G-Eskayo
@G-Eskayo
Public Skills
portfolio-page
by G-Eskayo
Build or rebuild a project page on gileskayo.me from a content template, end to end on the dev site — verified facts, real evidence (screenshots, outputs, charts, diagrams), copy in Gil's portfolio voice, authored as a long-form page and checked by the pipeline. Use for "add X to the portfolio", "rewrite the X page", "this page needs evidence/diagrams", "make the portfolio page for X".
triage
by G-Eskayo
Triage issues through a state machine driven by triage roles. Use when user wants to create an issue, triage issues, review incoming bugs or feature requests, prepare issues for an AFK agent, or manage issue workflow.
to-issues
by G-Eskayo
Break a plan, spec, or PRD into independently-grabbable issues on the project issue tracker using tracer-bullet vertical slices. Use when user wants to convert a plan into issues, create implementation tickets, or break down work into issues.
resume-tailor
by G-Eskayo
Maintain a master resume and generate tailored 1-page application packages from job descriptions. Domain-siloed — job-search specific, local-only (master resume never leaves the machine). Use for /tailor, /add-to-master, /merge-resume, /parse-transcript.
diagnose
by G-Eskayo
Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression.
writing-style
by G-Eskayo
Rewrites or drafts text so it reads in Gil's actual voice instead of generic AI writing — for portfolio project pages, resume text, outreach emails, and any other user-facing prose. Use when the user asks to draft/write copy for them, asks whether something "sounds like AI," or wants a pass to make existing text sound more like them.
handoff
by G-Eskayo
Produce a structured handoff document before any context switch, topic change, or when a fresh model would improve outputs. Run autonomously — do not wait to be asked. Trigger when: conversation topic shifts significantly, context window is getting long, repeated mistakes suggest context confusion, a complete unit of work just finished, or the user opens a new project. Save to ~/.claude/handoffs/ with a timestamped filename.
readme
by G-Eskayo
Write, update or review a repository README so it does its job: orient a stranger, show the project working (not just describe it), get them to a first success, link everything, and stay true. Use when asked to write or update a README, review READMEs across repos, check a README is accurate and linked, or when a project's front page needs to show more than tell.
index
by G-Eskayo
Context retrieval using the manifest index. Pull only the files relevant to the current task — no full context load. Use at the start of any task to identify which boxes to open. Consult ~/.claude/manifest.json to match task keywords to tags, then read only those files.
write-a-skill
by G-Eskayo
Create new agent skills with proper structure, progressive disclosure, and bundled resources. Use when user wants to create, write, or build a new skill.
variable-tracker
by G-Eskayo
Track Python variable definitions, uses, and scope across files
research
by G-Eskayo
Deep research with epistemic rigor — triangulates sources, tiers confidence levels, separates facts from interpretations, and steelmans competing views. Use when user asks to research a topic, investigate a claim, compare approaches, evaluate a technology or theory, or needs a thorough evidence-based analysis.
paper-dive
by G-Eskayo
Socratic reading partner for scientific papers. Ingests a paper and walks the user from plain-English understanding (L0) up to expert-level synthesis (L5) through conversation. Adaptive to the user's analogy domains. Builds a shared lexicon. Pulls related papers from Semantic Scholar / arXiv / scihub.org. Runs challenge and compare modes.
to-tasklist
by G-Eskayo
Convert a confirmed design/requirements doc into a persisted markdown task list with per-task status checkboxes, for projects that have no issue tracker (private/local-only repos). Use when a design doc has just been approved and the project has no GitHub issues (e.g. local-only git repos holding sensitive data), instead of to-issues/to-prd.
to-prd
by G-Eskayo
Turn the current conversation context into a PRD and publish it to the project issue tracker. Use when user wants to create a PRD from the current context.
graphify
by G-Eskayo
"Use for any question about a codebase, its architecture, file relationships, or project content — especially when graphify-out/ exists, where the question should be treated as a graphify query first. Turns any input (code, docs, papers, images, videos) into a persistent knowledge graph with god nodes, community detection, and query/path/explain tools."
safety-monitor
by G-Eskayo
Calibrated verifier that scores an autonomous loop's output for risk (fabricated references, unsupported claims) before it ships, quarantining flagged artifacts for review instead of blocking the loop. Use when wiring a new autonomous/background loop, reviewing quarantined items, or asked about the safety/drift monitor.
audit
by G-Eskayo
Reactive background investigation comparing documented intent (docstrings, ADRs, roadmap markers) against actual system state to find gaps nobody's reported yet. Distinct from diagnose, which needs a known symptom to root-cause. Trigger when a direct verification question can't be confidently answered, a mismatch is noticed between something just read and observed state, or Gil expresses suspicion something isn't working as documented. Dispatched via the Agent tool in background mode — not standalone infrastructure.
architecture-review
by G-Eskayo
Autonomously reviews the WHOLE agent system (not just meta-config — every skill, script, hook, and cron job under ~/.agents and ~/.claude) and generates actionable optimization suggestions queued for user authorization in priority order. Run regularly — after every 3–5 sessions, when CLAUDE.md grows significantly, when a skill feels slow or redundant, or when routing table gets noisy. Never implements without approval. Focuses on: token reduction, retrieval speed, file organization, pipeline logic, reliability, robustness.
self-improve
by G-Eskayo
Meta-skill that runs autonomously after every non-trivial task — no user prompt needed. Observes patterns, failures, and recurring approaches, then creates or updates skills that pass a 3-gate quality filter (recurrence, evidence, value). Auto-wires new skills into commands and routing table directly (Write + Edit, no script).
research-colony
by G-Eskayo
Autonomous research agent that monitors arXiv, GitHub, and HN daily, cross-references findings against MARVIN's knowledge base, and synthesises a digest. Use when the user wants today's research digest, what's new in AI/research, or wants the colony run/fetched manually.
grill-with-docs
by G-Eskayo
Grilling session that challenges your plan against the existing domain model, sharpens terminology, and updates documentation (CONTEXT.md, ADRs) inline as decisions crystallise. Preferred default for any grilling request whenever a project/repo exists — a strict superset of grill-me. Use grill-me alone only when there's no codebase to attach docs to.
grill-me
by G-Eskayo
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Fallback only — prefer grill-with-docs (a strict superset) whenever a project/repo exists to attach docs to; use this alone only for doc-less/abstract planning or an explicit request to skip doc-tracking.
caveman
by G-Eskayo
Ultra-compressed communication mode. Cuts output tokens ~73% (verified: bench Run 16, N=3, judge-confirmed zero correctness loss — see ~/.agents/bench/RESULTS.md) by dropping filler, articles, and pleasantries while keeping full technical accuracy. Must be explicitly triggered per-request (opt-in only, not session-wide) — an earlier always-on configuration appeared to backfire but that test was later found to be confounded, not a real flaw in this mechanism. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman.