SkillRanker: picks the right skill for the next step
SkillRanker
The Rust sr CLI evaluates live session context, compares skills and suggests one fitting skill to the agent, or none if unsure
- 141 ★ on the repository
- Last commit
Extended access
Why this level
- The context of your conversations with the agent leaves the machine for each evaluation, and data redaction is fallible
- The hook runs on every prompt and spends the provider's paid attempts
Install
- Works in
Manual install
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/skillranker/main/install.sh?$(date +%s)" | bash -s -- --verifyInstalls sr into ~/.local/bin with SHA256 verification. From source: cargo install --locked --path . --bin sr.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The context of your conversations with the agent leaves the machine for each evaluation, and data redaction is fallible
- The hook runs on every prompt and spends the provider's paid attempts
This is third-party code. Review the repository files before installing.
What it does
SkillRanker (`sr`) is a standalone Rust utility that answers which skill from a large library the agent needs for the next step. It takes the recent conversation, the current request, workspace signals and the list of skills visible to the chosen agent harness. The evaluation comes from a Jev-compatible provider: TypeSafe.ai by default, or TypeSafe Jev on Cloudflare. It first compares candidates broadly, then reads richer excerpts from a shortlist and checks whether each fits. Both comparisons include an explicit none-fits option. For libraries over 254 skills, local Quill search narrows the candidates. An explicit skill request resolves locally before any calls. The result is advisory: the agent decides. For Claude Code, `sr install-hook claude` installs a `UserPromptSubmit` hook that names at most one validated skill and never blocks work on failure. The `sr rank`, `sr demo` and `sr capabilities --json` commands and the `--why-not` flag show why a skill did not make the recommendation.
Who it is for. People whose Claude Code or Codex has accumulated dozens of skills with similar descriptions and who lose time and context choosing the right one.
Good fit when
- The library holds hundreds of skills and the agent loads the wrong ones, wasting context
- After several turns the previous skill is stale and a new suggestion for the current step is wanted
- You want to know why the right skill was not suggested: `--why-not` shows the threshold and a hint
- Before attaching to a session you want to watch it work on an offline fixture with `sr demo`
- Spending on evaluations must be capped with a shared attempt allowance and a provider circuit breaker
Not a fit when
- You need fully offline work: a fresh evaluation requires your provider key and network
- You have no TypeSafe.ai key or Cloudflare account: recommendations will not appear without them
- The library holds three or four skills: evaluation costs outweigh the benefit of a hint
- You need guaranteed automatic skill loading: the result is only advice and the agent chooses
Example request
See which skills fit the next step in this session and explain with --why-not why the testing skill did not make the list.Limitations
Fresh ranking needs your own provider key (`TYPESAFE_API_KEY` or Cloudflare credentials) and explicit consent to the network with the `--allow-network` flag or a setting. The program has no local model. A third-party service computes the evaluations, and the program only prepares data for it.
Quality is bounded by the visible skill list and by Quill search, which can miss paraphrases and multilingual matches. Native harness support needs a verified event contract, so Claude Code and Codex work most reliably. Installation supports Linux and macOS, and on macOS the state must sit on a local APFS or HFS+ filesystem.
How to disable. Remove the hook with the command from the docs and delete the `~/.local/bin/sr` binary. The installer adds a small usage skill to detected Claude Code and Codex installs, so delete its folder by hand; with `--no-configure` there are no extra files.
FAQ
How do I install SkillRanker?
An installer with SHA256 verification puts `sr` into `~/.local/bin`: `curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/skillranker/main/install.sh?$(date +%s)" | bash -s -- --verify`. From source: `cargo install --locked --path . --bin sr`.
What is Jev?
It is an evaluation provider that compares candidates by meaning and can say that none fits. SkillRanker supplies the session integration, local safeguards and feedback around it. Without a provider key, fresh ranking is unavailable.
What leaves the machine?
Fresh evaluations send redacted context and skill excerpts to the provider. The network is switched on only by your decision, and a key alone is not enough. The `--offline` flag guarantees zero requests.
Does the hook block the agent?
No. If a recommendation fails, the hook prints nothing, writes a short diagnostic and exits with code zero. It names at most one skill and stays silent on an ordinary abstention.
Which agents are supported?
The hook is implemented for Claude Code through `sr install-hook claude`, and the installer adds a usage skill to Codex too. For other harnesses the portable boundary is normalized-context input.
Discussion
No questions or reviews yet.
Related
CLI toolsA plugin and skill that pick three vetted skills for a task from an offline registry, with a safety grade and reasons.
A doctor skill for agent skills that picks an architecture, audits against 21 red lines and returns a report with line references
A Rust tool with its own window: scans Codex, Claude Code and Gemini CLI skills, checks risks and enables or disables them
A skill that designs, tests, evaluates and packages other agent skills, starting from architecture
The clawskills CLI searches an index of 125,000 skills and installs matches into Claude Code, Cursor, Codex and OpenClaw folders.
A skill that finds other skills for your task through the DeepSkill Market service and ranks them using reviews from people who ran them
README in short
The README lays out the problem of choosing among similar skills and the two-stage ranking with local Quill search. Sections cover commands, configuration, explaining and replaying a result, local feedback, budgeted evaluation, hooks, the inline TUI, privacy and limitations. The licence is at the end of the file.