anorth/agent-skills

model-tiers

Resolves which model a subagent runs on from three tiers (cheap, standard, strong), listed in the user's ~/.agents/model-tiers.md and matched to what the current harness offers.

ソースを見る

使い方

  1. インストールコマンドをコピーし、自分の環境へ導入する前にソースを確認します。
  2. 必要なツール、認証情報、対応環境を原文で確認します。
  3. 入力と成果物を指定し、出力を資料と照合します。
npx skills add https://github.com/anorth/agent-skills --skill model-tiers

開始前の確認

  • 導入にはモデル契約、API枠、ファイルへのアクセス許可は含まれません。
  • スター数や導入数は活動の指標であり、独立した品質・安全性試験ではありません。
リポジトリの原文

見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。

Model tiers

Skills name a tier, not a model. This skill turns a tier, or a model the user names, into a model the current harness accepts.

Tiers

  • cheap: high-volume, narrow work. One theme of a themed review, scanning, summarising.
  • standard: routine implementation and review.
  • strong: judgment. Design, coordination, triage, reviewing or verifying others' work, the hardest changes.

Config

The user keeps one list for all harnesses in ~/.agents/model-tiers.md. Each tier line names models in preference order, by family and version as a person would write them. An optional effort in parentheses applies where the harness supports it.

Example:

markdown
cheap: Claude Haiku 5.5, Gemini 3.5 Flash
standard: Claude Sonnet 5.5, GPT-5.6 Terra
strong: Claude Opus 5.5 (high), GPT-5.6 Sol (high)

Not every harness offers every model. Entries the harness can't run are skipped.

Spawning

Spawn a fresh general-purpose subagent, not a fork of this conversation. A fork may ignore the model and carries the implementer's context. If the harness has no subagent tool, tell the user and stop; don't do the subagent's work inline.

Match a model name

To turn a config entry or a name the user gives ("Opus", "GPT") into a model, match it against the values the subagent tool accepts: its schema, or the list in a rejection message.

  • Cursor-style tools take full slugs, which encode effort: pick the slug for that model and effort.
  • Claude Code-style tools take an alias (opus, sonnet, …) and a separate effort parameter.
  • Match family and version.

If that version isn't offered, take the nearest newer one in the same family. Never substitute a different family.

  • An entry with no match is unavailable here; skip it.

Resolve a tier

  1. Read the config once per session and take the tier's line.
  2. Drop entries that don't match an available model.
  3. For a reviewer or verifier, take the first remaining entry that passes Reviewers differ from the implementer.

For anything else, take the first.

  1. If the tool rejects a model, try the next.

If the line runs out, omit the model so the subagent inherits the parent's, and say so.

  1. If the file or line is missing: ask the user once which model to use for that tier this session,

and offer to write it to the config. If you can't ask, omit the model and say so.

A model the user names explicitly wins over the config. If it has no match, or the match is ambiguous, ask. When the user gives a count larger than the models they name, resolve the rest from the tier the calling skill uses.

Reviewers differ from the implementer

A reviewer or verifier of work runs on a different model from the one that wrote it, so it doesn't share the author's blind spots.

  • The implementer is whichever model wrote the change:

this session's model if you wrote it, or the model given to the implementing subagent. If you don't know, say so.

  • Prefer a different model family (Claude, GPT, Gemini, …).

Where the harness offers one family, a different model in it.

  • Several reviewers doing the same thing each get a distinct model, all different from the implementer.
  • If the tier's line has no qualifying model, look in higher tiers, then lower tiers.

If none qualify at all, use the implementer's model. Say which in the report.

  • Flag a user-named model that matches the implementer.

Set up or change tiers

  1. List the models the subagent tool accepts. If you can't, ask the user to paste them.
  2. Propose a line per tier, keeping entries the user already has for models this harness doesn't offer.

Give standard and strong at least two families where possible, so a reviewer always has an alternative.

  1. Confirm with the user, then write the file.
同じリポジトリから

関連する Skills

すべての Skills
anorth
コミュニティ

build-ticket

Builds one ready ticket through to a reviewed result with evidence, without stopping for routine decisions: checks it's ready, builds to its design, runs its done checks, reflects, reviews, and reports. Use when the user asks to build, implement, or pick up a ticket.

1時間の導入数
2
GitHub Stars
1
更新日
11月3日
anorth
コミュニティ

deep-review

Deep, themed review of a diff: the review-diff flow and its standard lenses, plus one cheap reviewer subagent per theme, using distinctive preferences distilled from Alex North's public GitHub reviews (naming, API shape, errors, tests, coupling, and related themes). Needs the review-diff skill. Use when the user asks for a deep review, themed review, or anorth-style review.

1時間の導入数
2
GitHub Stars
1
更新日
11月3日
anorth
コミュニティ

domain-modeling

Build and sharpen a project's domain model. Use when discussing codebase terminology, writing or editing a GLOSSARY.md, or recording or editing an ADR.

1時間の導入数
2
GitHub Stars
1
更新日
11月3日
anorth
コミュニティ

grilling

Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.

1時間の導入数
2
GitHub Stars
1
更新日
11月3日