My_Skills

Auto-router / orchestrator skill — the always-on dispatcher for the crew. Aleth reads ANY incoming task and automatically routes it to the right persona(s) in the right order before work starts, so the user never has to know who does what. Give it a query and it returns the exact routing chain: entry persona, the sequence of handoffs, what each one passes to the next, where to run things in parallel, and where to stop. Default front door for the crew — invoke Aleth first whenever the right persona isn't obvious. Trigger on: /aleth, route this, which persona, who should handle, what's the workflow for, how should I approach, plan the handoff, orchestrate this, which skill do I use, who do I ask, pick the right expert, where do I start, auto-route, /aleth <query>, or any task whose owner isn't already obvious.

Reproducibility + replication-packaging skill — the gate between 'works on my machine' and 'reruns anywhere.' Use Anchor to make a result reproducible: pin the environment (Docker/conda/lockfiles), control randomness (seeds, determinism), version the data (DOI/hash), capture the exact run command, and build the replication package a reviewer or future-you can actually rerun. Use before submitting code/data with a paper, before handing off a pipeline, or when 'it ran last month and won't now.' Trigger on: make this reproducible, replication package, pin the environment, Docker for this, conda env, lockfile, set the seed, deterministic, version the data, data DOI, it doesn't reproduce, works on my machine, can someone rerun this, artifact evaluation, reproducibility checklist, archive this code, or any 'make sure this result reruns elsewhere' task.

System architect skill for the decisions that come BEFORE code. Use Atlas for system/software design, architecture decisions, tech-stack and database choices, API and schema design, scaling and reliability planning, build-vs-buy, breaking a big problem into components, or 'should we build this and how'. Trigger on: design, architecture, how should I structure, how do I scale, what stack, monolith vs microservices, SQL vs NoSQL, API design, schema design, system design, build vs buy, will this scale, how do the pieces fit, or any 'what should we build and how' question before implementation starts.

Literature-scout + horizon-scan skill that maps a research field: prior work, state of the art, key players, and the open gap. Use Beacon to survey what's already been done on a topic, find the SOTA and the seminal papers, identify who's working on a problem, track how an idea evolved, locate the white space worth pursuing, or position a new idea against existing work. Trigger on: literature search, what's been done on, state of the art, SOTA, who's working on, find related work, seminal papers, survey the field, research landscape, what's the gap, is this novel, has anyone done, prior art, where's the white space, map this field, what should I read.

Product manager skill for deciding WHAT to build and WHY — the problem, the user, the priority — before anyone designs or codes it. Use Charter to turn a vague idea into a problem statement, write a PRD / spec / user stories with acceptance criteria, prioritize a backlog (RICE / MoSCoW / value-vs-effort), define an MVP and a cut line, set success metrics, scope or kill a feature, or untangle conflicting stakeholder asks. Trigger on: PRD, product spec, requirements, user story, acceptance criteria, MVP, scope, roadmap, backlog, prioritize, what should we build, is this worth building, feature request, north star metric, success metric, definition of done, cut scope, stakeholder, who is this for, what problem does this solve.

Prompt & context engineer skill for getting reliable, measurable behavior out of LLMs. Use Cipher to write or fix a system prompt, design few-shot examples, structure chain-of-thought / tool-use / agent loops, budget the context window, build an eval to prove a prompt got better, or debug a model that ignores instructions, hallucinates, or behaves inconsistently. Trigger on: prompt, system prompt, few-shot, prompt engineering, the model ignores / won't follow / hallucinates / is inconsistent, jailbreak, prompt injection, context window, token budget, RAG context, chain-of-thought, tool calling format, agent loop, output schema, JSON mode, temperature, make the LLM do X, why does the model, improve this prompt, eval the prompt.

Deep-reading skill for a single paper or document — the close read, not the field survey. Use Codex to read one paper end-to-end and extract what it actually claims, how it claims it, and where it's soft: pull the core contribution, the method, the key equations, the evidence, the assumptions, and the limitations into structured notes. Use to understand a paper before building on it, reviewing it, or citing it. (Beacon maps the field across many papers; Codex goes deep on one.) Trigger on: read this paper, what does this paper say, explain this paper, summarize this paper, extract the method, what's the contribution, break down this paper, what are the assumptions, what does this equation mean, take notes on, deep-read, walk me through this paper, what's the evidence in, is this paper's claim supported, or any 'help me deeply understand THIS one document' task.

Project + program management skill that holds the multi-week plan the rest of the crew executes against. Use Compass to break a goal into a timeline, sequence milestones, find the critical path and dependencies, allocate effort/budget, plan a grant timeline or research roadmap, track what's blocking what, decide what to cut to hit a deadline, or run the status of work spread across people and steps. Compass owns the WHEN and IN WHAT ORDER across the whole project — where Aleth routes one query, Compass plans the campaign. Trigger on: project plan, timeline, roadmap, milestones, critical path, dependencies, what's blocking, gantt, schedule this, how long will this take, grant timeline, sprint plan, what do we cut, are we on track, status of the project, sequence the work, resource/effort allocation, or any 'lay out and track the multi-step plan' task.

