Skip to main content

Agents now surface the best-fit skill before they start typing

· 3 min read
Carefree Investments LLC
Anchor maintainers

The best feature in your toolchain is worthless if nobody knows the command for it. A new doctrine rule closes that gap: before acting on a request, an Anchor agent checks whether a skill or slash-command already available in the session would do the job faster or more correctly — and if one clearly fits, it says so in a single line and offers to use it, then gets on with the work.

The rule​

This lands as mythos-core rule 14, mirrored into the Claude Code standing rules and the Grok Build hard rules so it holds on every model tier, not just the frontier ones:

Before acting on a request, judge whether a skill, slash-command, or feature available in this session would do it faster or more correctly than working by hand. If one clearly fits with a nameable benefit, prepend a single line naming it and offering to use it, then proceed with the request the same turn.

In practice that looks like one unobtrusive line before the agent dives in:

Tip: this is what /work is for — want me to run it instead? Proceeding manually otherwise.

The user learns a feature exists at the exact moment it would have helped, without having to read a command reference first. That's the whole point: passive discovery of the cutting-edge, elusive features people otherwise never find.

Why it needs guardrails​

A nudge like this fails in two obvious ways — it invents commands that don't exist, or it turns into a nag on every prompt. Rule 14 is built to do neither:

  • Available-only. The agent suggests only a skill or command actually loaded in this session (the harness skill roster, .claude/commands/, .grok/skills/). It never names a command that isn't there and never invents a feature; when it isn't sure something exists, it stays silent. Hallucinating a /command is the sharpest failure mode, so it's the hardest limit.
  • Once per capability per session. A given skill is surfaced at most once — and not at all if you're already using it or have waved it off.
  • Materially better only. The tip fires only when the concrete win can be named in a few words. If it can't, the agent says nothing.

It is a suggestion, not a gate. The agent never blocks on it, never makes you "run this first and come back" — it does the work you asked for on the same turn. And it is deliberately not a pre-flight item: rule 13's fixed six-line block is untouched, so the check adds discipline without adding ceremony.

Where it lives​

  • anchor/system-prompts/mythos-core.md — rule 14 plus a new anti-pattern ("silently hand-doing what a purpose-built available skill is for without offering it once").
  • platforms/claude-code/CLAUDE.md — standing rule for Claude Code projects.
  • platforms/grok-build/GROK.md — hard rule 13 for Grok Build.
  • /doctrine — the shipped-behavior write-up.