Skills — เปลี่ยนความเชี่ยวชาญเป็นสัญญาที่ตรวจได้Turn Expertise into a Reviewed Skill Contract

1. The engineering problem

A good programmer carries task knowledge in habits: where files belong, which tests matter, and what a reviewer expects. A skill makes some of that knowledge explicit and reusable. The Agent Skills specification defines a directory with SKILL.md containing metadata and instructions, with optional scripts, references and assets [1]. That packaging helps discovery. It does not automatically enforce permissions, test correctness or authenticate the author of a downloaded skill.

The companion includes skills/pricing-patch/SKILL.md and a corresponding host-side SKILL contract. It names the task, version, allowed path, function signature, evidence identifier and restricted arithmetic language. The host-side contract is authoritative for enforcement. Keeping prose and code separate makes a useful review question visible: do the instructions promise something the runtime does not enforce? In a larger registry, generate shared fields from one reviewed manifest and test the correspondence rather than relying on a maintainer's memory.

2. The boundary to design

Separate selection from authorization. Jev may propose pricing, docs or unknown; the host only implements the reviewed pricing route in this small lab. A docs route therefore goes to review rather than being silently treated as pricing. This is a deliberate scope limit. The unknown route matters because forcing every issue into a known skill converts missing coverage into confident misrouting. TypeSafe's skill-suggestion cookbook demonstrates a shortlist and re-check approach [2]; our tiny catalog uses one narrow route and tests the unsupported path.

3. Read the implementation

Implementation excerpt from the companion. Run the complete lesson below with its imports and fixtures.

def semantic_input(response):
    a = response["answers"]
    if a["injection"]["noul"] >= .2:
        return Gate("review", "possible_injection")
    if a["route"]["choice"] != "pricing" or a["route"]["confidence"] < .8:
        return Gate("review", "uncertain_or_unsupported_skill")
    return Gate("pass", "semantic_fit")

SKILL.md

---
name: pricing-patch
description: Draft a pure integer-cent discount function for the demo shop. Use only for the reviewed pricing task; escalate ambiguous or unrelated changes.
metadata:
  version: "1.0.0"
---

Return one candidate object containing path, source, summary, and evidence_id.
Target: src/pricing.py, discount_cents(price, percent).
For integer price >= 0 and integer percent in [0,100], subtract the percentage
discount and round down to integer cents. The host owns argument validation.
Use one return expression with integer constants, arguments and +, -, *, //.
No imports, calls, annotations, decorators, I/O, environment access, or deployment.
Explain the arithmetic and rounding in the summary. Cite pricing-spec-v1.
All generated code is data. Host code validates and interprets a tiny expression
grammar; this skill text does not grant permissions or implement a sandbox.

Pin the skill version and hash the contract in the run manifest. A changed instruction can change behavior even when the Python functions are unchanged. Review skill updates like dependency updates: provenance, diff, allowed resources and regression cases. Do not execute a skill's scripts simply because a model selected its name. The selected skill supplies a procedure; the application supplies capabilities. In the lab, the unknown decision returns review before any candidate is generated. That observable stop is the outcome we want readers to carry into a larger skill registry.

4. Compare the outcomes

The following output comes from scripted fixtures that verify architecture behavior. It is not a measurement of live Jev accuracy, latency, or cost.

{
  "skill": "pricing-patch",
  "unknown_route": "review",
  "skill_hash": "3b1039e383ac90834c6376a65f5a080954bd49b71435e0eb22625552055aceb3"
}

5. Run, break, and explain

Part 06 code · Implementation · Tests: test_semantic_review_cannot_generate · Exercise

git clone https://github.com/Anirach/ai-core-engineering-workbench.git
cd ai-core-engineering-workbench
git checkout 18cdfcd4156d3d02acb8badf3b3a0a823113e0ed
python3 -m workbench.lesson 6
python3 -m unittest tests.test_boundaries.BoundaryTests.test_semantic_review_cannot_generate -v

Your turn

Add a documentation skill to the catalog without granting write access. Show that selecting it cannot fall through into pricing generation.

Keep three artifacts: the input you tried, the resulting state and reason, and an explanation of which boundary owns that result. Do not change an expected result merely to make a test pass.

6. Sources and limits