick out"; read function signatures and API definitions always, care less about bodies. Theo's Lakebed version: read every signature/API, use slop (the 10-dumb-agents test) to verify them.
+- "Writing 10,000 lines of code to organize 100 files is mental illness until the code is free to generate. Then all of a sudden it's totally reasonable."
+- Closing principles: code is useful for things other than shipping; draw the ship/no-ship line and apply discipline separately per side; if you can't find slop opportunities "you're not creative enough"; don't read faster — read only what's worth reading; have AI review code before humans do.
+
+## Connections
+
+- **Entity:** [[theo-browne]] (his second source in the vault)
+- **Concepts:** [[make-more-cheap-code]] (created from this source) · [[code-as-throwaway]] (this source answers its open question: the durable artifact includes the *verification harness*) · [[think-wider-not-bigger]] (same author-move: reframe a human-side habit as the bottleneck) · [[leave-less-room-for-imagination]] (tension: agent-summary-instead-of-diff vs Eugene's invisible collateral damage — logged there) · [[product-ownership]] (verifying as the human's remaining job) · [[context-as-scarce-resource]] (reading costs attention — the human-side analog)
+- **Related sources:** [[2026-07-14-everything-we-knew-about-software-has-changed]] (his earlier talk: code disposable, kill without guilt — this source supplies the *discipline* for that stance); [[2026-07-21-larysa-interview]] (Eugene's trust carve-out and invisible-drift worry — the audience this source's tier framework speaks to)
+- **Tools/people mentioned:** Codex, AWS, Grok-tier models, Dax, Shao, Lakebed
+
+## Open Questions
+
+- Shao's "~80% into harnesses, fewer mistakes overall" is cited without data — is there any measured before/after? Status: tentative.
+- The tier framework assumes you can cleanly separate ship-code from slop; where do skills/prompts/CLAUDE.md files (which *do* persist and steer future generations) fall on the A–D spectrum?
+- Does the "agent per-file summary instead of the diff" practice survive Eugene's objection that drift is precisely what a summary hides? (Logged as a tension in [[leave-less-room-for-imagination]].)
+
+## Change Impact on Wiki
+
+- Created [[make-more-cheap-code]] (concept — throwaway generation for verification/exploration; four tiers; always-another-layer).
+- Updated [[code-as-throwaway]] (verification as the durable use of cheap code; partial answer to its "what is the durable artifact" question), [[theo-browne]] (second source; sharpened stance — explicitly anti-slop-shipping), [[leave-less-room-for-imagination]] (new logged tension on diff summaries vs invisible drift), [[2026-07-14-everything-we-knew-about-software-has-changed]] (reciprocal related-source link), [[overview]] (8→9 sources; "verifying" half of the spine deepened).