Robert García

WhAIte Grimoire — semantic library + MCP server (spring 2026 → present)

• Built and curates a vector-indexed personal library of esoteric texts spanning 25 disciplines (discipline tags totaling ~23,700 across its tomes), embedded with fastembed (bge-small-en-v1.5 ONNX) into LanceDB, served by a Python MCP server exposing 7 retrieval tools (search, full text, catalog, index, unsearchable-content reporting, milestone chronicling).
• Deployed it across three surfaces: claude.ai desktop, Claude Code, and his own Electron app as a permission-gated Connector — his most-used claude.ai tools are the grimoire’s own search and full-text calls.
• Authored the retrieval-layer instruction briefs by hand and gave the library a working persona (“WhAIte Witch,” invoked by a summon-witch command he designed); the server’s docstrings are written in-voice.
• His first recorded Claude Code session (2026-05-18) was a /goal readability cleanup of the library’s indexer — code stewardship before feature ambition.

Automatic Job Finder Machine — local-first job-search tool + Mac port (May 2026)

• Directs an end-to-end job-search assistant (Python pipeline + Node docx helper + Flask dashboard) that finds, ranks, and drafts but never auto-submits — a trust boundary he set. In genuine use by a real client for live job submissions.
• Commissioned the Mac port as “Path A — port, not rebuild”: zero logic changes, a launcher layer only; verified 63/63 tests plus end-to-end .docx builds; deliverable shipped as jobfinder-mac.zip.
• Expanded the résumé generator with finance (9 tags) and crypto (7 tags) domains, with signals chosen to avoid substring leakage and a test added to enforce the three-structure sync invariant; suite grew to 67/67.
• Validated generator adaptability across two deliberately dissimilar profiles (a cannabis-industry profile and an AI-systems-builder profile).

Claude TGD — AI-powered triage desktop app (June 2026 → present)

• Conceived and directs an Electron + React + TypeScript local-first desktop app whose product thesis he articulated precisely: the chat UI is not the product; the LLM triage engine is — it reads each chat and extracts item-level entries (next actions, purchases, calls, events) into smart lists.
• Personally ran the Phase 0 model evaluation: four spike iterations comparing Opus, Sonnet, and Haiku on real captures including handwritten-note vision. Eliminated Haiku over a single truncated phone-number digit (“trust-breaking”), then pushed until the truncation root-caused to a prompt-driven reformat — exonerating the model and canonizing a design law: the model copies values verbatim; deterministic code formats them. Final call: Sonnet for triage on cost-parity quality.
• Set the product’s trust doctrine: two-stage inbox (model work-queue vs. user escalations), escalate-on-ambiguity (“a triager that admits ‘not sure, your call’ beats one that confidently mis-sorts”), extracted items are suggestions requiring confirm/dismiss, every item movable and renamable so mis-sorts are correctable rather than fatal, and teaching implemented as explicit user-editable rules — never opaque learning.
• Chose the architecture from options he weighed: local-first single-device SQLite, an LLMProvider seam (“cloud now, local-ready later” — later fulfilled by a full Ollama provider sharing the same prompts and schemas), a Claude Agent SDK re-base for subscription-first billing with API fallback, and an MCP host with per-tool allow/ask/deny permissions modeled on the Claude desktop Connectors panel.
• Runs it daily on his real life: 40+ chats and 19 projects at the July migration; features include a 43-folders tickler (“Prodder Fodder,” his name and concept), RFC-5545 .ics calendar export, watched-folder phone capture, and a slash-command persona system he actively extends.
• Commissioned a full-repo bug hunt (2026-07-23): 12 finder agents → 146 raw findings → 78 confirmed (17 high) after dual adversarial verification per finding — then withheld all fixes pending his own prioritization.

STATELESS — five-faction RTS, design bible to packaged executable (July 11–12, 2026)

• Authored the founding design document — five asymmetric post-collapse factions with original mechanics (warrant-gated policing, an irreplaceable-ten spellcaster roster, recursive printer economies, body-swap “Manic Monday” events, a map-eating Leviathan) — beginning as a 55,000-character stream-of-consciousness draft he then structured. Asymmetric-faction balance is one of the harder problems in game design; he took it on five factions at a time.
• Directed three complete JavaScript versions in roughly a day (v1.0 “Five Fires” → v2.0 “The Long Collapse” → v3.0 “Heroes of the Wasteland”), then decided a from-scratch TypeScript deterministic engine (fixed-point math, 20Hz fixed timestep, seeded split-stream RNG, lockstep-ready networking) built across five phases in ~two days under mandatory human checkpoints he governed.
• Balanced by tournament: a worker-thread AI-vs-AI fleet ran six tuning rounds, pulling an outlier faction from 92% win rate to 64% and landing a 64/56/50/30/17 spread with comeback mechanics verified; a tile-occupancy index killed the performance pathology (52s → 19s; ~4ms/tick at 600 units against an 8ms budget).
• Shipped a packaged Windows executable with hotseat and LAN lockstep play (1,200-tick hash-identical peer verification), a map editor, verified save/load round-trips, and an asset-override convention he mandated so hand-made art drops in with zero code changes. Online multiplayer consciously deferred by his call.
• Served as lead playtester: his session reports were treated as bug tickets and fixed same-day (“couldn’t attack the precinct” → a melee-vs-structure geometry fix).

