tool-definition-review
Skill15 starsChanged 23 days agoMITView on GitHub
Use when a diff touches src/tools/**, whether a tool definition (name, title, description, parameter descriptions, annotations) or the handler behind it, including from /code-review; when asked to review or improve how a tool describes itself to agents; or when a tool scores low on Glama.
In this repository5 other files it gives its agents.
- AGENTS.mdAGENTS.md
- functional-testingSkill
- functional-validation-planSkill
- probe-spikeSkill
- run-validation-planSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 4 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 24
Deterministic report: Run pnpm -s tool-definitions:report.
- Distinctive: line 40
If one could, that is a Usage Guidelines finding: the description must say when to use this tool rather than that one.
- Distinctive: line 43
If none could, there is nothing to report.
- Distinctive: line 59
keep the first sentence standalone, because proxy modes show only it.
- Distinctive: line 62
add no padding: Conciseness & Structure scores it down.
- Very specific: line 64
Do not ask for an outputSchema: tools return text shaped for the agent, by decision (docs/decisions/tool-definition-quality.md).
- Very specific: line 67
When asked to apply the rewrites, run pnpm test after: the tools/list snapshots will show the change, and each changed description must …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Tool definition review
An agent picks and calls a tool from its definition alone: what tools/list
returns. This review judges that text the way Glama's score does, on the part
no test can: whether it is good. pnpm test already enforces the floor
(rules A/B/C in tests/unit/tools/tool-quality.test.ts, the tools/list gate
in tests/tool-definitions/); never report what those already fail on.
The rubric (six tool dimensions with weights, four coherence dimensions) is
described in docs/mcp-metadata.md. The per-point anchors are Glama's own,
shipped in mcp-tdqs:
node -e "import('mcp-tdqs').then(m => console.log(JSON.stringify({tool: m.TOOL_DIMENSIONS, server: m.COHERENCE_DIMENSIONS}, null, 1)))"Steps
- Deterministic report. Run
pnpm -s tool-definitions:report. Any entry under "Problems" is a blocking finding: report it as is. If you cannot run commands, skip this step and judge from the diff alone. - Scope. Review the tools whose definition changed: the entries that
differ in
git diff <base> -- tests/integration/tools-list/(those snapshots pin every tool'stools/listentry), or the tools named in the request. Add the tools whose handler changed with an unchanged definition: check that what the description promises still holds. A new or renamed tool also opens the coherence dimensions for its namespace. - Score each tool in scope on the six dimensions, 1 to 5, against the anchors. Give no credit for text that repeats the name, the schema or the annotations. Report each dimension under 4, nothing else.
- Siblings. For a tool in scope that the report lists as naming no other
tool, ask whether an agent could pick a sibling in its place: a tool on the
same resource and the same side (two reads, or two writes with the same
verb), or a generic tool sharing its verb (
searchandsearch_users).- If one could, that is a Usage Guidelines finding: the description must say when to use this tool rather than that one. Naming the sibling is not enough; the arbitration is the point.
- If none could, there is nothing to report. Never add a sibling mention only to leave the list.
- A shadow candidate from the report is a finding only if both tools could answer the same request.
- Second opinion. Glama publishes a per-dimension justification for every tool of the released version at https://glama.ai/mcp/servers/fruggr/zendesk-mcp-server/score. When a tool in scope scores under 4.0 there, read its justification before writing yours.
Each finding
State the tool, the dimension, the sentence at fault (quoted), why an agent would misread or misuse it, and a rewrite. A rewrite must:
- drop nothing the old text said, unless it was false (
AGENTS.md, multi-agent rule); saying the same in fewer words is a win; - keep the first sentence standalone, because proxy modes show only it;
- state only Zendesk behaviour that the code, a test or the Zendesk docs back; otherwise ask instead of asserting;
- add no padding: Conciseness & Structure scores it down.
Do not ask for an outputSchema: tools return text shaped for the agent, by
decision (docs/decisions/tool-definition-quality.md).
When asked to apply the rewrites, run pnpm test after: the tools/list
snapshots will show the change, and each changed description must keep every
fact the old one stated.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

