Vol. I · No. 579United States, Saturday, September 26, 2026Agent Tooling · Page B4
All workSection B · Page B4

Agent Toolingskill-issueNew

The Right Skill for Every Prompt, or None at All

A fast local router that tells your coding agent which Agent Skills to load for each prompt, and stays quiet when none fit.

Real output, real timings: a CI failure goes to gh-fix-ci, a rename gets no skill.Moving picture

Coding agents pick which skill to use by reading a one-line description of every skill in their context. That’s fine with ten skills. With hundreds it falls apart: Claude Code gives the whole skill list about 1% of the context window and starts dropping descriptions when it overflows. So the more skills you install, the worse your agent gets at using them.

skill-issue takes that decision out of the prompt. It runs as a warm daemon on your machine, and when you send a prompt it finds the skills that fit, or decides nothing fits, in a fraction of a second. Your agent only sees the zero to three skills that cleared the bar.

Scoring candidates one at a time was the key design choice.

It works in three steps. First, retrieval: BM25 over each skill’s name, description and full body, plus a small embedding model (bge-small-en-v1.5), fused with reciprocal rank fusion into a top 40. The body matters a lot, because that’s where a skill says what it actually does. Second, a gate: each of the top 12 candidates gets scored on its own with one question, would this skill help with this request. Third, a threshold: the scores are Platt-calibrated on held-out data so they behave like probabilities, and only skills above it get injected.

Scoring candidates one at a time was the key design choice. Asking one model to pick the right skill out of dozens of options gets worse as the list grows. Asking a yes or no question per skill doesn’t. The default gate is gte-reranker-modernbert-base, which I fine-tuned for exactly this question, with a fine-tuned MiniLM-L6 for machines without a GPU. Fine-tuning is what made it good: the gte gate went from 75.1% to 85.8% exact match at 100 skills.

The Claude Code side is a UserPromptSubmit hook written with only the Python standard library and a hard 1.5 second cap. It talks to the daemon over localhost with a token. If the daemon is starting, slow or broken, the hook prints nothing and your prompt goes through untouched, so the worst case is the same as not having it. The model weights download once from the repo’s release and get checked against a pinned SHA-256.

I benchmarked it properly because I wanted to know if it actually helps. With 1,000 skills installed, Claude Code on its own loaded exactly the right skills for 77% of prompts. With skill-issue it was 93%. On the same 100 prompts the router fixed 16 and broke none (exact McNemar p = 3e-5), cost per prompt didn’t change, and the median prompt took 0.6 seconds longer. I also compared it against SkillRouter, which uses two 0.6B models: my gate is ahead at 100 and 1,000 skills, about even at 10,000, and behind on the SkillRet benchmark.

It isn’t magic, and the README says so. Every system gets worse as the catalog grows, and at 18,719 skills the best ones land around 44% exact match. The gate is also cautious on purpose: it would rather load nothing than load the wrong skill. On a laptop RTX 4050 the whole hook takes 282 ms at the median on Windows, and about half of that is just starting Python.

It ships as a Claude Code plugin, an MCP server for Codex, Gemini CLI and Cursor, and a companion skill for any agent that supports Agent Skills. It’s open source under Apache 2.0.

Graphic

How it works

skill-issue, start to finish, in 5 steps.

  1. Step 1:

    Your prompt

    A stdlib-only hook with a 1.5 s cap sends it to the local daemon

  2. Step 2:

    Retrieval

    BM25 plus bge-small embeddings over every skill's full text, top 40

  3. Step 3:

    Gate

    A fine-tuned reranker scores the top 12, one yes or no each

  4. Step 4:

    Threshold

    Platt-calibrated scores, three skills at most

  5. Step 5:

    Agent

    Loads only those skills, or nothing at all

Graphic: The Sahil Bachu

More from The Work

See all 26

Voice AI · New

AI Agent Now Phones the Caterer for You

Dispatch finds local vendors for your event, calls and emails them, and brings back real quotes.

An AI event-planning agent. Tell it about your wedding, party or corporate event and it finds local vendors, reaches them by AI phone call or email, and turns what they say into quotes you can compare.

Continued on Page B2

Creator Tools

A Telegram Chat That Runs My Instagram

autoreel turns one talking-head clip into a finished, edited Instagram Reel and posts it.

The daily Reels pipeline behind @21stcentury.sahil. It finds topics, drafts scripts in my voice, edits my clip into a reel with Remotion, posts it and sends tool links to commenters. I run all of it from one Telegram chat.

Continued on Page B6

Browser Agents

The Job Hunt Agent That Refuses to Hit Submit

JobOps finds jobs, tailors a resume for each one and pre-fills the application, then stops and waits for me.

A job-search command center that runs on my laptop. It pulls postings from five sources, scores them with headless Claude, tailors a LaTeX resume per job and pre-fills applications in Chrome. It never submits. I review and send every one.

Continued on Page B10