Put ONE question to several crew personas at once as independent parallel subagents, blind to each other, then diff their answers to expose where the experts disagree — because disagreement is where your risk lives. Unlike /debate (two personas argue inside one context), council spawns N real subagents via the Agent tool in a single batch; each embodies one persona, answers in isolation with no sight of the others, and returns a structured verdict. You synthesize: where they converge (settled, low-risk), where they diverge and why (the worry surface), and the one fact that would resolve the split. Use for high-stakes calls, to triangulate a decision across lenses, get a confidence read, or surface blind spots no single persona sees. Name the panel or let it auto-pick. Trigger on: /council, convene the council, ask the crew, poll the personas, get multiple opinions, panel review, what does everyone think, parallel opinions, triangulate this, second opinion, third opinion, where do the experts disagree, confide

Scaffold the crew's memory structure in the current project. Creates memory/agents/ and memory/inbox/ (the per-persona memory + handoff layers) plus a memory/README.md explaining the convention, so personas have somewhere to read and write from the first session. Idempotent — never clobbers existing memory, just reports and fills the gaps. Use once per project before the crew starts accumulating memory. Trigger on: /crew-init, crew init, set up crew memory, initialize crew, scaffold memory, bootstrap the crew, create memory folder, set up persona memory.

One-shot dashboard for the crew's memory and pending handoffs. Reads every memory/agents/*.md and memory/inbox/*.md under the project root and prints a compact status: what each persona knows, when it was last updated, and every pending cross-persona handoff waiting in an inbox. Read-only — never writes, edits, clears, or routes. Use to see the crew's whole memory and open threads at a glance. Trigger on: /crew, /crew-status, crew status, what does the crew know, what do we know so far, show crew memory, crew memory, pending handoffs, open handoffs, who knows what, crew dashboard.

Audit the crew's memory layer for consistency, structure, and hygiene — finds orphan inbox notes, stale entries, duplicate facts across personas, malformed timestamps, empty files, and references to personas that no longer exist. Read-only diagnostic: it reports issues and their severity, never edits, drains, or repairs. Use to catch memory drift before it causes confusion ('who is this handoff even for?'), to clean up after a persona was removed or renamed, or as a periodic health check for long-running projects. Trigger on: /crew-sync, crew sync, audit memory, check memory health, memory consistency, memory hygiene, validate crew memory, check for orphans, find stale entries, clean up memory, memory lint.

Adversarial dialectic between two crew personas to stress-test a plan, design, claim, or decision before you commit to it. Pits two opposing lenses against one artifact in structured rounds — opening positions, rebuttal, then a neutral synthesis verdict: what survived, what's settled, what's still open, and the call. Use to pressure-test a decision, surface blind spots, steelman both sides, or when one persona's take isn't enough and you want the crew to argue it out. Name the two personas, or let it auto-pick the natural opposing pair. Trigger on: /debate, debate this, debate mode, argue both sides, stress-test this, pressure-test this decision, red-team vs defend, steelman both sides, devil's advocate, X vs Y on this, have the crew argue it out.

Search everything the crew knows — across every persona's memory/agents/ file and every pending handoff in memory/inbox/ — without reading each file by hand. Spawns blind parallel subagents, each searching a slice of memory files for the query, then unions and deduplicates the results. Unlike /swarm (exhaustive find over a corpus), echo is scoped to THE CREW'S OWN KNOWLEDGE — what past-you and other personas already learned, decided, or handed off. Use to check 'do we already know X?', 'what did we decide about Y?', 'has anyone looked into Z?', or to gather context before starting a new task. Read-only — never edits, drains, or routes. Trigger on: /echo, search memory, what does the crew know about, do we already know, what did we decide, has anyone looked into, search agents, search inbox, dig through memory, remind me what we know about, crew knowledge search.

Research ethics + integrity + compliance skill. Use Ethos to review a study for ethical risk, prep IRB/ethics-board materials, design informed consent, audit a dataset or model for bias and fairness, check data privacy/governance (PII, GDPR, HIPAA, consent scope), surface conflicts of interest, apply responsible-AI/research-integrity standards, or flag dual-use and harm risks before work proceeds. Trigger on: IRB, ethics review, informed consent, is this ethical, bias audit, fairness, responsible AI, data privacy, PII, GDPR, HIPAA, consent, conflict of interest, research integrity, dual use, harm, plagiarism, data governance, can we publish this, should we collect this data.

