Polish Docs Meta

When to Use

• Server implementation is functionally complete (tools, resources, prompts, services all working)
• bun run devcheck passes, tests pass
• You're preparing for first commit, first release, or making the repo public
• User says "polish", "polish docs", "finalize", "make it ship-ready", "clean up docs", or similar
• Re-running after adding/removing tools, resources, or other surface area changes

Prefer running after implementation is complete, but safe to re-run at any point — steps are idempotent.

Companion: pair with security-pass for a full pre-ship review — this skill polishes docs and metadata; security-pass audits handlers for MCP-specific security gaps.

Prerequisites

• All tools/resources/prompts implemented and registered
• bun run devcheck passes
• Tests pass (bun run test)

If these aren't met, address them first.

Steps

1. Audit the Surface Area

Read all tool, resource, and prompt definitions. Build a mental model of what the server actually does — names, descriptions, input/output shapes, auth scopes. This inventory drives every document below.

Read:

• src/index.ts (what's registered in createApp())
• All files in src/mcp-server/tools/definitions/
• All files in src/mcp-server/resources/definitions/
• All files in src/mcp-server/prompts/definitions/
• All files in src/services/ (if any)
• src/config/server-config.ts (if any)

Capture: tool count, resource count, prompt count, service count, required env vars.

2. README.md

Read the gold standard first: the pubmed-mcp-server README — https://github.com/cyanheads/pubmed-mcp-server/blob/main/README.md (or ../pubmed-mcp-server/README.md when that repo is checked out beside this one). Mirror its structure, section order, heading forms, and density. Reuse its non-server-specific content as-is — the framework line under Features, and the Getting started / Configuration / Running the server / Development guide / Contributing boilerplate — and tune only what is server-specific: the header, the Overview description, the primitives and their Capability reference entries, the domain-specific and agent-friendly bullets, env vars, project structure. Then read references/readme.md for the conventions spelled out; where it and the pubmed README disagree, the README wins — note the discrepancy in your report rather than editing the reference.

If README.md doesn't exist, create it from scratch. If it exists, diff the current content against the audit — update tool/resource/prompt tables, env var lists, and descriptions to match the actual surface area. Don't restructure sections that are already accurate and already in the gold-standard shape.

Every run is also a concision pass. READMEs accrete: each release adds a bullet, and nobody removes one. Accurate is not the same as done — on every run, tighten what is already there, especially the Capability reference. Read references/readme.md § Concision for what to keep and what to cut. The bar is natural prose a human reads once, that still carries every fact a caller needs before the first call, and where every retained claim has been checked against the definition it describes.

The bold header tagline (the <b> text inside the first <p>) must match the package.json description. The surface count is a nested <div> inside the same <p>, separated by •.

3. Agent Protocol (CLAUDE.md / AGENTS.md)

Update the project's agent protocol file to reflect the actual server. Scope is the project-root CLAUDE.md / AGENTS.md only — do not edit framework-skills/*/SKILL.md or their references/ files. Those are external skill files synced from @cyanheads/mcp-ts-core and get overwritten on the next maintenance refresh.

Read references/agent-protocol.md for the full update checklist, then review the current file and address what's stale or missing:

• If a "First Session" onboarding block is still present and onboarding is complete, it can go
• If example patterns still use generic/template names (e.g., searchItems, itemData), replace with real definitions from this server
• If server-specific skills were added, update the skills table
• Verify the structure diagram matches the actual directory layout
• If custom scripts were added to package.json, update the commands table

4. .env.example

Compare .env.example against the server config Zod schema. Add any missing server-specific vars with a comment and default (if any). Remove vars for features that no longer exist. Group by category. Preserve existing framework vars that are still relevant.

5. package.json Metadata

Check for empty or placeholder metadata fields. Read references/package-meta.md for which fields matter and why. Fill in anything still missing — skip fields that are already correct.

Key fields: name, description, repository, author, homepage, bugs, keywords.

name must communicate the server's domain at a glance. See references/package-meta.md for the naming convention — ambiguous abbreviations and acronym-only names fail the scannability test for humans and agents alike.

name and title in createApp() / createWorkerHandler() must match the unscoped package.json name — display identity is the machine name on every surface; lint:packaging (run by devcheck) enforces the match and warns when the pair is partial. description is never duplicated into the entrypoint — package.json is the canonical source (the framework derives the served description from it). Adopting the pair also seeds OTEL_SERVICE_NAME when unset, so telemetry's service.name switches to the machine name on first boot — expect a one-time series split in backends keyed on the old scoped label.

description is the canonical source. Every other surface (README header, server.json, Dockerfile OCI label, GitHub repo description) derives from it. Write it here first, then propagate.

6. server.json

Read references/server-json.md for the official MCP server manifest schema. If server.json doesn't exist, create it from the surface area audit. If it exists, diff against current state and update stale fields.

Key sync points:

• $schema set to https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json
• name matches mcpName from package.json (reverse-domain: io.github.{owner}/{repo})
• version matches package.json version (in all three places: top-level + each package entry)
• description matches package.json description
• environmentVariables reflect the server config Zod schema — server-specific required vars in both entries, transport vars only in HTTP entry
• Two package entries: one for stdio, one for HTTP (if both transports supported)
• npm entries carry runtimeHint: "npx" and no packageArguments — a registry client runs the package's bin, which ignores run start:* arguments; the HTTP entry fixes MCP_TRANSPORT_TYPE as "value": "http", never as a default. lint:packaging fails the arguments and a missing or wrong transport value

7. GitHub Repository Metadata

Sync the GitHub repo with package.json using the gh CLI. Skip if the repo isn't hosted on GitHub or gh isn't available.

