Ask HN: How do you manage skills files?
AI users are split on whether “skills” — reusable prompt and workflow files for LLM agents — are essential infrastructure or unnecessary complexity as models improve. Many find them valuable as cached workflows and project‑specific runbooks that encode tooling conventions, reduce token use, and make agents more reliable, especially in proprietary or complex environments. Others argue that generic, internet‑downloaded skills are often snake oil, and recommend keeping only a small, curated, version‑controlled set (often managed via Git, symlinks, or custom CLIs) and relying on deterministic scripts or good AGENTS.md/docs for everything else.
What “skills” are and where they help
- Many describe skills as small, reusable prompt files or runbooks that tell agents how to perform repeatable tasks, often combining deterministic scripts with higher‑level guidance.
- Common uses:
- Macros / shortcuts for frequent prompts (e.g., rebasing, test runs, Jira workflows).
- Project‑ or org‑specific conventions: coding style, commit messages, branching, deployment steps.
- Interfaces to niche tools/CLIs or proprietary systems not well covered in model training data.
- “Workflow caches” that skip rediscovering how to do multi‑step processes.
Skepticism and “skills are obsolete” arguments
- Several argue modern models can infer most “general” behaviors (design critique, basic code review) from repos and docs, making generic marketplace skills unnecessary or harmful.
- Some see large skill collections as a code smell, extra tech debt, and a product of influencer marketing.
- Others prefer putting almost everything in AGENTS.md/README and letting the model read code and docs directly.
- There’s concern about over‑trusting LLM judgment and building fragile, opaque systems around skills.
When skills are seen as essential
- Many report big wins in:
- Proprietary, complex, or cross‑system workflows (deployments, CI, security checks, compliance).
- Reducing token use and trial‑and‑error when using custom tools (e.g., obscure CLIs, DSLs).
- Capturing hard‑won debugging flows or “how to do X here” so agents don’t relearn them.
- Skills are framed as contextual guidance and process encoding, not “extra intelligence.”
Organization, sharing, and tooling
- Common patterns:
- Keep skills under version control (often in dotfiles or dedicated repos) and symlink into agent skill directories.
- Use package‑manager‑like CLIs or “marketplaces” (internal or public) to install, update, and pin versions.
- Separate global vs project‑specific skills; some use profiles or “skillsets” per task domain.
- Sync skills across machines and agents via scripts, Nix/Home Manager, or homegrown registries.
Design, maintenance, and evaluation practices
- Principles often suggested: start with zero skills; only add when you see repeated friction or wasted tokens; keep them few, small, and task‑specific; periodically prune.
- Some treat skills like code: run evals or behavioral tests, use integration‑style checks, and update skills when agents struggle.
- Security and authority of third‑party skills are flagged as concerns; many prefer homegrown or internally reviewed skills only.