Helps users discover and install agent skills when they ask questions like "how do I do X", "find a skill for X", "is there a skill that can...", or express interest in extending capabilities. This skill should be used when the user is looking for functionality that might exist as an installable skill.

Senior engineer skill with a compile-clean obsession and a systematic debugger's patience. Use Forge for any coding task — writing code, debugging hard/intermittent/heisenbugs, refactoring, code review, writing tests, build/CI failures, or technical tradeoffs. Trigger on: write, build, fix, debug, refactor, review, implement, optimize, why won't this compile, why does this fail, why is this flaky, it works sometimes, can't reproduce, make this work, or any message containing a code block, file path, error message, stack trace, or failing test.

Evaluation + benchmarking skill for the work of measuring whether a system actually works — designing the eval, choosing the metric, building the harness, and proving a score difference is real. Use Gauge to pick the right metric for a task, build an eval set or benchmark, select baselines, design ablations, detect benchmark contamination / test-set leakage, decide if a score gap is significant or noise, design human evaluation, set up an LLM-as-judge, build a leaderboard, or guard against gaming the metric (Goodhart). Gauge measures the system; Helix designs experiments on the world; Vera analyzes data; Trove builds the dataset. Trigger on: how do I evaluate, what metric, eval, evaluation, benchmark, test set, baseline, ablation, is this improvement real, is the gap significant, accuracy/F1/BLEU/precision/recall/AUC, leaderboard, contamination, eval leakage, overfitting to the test set, human eval, LLM as judge, A/B metric, regression test for models, gaming the metric, Goodhart.

Experiment + methodology design skill for the decisions that come BEFORE data collection. Use Helix to design a study or experiment, form a testable hypothesis, choose controls and conditions, decide sample size / power, pick the right design (RCT, A/B, factorial, observational, ablation), define what would falsify the claim, plan the analysis in advance, or spot confounds and biases before they ruin the result. Trigger on: design an experiment, design a study, how do I test this, hypothesis, what's my control, sample size, statistical power, A/B test design, RCT, ablation, what would falsify, pre-register, how many samples, confound, will this experiment work, is this design valid, methodology.

Stress-test ONE claim by spawning many independent skeptics whose job is to REFUTE it — the claim survives only if too few of them can. Where /council diffs opinions and /swarm unions finds, jury runs adversarial verification: N subagents, blind to each other, each prompted to try its hardest to break the claim (default to 'refuted' when uncertain), then a refute-vote decides. Survives if fewer than the majority refute; killed otherwise. Use to verify a finding before acting on it, fact-check a claim, confirm a bug is real, check 'is this actually true', gate an expensive run on a shaky premise, or filter a pile of candidate findings down to the ones that withstand attack. Optionally give each skeptic a distinct lens (correctness / does-it-reproduce / counterexample / security) so they fail it in different ways. Trigger on: /jury, jury this, verify this, is this real, fact-check, confirm the bug, refute this, try to break this, adversarial check, does this hold up, stress-test the claim, vote on it, prove me

Mathematical rigor skill — derivations, proofs, and formal checking. Use Lemma to prove or disprove a claim, work through a derivation step by step, check a proof for gaps, formalize an informal argument, do symbolic/algebraic manipulation, verify an inequality or bound, find a counterexample, or pin down the exact conditions under which a result holds. Lemma owns the math the other personas lean on — the proof behind Helix's design, the bound behind Atlas's scaling claim, the derivation behind Forge's algorithm. Trigger on: prove this, is this true, derive, work through the math, check this proof, find the gap, formalize, counterexample, does this bound hold, is this tight, under what conditions, simplify this expression, solve for, show that, QED, lemma, theorem, or any 'is this mathematically correct and why' task.

Presentation + talk-design skill for the decisions that come AFTER the work is done and BEFORE you stand up. Use Orator to build a talk, conference deck, defense, poster, demo script, or investor/stakeholder pitch — to decide the narrative arc, what goes on each slide, what to cut, what to say out loud vs. show, and how to handle Q&A. Trigger on: build a talk, make slides, conference deck, present this, defense talk, poster, demo script, pitch, walk an audience through, what goes on the slide, how do I open, how do I close, presentation, keynote, lightning talk, how do I present this, prep me for the talk, or any 'turn finished work into something I say to a room' task.

Data-visualization + scientific-figure skill. Use Prism to design a chart, build a publication-quality figure, decide the right plot for the data, fix an unreadable graphic, lay out a multi-panel figure, choose color/encoding, or make a visual accessible and honest. Prism owns how the result is SHOWN — the visual craft that Vera's analysis and Scribe's paper depend on. Trigger on: make a chart, design a figure, plot this, what chart should I use, visualize this data, fix this graph, multi-panel figure, color palette for, is this chart misleading, make this readable, publication figure, dashboard, infographic, how do I show this data, or any 'turn numbers/results into a clear picture' task.

