Alterlab Skill Finder

AlterLab Skill Finder — Name the Task, Get the Right Skill (or the Whole Workflow)

Skill type: ROUTER / LAUNCHER. Users say "use AlterLab skills" or fire alterflow without knowing the 230+ skill names. This is the front door: it reads the task, picks the AlterLab skill(s) that fit, and either uses one skill or launches a clarified, multi-agent workflow.

Core Mission

Copy code
WHEN THE USER NAMES ALTERLAB BUT NOT A SKILL, YOU PICK THE SKILL(S).
SIMPLE TASK → ROUTE TO ONE SKILL.   BIG / KEYWORD TASK → CLARIFY, THEN ORCHESTRATE MANY.
ALWAYS ASK YOUR QUESTIONS BEFORE YOU START A MULTI-STEP RUN.

Two modes

When to Use This Skill

• "Use AlterLab skills to [do X]." / "Which AlterLab skill should I use for [task]?"
• "Is there an AlterLab skill / workflow for [task]?"
• "alterflow — investigate [topic] and draft a paper." (→ Orchestrate mode)
• Any request that references AlterLab by name without specifying a skill.

Does NOT Trigger

Route mode — the method

1. Classify the task — what is the user doing (find literature? clean data? fit a model? write a section? query a database? design a study?).
2. Map to a domain (table below).
3. Select the skill(s) — name the best fit. For an exact niche name, consult references/skill_index.md (every AlterLab skill, grouped by domain, one-liner each) or match the installed AlterLab skill descriptions.
4. Check it is installed. The index lists all AlterLab skills, but the host can only run the ones installed there. In Claude Code a skill is available when it appears in the host's skill list as alterlab-<domain>:<skill>; a Skill call that returns Unknown skill means its plugin is not installed. Then stop and use the not-installed reply below — do not retry, and do not search for other tools.
5. Apply it — invoke the skill and do the work. Tell the user which skill you picked and why (one line).
6. If the routing genuinely forks, ask ONE clarifying question before committing.

Not-installed reply

Name the skill and why it fits, then how to get it. Its plugin is alterlab- plus the domain id shown in backticks in the index heading (for example turkish-academia → alterlab-turkish-academia):

• Claude Code: /plugin install alterlab-<domain>@alterlab-academic-skills, then /reload-plugins (or alterlab-complete@alterlab-academic-skills for every domain).
• Claude app (claude.ai): upload skill-<skill-name>-<version>.zip from the latest GitHub release (Customize ▸ Skills → Upload a skill).

If an installed skill covers part of the task, offer it as a stopgap and say what it will not cover; otherwise wait for the user to install the plugin.

Orchestrate mode — alterflow (clarify FIRST, then multi-agent)

Fires when the task spans several dependent stages (e.g. "investigate X and write it up", "research → analyze → publish") — the alterflow keyword is just an explicit shortcut into this mode, not the only way in; detect the multi-stage intent even when the user never types it. This is Anthropic's orchestrator-workers pattern (a lead agent dynamically decomposes the task, delegates to workers, and synthesizes) with a human-in-the-loop clarify/confirm gate in front. Do not start executing — run this sequence:

1. CLARIFY (mandatory, before any execution). Ask 2–4 sharp scoping questions — e.g. the research question / deliverable, quantitative vs qualitative, the dataset or corpus, the target output (report, paper, model, figures), depth/time budget, and language. Do not guess when the answer changes the plan.
2. SELECT. From the answers, choose the AlterLab skills the job needs (use the domain map + the index). List them. First ask whether it should be multi-agent at all: if the work is tightly coupled, needs one shared context, or has little that runs in parallel (Anthropic's caution — most coding-like, dependency-heavy tasks), prefer a sequential pipeline (prompt-chaining) or a single Route-mode skill over a fan-out.
3. PLAN. Lay out the dynamic multi-agent workflow — the phases, which skill runs in each, where subagents fan out in parallel, and the evaluator-optimizer verification/critique step. Scale the subagent count to complexity and default low (Anthropic's rule of thumb: simple fact-finding ≈ 1 agent; a focused comparison ≈ 2–4; only genuinely broad work ≈ 10+). Reuse the existing orchestrators rather than reinventing them:
   • a packaged job — whole-manuscript citation audit, independent review panel, claim stress-test, PRISMA dual screening, response to reviewers, grant mock panel, literature map → alterlab-research-workflows (in Claude Code: /alterlab-workflows:<name>)
   • end-to-end research→publish → alterlab-research-pipeline (deep-research → paper-writer → paper-reviewer, revision loops)
   • a whole social-science study → alterlab-ssci-orchestrator (design → measurement/reflexivity → sampling → analysis module → inference)
   • a bespoke fan-out / judge-panel / adversarial-verify workflow → alterlab-workflow-orchestration
4. CONFIRM — plan and its rough cost. Show the selected skills + the phases in a few lines, plus an effort estimate (≈ how many phases / subagents). Multi-agent runs spend far more tokens than a single pass, so let the user opt in knowingly; get a go-ahead (or incorporate a correction).
5. EXECUTE. The user's go-ahead in step 4 is their opt-in to a multi-agent run. In Claude Code with dynamic workflows, launch a matching packaged workflow (/alterlab-workflows:<name>), or write a workflow script for a bespoke plan that fans out beyond a handful of workers; for a few workers, spawn subagents (the Agent tool). On surfaces without these tools, execute the phases sequentially and keep the same hand-offs. Give every spawned subagent a complete task spec — an objective, an output format, which skills/tools/sources to use, and clear boundaries — or workers duplicate work and leave gaps (alterlab-workflow-orchestration models these specs). Carry each stage's artifact to the next (the pipelines define the hand-off contracts).

The one rule: questions before execution. A short clarify step beats a wrong 20-agent run.

Domain routing map (18 domains)

Full per-domain listing of every skill with a one-liner: references/skill_index.md.

Output pattern

Copy code
Task: <one-line restatement>
Mode: route | orchestrate
AlterLab skill(s): <name(s)> — <why this fits>
[orchestrate] Questions first: <2–4 scoping questions>   ← ask, then wait
[orchestrate] Plan: <phase 1 skill → phase 2 skills (fan-out) → verify → deliver>
→ on confirmation, execute.

References

• references/skill_index.md — every AlterLab skill, grouped by domain, one-liner each (generated, kept in sync with the catalog).

Part of the AlterLab Academic Skills suite.