Description:

gh repo edit <owner>/<repo> --description "<package.json description>"

Topics ↔ Keywords:

Compare GitHub topics (gh repo view --json repositoryTopics) against package.json keywords. They should be the union — add any that exist in one but not the other:

• Missing from GitHub → gh repo edit --add-topic <topic>
• Missing from package.json → add to keywords array

Common keywords shared across MCP servers (e.g., mcp, mcp-server, model-context-protocol, typescript) should appear in both. Domain-specific keywords should also be present in both.

8. bunfig.toml

Verify a bunfig.toml exists at the project root. If not, create one:

[install]
auto = "fallback"
frozenLockfile = false

[run]
bun = true

9. Changelog

Two patterns are supported — pick one and stay consistent.

Both are acceptable. The template scaffolds the directory-based structure by default; collapse to monolithic only if the rollup tooling is pure ceremony for this project.

Directory-based — per-version files live at changelog/<major.minor>.x/<version>.md (e.g. changelog/0.1.x/0.1.0.md), and CHANGELOG.md is a rollup regenerated by bun run changelog:build. Devcheck's Changelog Sync step enforces drift protection. changelog/template.md is a pristine format reference — never edited, never moved, never renamed. Read it to remember the frontmatter + section layout when scaffolding a new per-version file.

If the structure doesn't exist yet:

1. Make the changelog/ directory
2. Create changelog/template.md once from the template (frontmatter stub + H1 # <version> — YYYY-MM-DD placeholder + empty Added/Changed/Fixed sections) — this file is a format reference only and stays as-is after creation
3. If the server already has a shipped version (e.g. 0.1.0), create the series directory and initial entry: changelog/0.1.x/0.1.0.md with H1 # 0.1.0 — YYYY-MM-DD, concrete version and date — do not rename or move template.md to create the version file; author the per-version file directly
4. Run bun run changelog:build to generate CHANGELOG.md

Per-version file format:

---
summary: One-line headline for the rollup index — ≤350 chars, no markdown
breaking: false
---

# 0.1.0 — YYYY-MM-DD

Optional narrative intro (1-3 sentences).

## Added

- [list tools, resources, prompts, key capabilities]

Frontmatter: summary is required (powers the CHANGELOG.md index), breaking is optional and defaults to false (set true for releases requiring consumer code changes).

Never hand-edit CHANGELOG.md when using this pattern — it's a build artifact. Never edit changelog/template.md — it's the format reference. Never use [Unreleased] as a version header in a released file.

Monolithic — maintain CHANGELOG.md directly in Keep a Changelog format. To collapse from the template default: delete the changelog/ directory, remove changelog:build and changelog:check from package.json scripts (and from devcheck.config.json if referenced), and drop "changelog/" from the files array. The release skill's directory-specific steps then don't apply — just edit CHANGELOG.md and bump version at release time.

10. Plugin Metadata (Codex / Claude Code)

lint:packaging (run by devcheck) now enforces the high-value subset automatically when these manifests are present: non-empty descriptions, version equal to package.json's, identity/install correctness — display fields (name, server key, interface.displayName) must be the unscoped machine name, while the npx -y install arg must be the full package.json name (scoped if scoped) — and the env contract below (no "" values; every ${user_config.*} reference declared). Opt out per project with "packaging": { "pluginManifests": false } in devcheck.config.json. The checks below cover the fields the gate doesn't (repository / license sync, category, the wording of each option).

How user-supplied values reach the server. Neither client passes the user's shell environment through untouched, so an env entry of "KEY": "" is not a hint — it is the value the server receives, and the framework reads an empty string as unset. Claude Code prompts for values declared under userConfig at enable time and substitutes ${user_config.<option>} into env (sensitive values go to the Keychain). Codex starts stdio servers with a whitelisted environment and forwards only the host variables named in env_vars. Mirror manifest.json's user_config block: same options, same titles and descriptions.

If .codex-plugin/plugin.json exists, verify it's populated and in sync with package.json and server.json:

• name is the unscoped package.json name (display identity is the machine name on every surface)
• version matches package.json version
• description matches package.json description
• repository matches package.json repository.url
• license matches package.json license
• interface.displayName is the unscoped package.json name
• interface.shortDescription matches package.json description
• interface.category is set to a meaningful category

If .codex-plugin/mcp.json exists, verify the server-name key is the unscoped package.json name, the npx -y install arg is the full package.json name, env carries only fixed values (MCP_TRANSPORT_TYPE), and env_vars lists every user-supplied variable from the server config schema (API keys, contact emails, instance URLs) so Codex forwards it from the user's environment.

If .claude-plugin/plugin.json exists, apply the same checks: name (unscoped), version, description, author, repository, license, keywords from package.json, $schema set to https://json.schemastore.org/claude-code-plugin-manifest.json. Verify the inline mcpServers entry key is the unscoped name and its npx -y install arg is the full package.json name. Every user-supplied variable is declared under userConfig — type: "string", title, description, sensitive: true for keys and tokens, and either required: true or default: "" so a blank answer reaches the server as empty rather than as the literal placeholder — and referenced from env as "KEY": "${user_config.<option>}". Run claude plugin validate . after editing; the CLAUDE.md-at-root warning is expected, anything else is not.