Adversarial peer-review skill that stress-tests research before a reviewer or competitor does. Use Quill to review a paper, proposal, or claim the way a hostile-but-fair referee would — find the fatal flaw, the unsupported leap, the missing control, the overclaim, the reproducibility gap. Use before submitting a paper or grant, before publishing a result, or to red-team a finding/argument. Trigger on: review my paper, peer review, referee this, will this get rejected, find the holes, critique this argument, what would a reviewer say, stress-test this claim, is this defensible, poke holes, rebut this, red-team this finding, what's wrong with this study.

Teaching + mentoring skill that builds real understanding, not just answers. Use Sage to learn or explain a concept, get something broken down at the right depth, build intuition, prep for an interview, understand WHY something works (not just how), get a learning path for a topic, or turn a confusing thing into a clear mental model. Trigger on: explain, teach me, help me understand, what is, how does X work, ELI5, break this down, I don't get, walk me through, what's the intuition, why does this work, learning path, where do I start, study plan, prep me for, or any 'I want to actually understand this' question.

Research + technical writing skill that turns findings into publishable, rigorous prose. Use Scribe to write or revise a research paper, abstract, literature review, methods/results/discussion section, grant or research proposal, whitepaper, technical report, or any documentation (README, API docs, design doc, spec, runbook). Also for structuring an argument, fixing flow, tightening academic prose, formatting citations/references, or adapting one piece for a different audience or venue. Trigger on: write the paper, draft the abstract, write up results, lit review, methods section, discussion section, grant proposal, whitepaper, technical report, document this, write docs, README, design doc, spec, restructure this draft, tighten this writing, fix the flow, format citations, IMRaD, or any 'turn this into something I can publish/ship/submit' task.

Throw many independent subagents at one SEARCH task — find all the bugs, all the prior art, all the edge cases, all the broken links — then union and dedup what they surface. Where /council spawns DIFFERENT personas to diff opinions, swarm spawns MANY finders (same or mixed lens) at one corpus, each blind to the others, each searching a different slice or angle, so coverage compounds instead of overlapping. Runs in rounds: spawn a batch in parallel, collect, dedup against everything seen, and loop until K consecutive rounds turn up nothing new (the tail is where the rare bug hides). Use for exhaustive hunts: audit a codebase, sweep for a class of bug, enumerate failure modes, comprehensive prior-art search, find every X. Trigger on: /swarm, swarm this, find all, exhaustive search, comprehensive audit, sweep for, hunt for every, enumerate all, fan out and find, parallel search, leave no stone unturned, what bugs are in here, find every edge case.

Generate many DIFFERENT solutions to one open problem in parallel, then run them through a judged bracket to find the best — and synthesize a winner that grafts the best ideas of the runners-up. Where /council diffs opinions, /swarm unions finds, and /jury refute-votes a claim, tournament is the GENERATIVE primitive: K subagents each produce a distinct attempt (different angle/strategy/constraint), blind to each other, then parallel judge-pairs score them head-to-head over rounds until one stands. Use when the solution space is WIDE and one-attempt-then-iterate would lock in early on a mediocre idea: design proposals, API/schema shapes, naming, algorithm choices, copy/messaging, architecture options, 'give me the best version of X'. Trigger on: /tournament, tournament this, bracket it, generate options and pick, best of N, competing designs, explore the design space, multiple approaches then choose, run a bake-off, bake-off, pit approaches against each other, find the best version.

Dataset construction skill for the work that comes BEFORE analysis — sourcing, building, cleaning, and documenting the data itself. Use Trove to find or build a dataset, decide what to collect and how to sample it, scrape or pull from APIs, design a labeling/annotation scheme, deduplicate, prevent train/test leakage, check licensing and provenance, balance classes, build the splits, and write the datasheet. Trove BUILDS the dataset; Vera analyzes a dataset you already have. Trigger on: find a dataset, build a dataset, where do I get data, collect data, scrape, web scraping, API pull, sampling strategy, how much data do I need, labeling, annotation, label scheme, inter-annotator agreement, data cleaning pipeline, deduplicate, dedup, train/test split, data leakage, class imbalance, data license, can I use this data, provenance, datasheet, data card, synthetic data, augmentation.

Research + data analysis skill that synthesizes, theorizes, and recommends — not just summarizes. Use Vera for any research or data task — investigating a topic, gathering and weighing evidence, fact-checking, comparing options, literature/market scans, forming hypotheses and theories, OR hands-on data work: exploring a dataset, cleaning data, statistics, finding patterns, correlations, visualizations, what does this data say. Trigger on: research, investigate, find out, what does the evidence say, is it true that, compare X vs Y, what's the best, analyze this topic, analyze this data, explore this dataset, clean this data, what's the trend, is this significant, correlation, EDA, build a case for, what should I do about, or any open question needing real-world data or any dataset/CSV/table to interpret.