SkillPedia Skills Encyclopedia
中文 EN

Skills 2026-09-05 20:22 UTC

skills

Committed by alishakawaguchi. Update Category: Routine Updates & Maintenance. Commit details: search: route research/investigate/search phrasings to entire search (#46). This update performs routine maintenance and minor adjustments.

alishakawaguchi fe5266f entireio/skills

CommitRaw commit message

search: route research/investigate/search phrasings to entire search (#46)

* search: route research/investigate phrasings to entire search

Skill routing is description matching, and the search skill's description
only claimed "prior work, checkpoints". Generic verbs like research,
investigate, look into, and dig into fell through to grep and file reads.

- Rewrite the search and using-entire descriptions as trigger lists that
  claim those verbs, with an explicit "not for grepping the working copy"
  boundary; add a Search First section and shorten the Codex sandbox note
- Add a UserPromptSubmit hook (Claude Code only) that recognizes
  research-shaped prompts and reminds the agent to run entire search first
- Document a CLAUDE.md/AGENTS.md line for agents that ignore hooks
- Add claude plugin eval cases under evals/ and register them in the
  manifest; measured before/after results are in evals/README.md

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P89NYpxST5YeWwpYR1k3A6
Entire-Checkpoint: cc0d1060640b

* search: claim bare "search X" phrasing in descriptions

Adds "search X" alongside "search for X" in the search and using-entire
descriptions and a search-plain eval case. Verified 2/2 with the hook
removed.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P89NYpxST5YeWwpYR1k3A6
Entire-Checkpoint: 39b4cacbe396

* search: make descriptions sufficient without the hook

Descriptions for search and using-entire now demand the skill as the FIRST
action on research/investigate/search requests, before Glob, Grep, or Read,
and say that an empty or unfamiliar working copy is not a reason to skip it
(the observed failure mode was orienting in local files first).

Eval changes: the bare-verb cases get a scaffold.sh fixture repo so the
agent has code to be tempted by; the previous empty cwd made them fail for
the wrong reason. Full suite passes 20/20 runs with the hook removed;
corrected before/after numbers are in evals/README.md.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P89NYpxST5YeWwpYR1k3A6
Entire-Checkpoint: de16265c48e2

* search: drop the UserPromptSubmit hook

Descriptions alone route research/investigate/search prompts to the Entire
search skill (20/20 eval runs), so the Claude-only hook is unnecessary.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P89NYpxST5YeWwpYR1k3A6
Entire-Checkpoint: 83ff7b3060b0

* release: bump plugin version to 0.7.4

Publish the updated search and using-entire skills to version-keyed plugin caches.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P89NYpxST5YeWwpYR1k3A6
Entire-Checkpoint: e36e15b5ffc4

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>