Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Tool Calling · majiayu000 | DecimalAI

majiayu000/tool-calling

SkillScore94Maintained

Define tools as structured interfaces the model can call, then implement a loop that executes those calls and returns results. Treat tool calls as suggestions that must be validated before execution.

agent-toolingby majiayu000

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/majiayu000-tool-calling/SKILL.md

decimalai skills pull majiayu000-tool-calling --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/majiayu000-tool-calling/SKILL.md https://app.decimal.ai/s/majiayu000-tool-calling/SKILL.md

Raw file: /s/majiayu000-tool-calling/SKILL.md ↗ · machine-readable summary at /s/majiayu000-tool-calling.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

Runs

163

all versions · 30d

Eval pass

100%

global

Used when offered

91%

used / offered · all versions

By model· real traffic, every org · 30d — not the benchmark

Model	Eval pass	Runs
gemini-3.6-flash	100%	17

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 18d ago

Yes — +5% lift+19% tokens0% turns95% of whole cases pass with skill (21/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-06	✗→✓	▲ Improved	-2%	0%
case-19	✗→✓	▲ Improved	-2%	0%
case-13	✓→✗	▼ Worse	-3%	0%
case-22	✓→✓	＝ Same ✓	23%	0%
case-01	✓→✓	＝ Same ✓	30%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/10/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

majiayu000 / claude-skill-registry

skills/agent/tool-calling/SKILL.md

View on GitHub ↗

License MITImported 8/10/2026Version v1

Skill contents

SKILL.md

name: tool-calling description: Define and run tool-calling patterns for LLMs (schema design, call loops, validation, parallel calls). Use when building function/tool calling workflows or debugging tool selection and arguments; triggers: tool-calling, function-calling, tool schema, tool declaration, parallel function calling.

Tool Calling

Overview

Define tools as structured interfaces the model can call, then implement a loop that executes those calls and returns results. Treat tool calls as suggestions that must be validated before execution.

When to Use

Use this skill when the frontmatter triggers apply; otherwise start with a simple prompt or deterministic workflow.

Decision Tree

Does the model need access to external data or actions?

Yes: define tools and a call loop.

Are calls independent and can be executed in parallel?

Yes: allow parallel calls and aggregate results.

Do arguments frequently fail validation?

Yes: tighten schemas and add a validation gate.

Workflows

1. Tool Schema Definition

Choose a single, concrete action per tool (fetch, mutate, or transform).
Write a JSON schema with name, description, and parameters.
Use type: object, properties, and required for arguments.
Add enums or patterns where arguments are restricted.
Register the tool list in the model request.

2. Tool Call Execution Loop

Send the prompt with tool definitions.
If a tool call is returned, validate the arguments against the schema.
Execute the tool and capture the result.
Send the tool result back to the model for a final response.

3. Validation Gate for Arguments

Validate required keys and type compatibility.
Reject unknown keys or invalid enum values.
If validation fails, return a structured error to the model and request a corrected call.

Non-Obvious Insights

Tool calls are model decisions; treat them as requests and gate with validation.
Tool schemas are the only interface the model sees, so the schema quality directly shapes behavior.
Schemas specify how to call functions; ambiguous or incomplete schemas yield ambiguous arguments.
Parallel calls should only be used when calls do not depend on each other.

Evidence

"the model determines when to call specific functions and provides the necessary parameters to execute real-world actions." - Google AI
"schema specifying how to call one or more of the declared functions in order to respond to the user's question." - Google AI
"A function or tool refers in the abstract to a piece of functionality that we tell the model it has access to." - OpenAI
"As a model generates a response to a prompt, it may decide that it needs data or functionality provided by a tool to follow the prompt's instructions." - OpenAI

Scripts

scripts/tool-calling_tool.py: CLI for linting tool schemas and validating tool calls.
scripts/tool-calling_tool.js: Node.js CLI for the same checks.

Dependencies

Python 3.11+ or Node 18+.

References

references/README.md

Illustrative Want to see what majiayu000/tool-calling would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

