mcpField
S
setup-sakana-skills

setup-sakana-skills

github.com/sakana-yOther239 installsUpdated today
Repository

About setup-sakana-skills

setup-sakana-skills helps configure shared working-record and issue-tracker guidance for selected consumers such as Claude Code and Codex. It inspects existing instructions, resolves only missing choices, shows proposed text for approval, and verifies the applied configuration.

Key features

  1. 01Inspects existing guidance and pointers
  2. 02Resolves settings by precedence
  3. 03Configures selected clients and scope
  4. 04Shows proposed text before changes
  5. 05Verifies approved configuration

Use cases

  1. 01Setting shared project conventions for Claude Code and Codex
  2. 02Documenting working-record locations and permissions
  3. 03Configuring a canonical issue tracker
  4. 04Checking that setup introduces no duplicate guidance

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/sakana-y/skills --skill setup-sakana-skills

After installing, type/setup-sakana-skillsin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

agents/
openai.yaml
references/
issue-tracker-github.md
issue-tracker-local.md
working-records.md

Details

From the project README

Setup Sakana Skills

Prepare conventions that installed consumers can read directly. Run only when the user invokes this skill. Reuse existing guidance; sufficient configuration works without setup. Setup owns configuration, not runtime scheduling, skill installation or record migration.

1. Inspect effective guidance

Read applicable user-level and project AGENTS.md, CLAUDE.md and their pointers, including docs/agents/ when present. Verify the selected clients' actual instruction locations rather than inventing a native configuration key. Inspect repository identity and installed consumers' requirements. Distinguish working records from canonical issues, evaluation fixtures and private evolution history.

Resolve settings from the current explicit instruction, then project overrides, then global defaults. Record the source of each effective value. Existing user-selected record locations and commit permissions remain valid within their stated scope. Detect conflicting guidance, duplicate pointers and shared or symlinked files before editing.

Complete when: every required setting has an effective value, a missing choice or a specific conflict. With no missing or requested change, report that setup is unnecessary.

2. Resolve only missing choices

Carry forward decisions already made. Configure only supported consumers and the requested scope:

  • Scope and clients: global defaults, project overrides or both; Claude Code, Codex or both. Show how a project inherits or overrides values. Project tracker identity stays project-specific even when general conventions are inherited.
  • Working records: exact root, verified project/upstream/fork identity, existing index and individual-record layout, and useful author identities or explicit provenance for locating the user's PRs. Use one common format for observations and opportunities; reuse existing history without moving it.
  • Record authority: which kinds permit ordinary saves and substantive updates; separately, whether each new record and substantive update should receive a local commit. This policy may cover observations, opportunities, assessments and implementation progress when the user selects it. Amendments and pushes remain separate instructions. Do not infer broader permission from an Explore-only rule.
  • Canonical tracker: reuse established Tracker Guidance. When absent and needed, propose GitHub Issues for a verified GitHub project or a user-selected local Markdown directory. Preserve another established tracker. Confirm its exact destination and tool; supplemental remote evidence does not create a second canonical destination.

Read the working-record guidance template when configuring records. Its shared contract comes from Assess; declare that dependency for consumers using it. Read the existing GitHub or local tracker template only when configuring that tracker.

Complete when: required choices have exact existing or user-selected values. Missing access is an explicit limit, not a successful connection.

3. Show the concrete configuration

Keep one authoritative document per guidance topic. Reuse existing locations; for new project guidance, docs/agents/working-records.md and docs/agents/issue-tracker.md are suitable. For global defaults, propose an exact user-level document and pointer reachable from the selected client's verified global instructions. Global configuration here is agent-readable guidance, not a new application setting.

Show the exact text, paths, scope and resulting effective values before writing. Agent entrypoints carry concise, conditional pointers. The record pointer must cover ordinary chats: before starting implementation or a PR, read relevant records and claims; after delivery, reconcile links and update progress within configured authority. Preserve independent-audit history gates. Do not copy skill bodies or install runtime routers into entrypoints.

Preserve unrelated domain, label and project rules. Both clients may point to the same guidance. Resolve symlinks once. Include a narrow ignore rule only when the user has selected private local storage; do not assume a records repository is private. Creation of a destination or changes to existing records are separate from documenting its location.

Complete when: the user has approved the exact configuration changes. Prior approval for that same text and scope remains valid. Approval of the reusable skill's design is not approval to edit a user's global settings.

4. Apply and verify

Apply only approved configuration. Verify pointers from each selected entrypoint, effective precedence, project identity, the distinction between record and tracker authority, and preservation of unrelated content. Confirm any authorized directory creation and report inaccessible paths separately.

Re-read the result as a second setup pass: unchanged guidance should require no further edits. Report changed files, global versus project effect, consumers and unverified access. Do not create findings, publish issues, migrate history or schedule synchronization as part of setup.

Complete when: approved guidance is readable, effective choices are unambiguous, and a second pass finds no duplicate or necessary rewrite.

Comments

Markdown supported