Skip to main content
Glama

promptpack

A pack of prompt templates tends to drift quietly: someone renames a variable in one template and forgets its declaration, a template declares variables it no longer uses, a version string gets typed as 1.2 instead of 1.2.0, and nobody finds out until a render fails in production. promptpack checks the whole pack at once, before that happens. You supply each template's name, version, body, and the variables it declares, plus any test renders you want exercised. promptpack compares the placeholders actually used in each body against its declarations, checks every version string, runs the renders deterministically with the values you gave, and hands back one report. Nothing is published, written, or fetched; the pack you describe is the only input.

Placeholders are written {{name}}. A span counts as a placeholder only when its inner text, once trimmed, is a plain name made of letters, digits, and underscores — so {{ name }} is a placeholder for name, while {{}}, {{first name}}, {{user-name}}, and an unterminated {{ are ordinary text: left untouched in renders and never mentioned in the report. Each template's placeholders are the distinct names it uses, in the order they first appear in the body.

For every template the report echoes its declared variables and lists two findings: placeholders with no declaration (missing_declarations, in placeholder order) and declarations no placeholder uses (unused_declarations, in declaration order). A template with neither finding is ok; any finding makes it issues. Version trouble is stricter than a finding: a version that is not exactly three dot-separated non-negative integers, a duplicated template name, or a duplicated declared variable rejects the whole input with an error naming the offending path — there are no partial results for malformed input.

Test renders run in the order you supply them. Each one names a template and gives values for its variables. Every placeholder in that template is replaced by its supplied value, and the replacement text itself is never rescanned, so a value that happens to contain {{other}} stays exactly as given. Extra values are ignored. If a placeholder has no supplied value, that render does not blow up the check: its entry reports ok: false with the distinct missing names in placeholder order, and the remaining templates and renders still report. Naming a template that is not in the pack, or giving a non-string value, is malformed input and errors like any other validation failure.

CLI

npx @reneturcios/promptpack check --input args.json
npx @reneturcios/promptpack --version

args.json contains the full arguments object:

{
  "templates": [
    {
      "name": "welcome",
      "version": "1.2.0",
      "body": "Hi {{name}}, welcome to {{ product }}!",
      "variables": ["name", "product"]
    },
    {
      "name": "reminder",
      "version": "0.3.1",
      "body": "Hi {{name}}, your {{plan}} renews soon.",
      "variables": ["name"]
    }
  ],
  "renders": [
    {
      "template": "welcome",
      "values": { "name": "Ada", "product": "Promptpacks" }
    },
    {
      "template": "reminder",
      "values": { "name": "Ada" }
    }
  ]
}

The report is printed as JSON:

{
  "templates": [
    {
      "name": "welcome",
      "version": "1.2.0",
      "placeholders": ["name", "product"],
      "declared": ["name", "product"],
      "missing_declarations": [],
      "unused_declarations": [],
      "status": "ok"
    },
    {
      "name": "reminder",
      "version": "0.3.1",
      "placeholders": ["name", "plan"],
      "declared": ["name"],
      "missing_declarations": ["plan"],
      "unused_declarations": [],
      "status": "issues"
    }
  ],
  "renders": [
    {
      "template": "welcome",
      "ok": true,
      "rendered": "Hi Ada, welcome to Promptpacks!"
    },
    {
      "template": "reminder",
      "ok": false,
      "missing_values": ["plan"]
    }
  ],
  "counts": {
    "templates": 2,
    "with_issues": 1,
    "renders": 2,
    "renders_ok": 1,
    "renders_failed": 1
  }
}

Exit codes: 0 on success, 1 on a runtime error (including an unreadable or invalid input file, or invalid input), 2 on a usage error.

Related MCP server: deepbom-review

MCP server

npx -y @reneturcios/promptpack mcp

MCP client config:

{
  "mcpServers": {
    "promptpack": {
      "command": "npx",
      "args": ["-y", "@reneturcios/promptpack", "mcp"]
    }
  }
}

Tools

  • check_prompt_pack — Check a caller-supplied pack of prompt templates: placeholders used vs variables declared, version format, and deterministic test renders with supplied values.

    • templates (array, required, at least 1 item) — prompt templates to check, in pack order; each item:

      • name (string, required, non-empty) — template name, unique within the pack

      • version (string, required) — template version, exactly three dot-separated non-negative integers (e.g. "1.2.0")

      • body (string, required, may be empty) — template text; placeholders are written {{name}}

      • variables (array of strings, optional) — variable names the template declares; absent means []

    • renders (array, optional) — test renders to run against the pack; absent means []; each item:

      • template (string, required, non-empty) — name of the template to render

      • values (object, required) — values for the render: a map from variable name to replacement string

    Returns {templates, renders, counts}, where templates holds one {name, version, placeholders, declared, missing_declarations, unused_declarations, status} per input template in pack order, renders holds one {template, ok, rendered} or {template, ok, missing_values} per input render in input order, and counts is {templates, with_issues, renders, renders_ok, renders_failed}.

Docker

docker run -i ghcr.io/cloudcorprecords/promptpack
docker run --rm -v "$PWD:/w" ghcr.io/cloudcorprecords/promptpack check --input /w/args.json

Development

npm install
npm run build   # tsc -> dist/
npm test        # tsx --test test/*.test.ts (pure local, no network)
npm run dev     # run the CLI from source

License

MIT — Copyright (c) 2026 Rene Turcios

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    C
    maintenance
    Enables local candidate review of ONNX or TFLite models by checking static evidence and externally measured results against a fixed policy, producing accept/reject/hold evidence ZIPs for AI agent workflows.
    Apache 2.0
  • A
    license
    Not graded
    quality
    C
    maintenance
    Enables checking Minecraft Bedrock content before it ships by validating JSON, commands, scripts, Python and whole packs against official schemas, a real tsc, and version-pinned game data, and by looking up whether identifiers, versions and fields actually exist. It reports what will silently fail to load — including errors schemas structurally cannot catch — without generating or writing anything itself.
    1 npm
    Apache 2.0
  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables AI coding agents and MCP clients to run a read-only environment check against a project's declared requirements and receive structured results identifying exactly what needs fixing.
    260 PyPI
    3
    MIT