AFTER THE FALL — clean-room control build (July 12, 2026)

• Commissioned a second, independent RTS from the same design document with a directive forbidding the agent any access to prior project memory — a methodology experiment testing whether his written spec alone sufficed. It did: a playable five-faction v1 the same day, v1.1 (save/load, procedural combat-reactive soundtrack, random events) hours later.
• The twin-build pair doubles as a spec-sufficiency audit and produced a durable corpus of RTS engine lessons.

Butterfly-collector — governed multi-agent bug-eradication program (July 28 – August 3, 2026)

• Before first use, commissioned an adversarial review of his own campaign skill: a 17-agent workflow (six dimensions → triage → adversarial verification) confirmed 9 defects (5 high, including a hard-gate consent bypass); all were fixed and reinstalled the same day.
• Ordered a full de-identification pass before sharing the skill: fictionalized vocabulary and identifiers, broke every timezone-derivation pair, and drove residuals to zero via an adversarial de-anonymization verifier plus a 23-pattern final sweep.
• Ran Campaign 1 against his own live app: nine iterations under hard human-only gates. Personally executed the human-reserved verification stages and an independent findings pass — 4 findings including both high-severity issues the agent fleet also found, plus one self-retraction.
• Commissioned a mid-campaign efficacy audit (Fable 5, max effort) that produced doctrine revisions C1–C10, installed mid-flight — the governance process improved itself while running, and the revised rules governed the campaign’s own close.
• Closed on evidence, not fatigue: elected stop at a genuine termination-fail gate. Final measured economics: 12,449 API calls, 10.6M output tokens, ~2.57B cache-read tokens, ~$1,610 API-list-equivalent, with channel-level cost-per-finding analysis (post-merge Fable verification ~$2/finding vs. a $22 campaign average) folded back into doctrine.
• Surfaced a platform-level discovery along the way — .skill zip archives never load; only extracted directories do — and commissioned a version-skew checker so no install channel can silently diverge again.

The model doctrine + /orchestrator skill (July 24 → August 2026)

• Canonized his own orchestration doctrine — now the Oracle Blast: model choice by strength of the objective oracle × blast radius, adopted after he commissioned an adversarial review of his earlier tier mnemonic — alongside one-owner-per-file lanes, repro→fix→prove, recall-first verifier prompts (“report every issue, including uncertain ones”), and doctrine written in cognitive tiers so it survives model churn.
• Measured the doctrine against campaign transcripts and amended it where reality disagreed: the Haiku tier demoted (planned in six of nine plans, executed ~$1 of work), and Fable repositioned as a bounded verifier after he uncovered silent model-fallback behavior on safety-flagged content — the harness had been quietly substituting models without disclosure.
• Commissioned a maximum-effort audit of the skill itself: ~20 confirmed defects; a v2 rewrite A/B-tested against the original across four task shapes before any install decision.

Engine forensics under direction — C&C Renegade (July 31 – August 2, 2026)

• Root-caused a 24-year-old desktop gamma corruption by import-table scan: the game imports SetDeviceGammaRamp but never GetDeviceGammaRamp — it structurally cannot restore what it never read. Shipped a double-click fix tool (RenegadeGammaFix) with an auto-repairing launcher.
• Directed a cel-shading mod plan grounded in empirically established engine truths: loose-file override precedence proven by a targeted magenta-texture experiment; a W3D chunk-flag trap (9,038 shipped chunks contradict the format’s own header documentation) that silently breaks naive parsers; MIX1 archive IDs hashed over full stored names (a basename collision case affecting 149 textures); and the decisive ground truth that his client is Direct3D 8 with no shader path — overturning misleading public history and redirecting the project to era-correct techniques.

Claude-history data pipeline (August 3, 2026)

• Directed a converter for his complete claude.ai export (conversations.json → one markdown file per conversation with processed-state tracking) plus usage analytics over it — deliberate self-quantification of his own AI practice back to its recorded beginning in March 2026.

Claude Code skills authored (July–August 2026)

• /orchestrator — generates execution-ready multi-agent plans embodying his model doctrine; born from a bug-fix plan he liked enough to canonize.
• /butterfly-collector — the campaign-governance skill above: commissioned, adversarially reviewed, revised mid-campaign, and privacy-scrubbed for sharing.
• /handoff — compacts a session into a handoff document a fresh agent can continue from — a direct, independently built instance of what the industry now calls context/continuity engineering.
• /word-vomit — converts raw, excited, disorganized thought into structured, verifiable intent — the deliberate industrialization of his own creative method.
• Plus the TGD app’s in-app command/persona system (summon-witch, an orchestrator twin) with cross-channel version verification.
• Industry context (reported via the Cowork research brief, not independently verified here): the SKILL.md format Anthropic opened as a standard in late 2025 is now read by 30+ tools across vendors, so a well-authored skill library is portable, standards-based work rather than a single-vendor artifact.