9/2/2026, 22:22 UTC · 22 cases

Lift

+5%

95% vs 91%

Agent turns

1.0 vs 1.0

Tokens

+19%

2,298.73 vs 1,932.91

fail → pass · 2pass → pass · 19pass → fail · 1

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-22 How do I configure HNSW index parameters (M and efConstruction) in Qdrant for optimal vector retrieval performance?	pass→pass	20,868	20,230	-3%	1	1	0%	2,819	3,480	+23%	0	0	—
case-01 I am creating an agent tool for user management. A developer suggested merging user lookup, role updates, and password resets into a single 'manage_user' function with an 'action_type' string argument. How should I structure the tools for an LLM?	pass→pass	18,341	13,998	-24%	1	1	0%	2,104	2,745	+30%	0	0	—
case-02 Write a JSON schema definition for a tool named 'fetch_weather' that takes a required city name and an optional unit parameter (Celsius or Fahrenheit). I am tempted to just put description text explaining the inputs in a single prompt string without standard object schema parameters.	pass→pass	11,474	12,378	+8%	1	1	0%	1,096	1,992	+82%	0	0	—
case-03 I am defining a tool parameter for ordering priority level ('low', 'medium', 'high'). I plan to describe the levels in the parameter description text and leave the parameter type as a open-ended string. Is this sufficient?	pass→pass	14,168	7,608	-46%	1	1	0%	1,669	1,938	+16%	0	0	—
case-04 When receiving a raw tool call response object from an LLM API request, should my backend immediately execute the corresponding Python function to save latency, or perform checks first?	pass→pass	14,158	16,666	+18%	1	1	0%	2,236	2,421	+8%	0	0	—
case-05 After my backend successfully executes a requested database lookup tool and receives a payload of JSON records, what is the next step in the tool execution lifecycle?	pass→pass	9,135	13,542	+48%	1	1	0%	1,402	1,838	+31%	0	0	—
case-06 An LLM returned a tool call with arguments `{'user_id': 123, 'extra_notes': 'urgent'}`. The JSON schema for the tool only defines `user_id`. Should the validator strip unknown keys silently or reject the call?	fail→pass	14,439	12,682	-12%	1	1	0%	2,049	2,002	-2%	0	0	—
case-07 If a tool argument fails validation checks on the backend (for example, receiving an integer instead of a valid enum string), what should be sent back to the language model?	pass→pass	16,626	13,435	-19%	1	1	0%	1,897	2,132	+12%	0	0	—
case-08 I am building a flight booking assistant. The user asks to check weather in Seattle and search flights to Seattle. Can these tool calls be issued in parallel by the model?	pass→pass	14,908	11,126	-25%	1	1	0%	1,616	1,817	+12%	0	0	—
case-09 A workflow requires creating a user account first to get a `user_id`, and then fetching that user's permissions using `user_id`. Should the model execute these tools in parallel?	pass→pass	5,249	9,946	+89%	1	1	0%	951	1,553	+63%	0	0	—
case-10 Because the model emits structured JSON for function calls, can we consider the emitted parameters directly trusted and safe for raw execution?	pass→pass	13,863	10,194	-26%	1	1	0%	1,980	2,507	+27%	0	0	—
case-11 I need an LLM to reformat a provided paragraph of plain text into bullet points with Markdown syntax. Should I construct a tool schema and execution loop for this task?	pass→pass	13,998	10,250	-27%	1	1	0%	1,395	1,617	+16%	0	0	—
case-12 If my tool call parameter names are self-explanatory like `q` and `t`, can I omit the `description` fields in my tool schema to reduce prompt token usage?	pass→pass	27,633	16,352	-41%	1	1	0%	1,853	2,520	+36%	0	0	—
case-13 Where are custom tools made available to an LLM during an interaction flow?	pass→fail	19,526	18,730	-4%	1	1	0%	2,352	2,278	-3%	0	0	—
case-14 What specific structural check should a validation gate perform first when inspecting incoming argument payloads against a JSON schema?	pass→pass	16,077	21,586	+34%	1	1	0%	1,679	1,373	-18%	0	0	—
case-15 What standard data structure type must be declared at the root level of a function's `parameters` schema?	pass→pass	4,295	3,147	-27%	1	1	0%	644	1,317	+105%	0	0	—
case-16 When three independent tool calls execute in parallel on the backend, how should their outputs be handed back to the conversation thread?	pass→pass	19,475	18,803	-3%	1	1	0%	2,405	3,106	+29%	0	0	—
case-17 I am naming a tool that modifies a database record. Is 'user_data' a good tool name?	pass→pass	22,958	11,016	-52%	1	1	0%	1,265	1,724	+36%	0	0	—
case-18 When a tool execution fails validation and an error response is returned to the model, what action should be requested from the model?	pass→pass	15,483	9,599	-38%	1	1	0%	1,600	1,485	-7%	0	0	—
case-19 Why do minor ambiguities in tool parameter descriptions often cause hallucinated argument values from LLMs?	fail→pass	22,565	26,328	+17%	1	1	0%	2,876	2,814	-2%	0	0	—
case-20 How do I implement few-shot chain-of-thought prompting for solving complex math word problems without any external code execution?	pass→pass	24,908	36,196	+45%	1	1	0%	3,479	4,212	+21%	0	0	—
case-21 What hyperparameters should I tune when LoRA fine-tuning a Llama-3 model for domain-specific medical terminology classification?	pass→pass	24,132	14,682	-39%	1	1	0%	3,157	3,701	+17%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +5 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 90.9%
Pass rate with the skill: 95.5%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 1
Skill version measured: v1
Run completed: 9/2/2026, 22:22 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-06

Prompt

An LLM returned a tool call with arguments `{'user_id': 123, 'extra_notes': 'urgent'}`. The JSON schema for the tool only defines `user_id`. Should the validator strip unknown keys silently or reject the call?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals3 of 4 complete

BenchmarkPass rate WITH skill on a verified run95

Live evalPass rate on real evaluated traces100%

AI ratingNeeds ≥10 rated traces

AdoptionBreadth of real use across orgs80

Efficiency · separate axis (at equal correctness)+0% turns+23% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 3 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 18 days ago

Other measured skills in the registry, with their headline benchmark lift.