This is an independently written synthesis of public sources, not the MEGA course body. Cedar Cycle's agent files, tool contracts, and examples are original teaching artifacts. They are illustrative, API-neutral, and have not been executed against a live service.
Learning goals
- Assign distinct roles to AGENTS.md, MCP integrations, and Agent Skills.
- Package a repeatable procedure without confusing it with permission.
- Load useful context progressively and preserve evidence through compression.
- Evaluate the behavior of an added capability before expanding its scope.
Cedar Cycle's support assistant can already investigate a ticket and save a draft. The next request is to make its process repeatable across the support repository and several agent clients. Copying one long prompt into every chat will make updates hard to review. We need a place for project guidance, a reusable investigation procedure, and a reliable connection to support systems.
AGENTS.md establishes project guidance
The AGENTS.md project describes a predictable place for agent-focused instructions such as setup commands, tests, and conventions. It complements human documentation. It does not itself install tools or enforce operating-system permissions.
For our repository, AGENTS.md should explain how to run fixture checks, where fictional policies live, and how draft artifacts are validated. Keep instructions relevant to working in this repository. A complete history of support decisions belongs in a searchable record, not in the instruction file loaded for every task.
Here is an original miniature example:
# AGENTS.md
Use fictional fixtures for development.
Run the fixture checker before changing support tool contracts.
Keep evidence identifiers in draft artifacts.
Save proposed replies for staff review.
Never put customer records or credentials in committed fixtures.The most important design choice is what backs these sentences. A fixture checker should actually exist. Production credentials should be unavailable in the development environment. A draft validator should require evidence identifiers. Instructions become more dependable when executable structure supports them.
Instruction precedence and file discovery differ across clients. Verify which files your chosen client reads and how it treats nested guidance. Avoid writing a policy whose correctness depends on a discovery behavior you have never checked.
MCP connects the agent to implemented operations
The MCP architecture documentation describes hosts, clients, and servers and the exchange of tool and resource information. A server supplies implemented capabilities; the host decides how they participate in the user experience. Protocol compatibility is useful, but it does not prove that an individual server's implementation is safe or correct.
Our support server exposes read_ticket, read_account_summary, read_policy, and save_draft. The host may discover these capabilities and supply their schemas to the model. The server still authenticates the caller, checks scope, validates input, and returns well-defined results. A discovered function name is not authorization to run it.
A local server process and a remote service create different operational concerns. A local adapter may inherit filesystem or process access. A remote adapter introduces network identity, service availability, and data transfer. In either case, inspect the actual implementation and permissions needed for the task. Use a test tenant or fixtures before granting access to real records.
Do not freeze a tutorial to an assumed protocol handshake. Pin and test the versions your actual host and server support. The current documentation can change while your installed client remains older. Our example concerns capability boundaries, not a complete wire-level implementation.
Skills package a procedure
In Did Skills Kill MCP?, Angie Jones explains the complementary roles of workflow instructions and executable integrations. A support investigation skill can describe how to use the ticket tools; it cannot create the underlying service or confer account access.
The Agent Skills format packages metadata and instructions in SKILL.md with optional supporting resources. Progressive loading lets an agent discover a skill before reading every detailed reference. Good descriptions therefore matter: they help the agent decide whether the procedure applies.
Our original skill description should name both its trigger and its limits. “Investigate Cedar Cycle support tickets and prepare evidence-linked drafts for staff review” is more useful than “Makes support amazing.” It tells a caller what artifact to expect and avoids implying delivery or refund authority.
Jones's skill design principles distinguish decisions that need consistency from those that benefit from model judgment. She also argues for explicit constraints. Apply that distinction narrowly: put mandatory artifact validation in code and leave the wording of a clear explanation to the model.
A skill is executable guidance in a broad sense, so review its bundled scripts as code. Installing a folder from an unknown source can add more than harmless prose. The review should cover commands, inputs, network access, dependencies, and any files the procedure can modify.
Context management keeps capabilities usable
Every additional tool and instruction competes for attention. If the assistant sees hundreds of operations, selecting the correct one can become harder. Start by exposing the operations needed for the current role. Load detailed policy references after the ticket category is known.
For our assistant, keep the basic instruction and four support tools available. Load the billing investigation skill only for billing cases. Retrieve a policy section by identifier and effective date. Keep a short evidence ledger in the working context and leave full records addressable outside it.
Compression must preserve authority and uncertainty. A compact entry can say: “Account read A7: two pending authorizations; no duplicate settlement observed; read at 10:32.” It should not become “duplicate charge confirmed.” If a later draft needs exact amounts, rehydrate A7 or read the account again rather than guessing from the summary.
Cache procedure text and policy facts differently. A versioned skill may remain valid across many tasks. Account facts can change within minutes. Record the version or observation time so the assistant can recognize when a fresh read is needed.
Worked example
Create a skill folder containing these original files:
billing-investigation/
SKILL.md
references/draft-format.md
scripts/check-draftThe illustrative SKILL.md procedure reads the supplied ticket, retrieves account evidence, loads the applicable billing policy, drafts a reply, runs check-draft, and saves the validated artifact. Its failure rule says to identify missing evidence and stop before saving unsupported guidance.
The checker accepts a document with ticket_id, policy_version, evidence_ids, and draft_text. It verifies field presence and evidence ownership deterministically. It does not decide whether the explanation is empathetic; that requires a separate quality review.
For ticket C17, the model loads the billing skill and reference format, calls the MCP-backed read tools, and creates a draft. The checker rejects an evidence identifier from C99. The agent corrects the artifact using C17's actual observations and reruns validation. Expected result: one draft tied to the correct case, with the failure and correction visible in the trace. No capability expansion was required to complete the task.
Practice
Write your own AGENTS.md excerpt, one SKILL.md, and two draft fixtures. One fixture should be valid; the other should cite evidence from a different ticket. Describe the MCP tool contract needed to support the procedure, without implementing a live integration.
Acceptance criteria: repository guidance contains a reproducible check command; the skill has a precise trigger and stopping condition; validation rejects the cross-ticket fixture; the procedure explains what to do when a policy cannot be retrieved; no file claims that instructions grant permission.
A good answer separates concerns clearly. AGENTS.md explains repository work. The skill contains the investigation sequence. The tool server implements scoped operations. The validator enforces artifact structure. If removing the prose instruction would allow a cross-tenant read, the security design is incomplete.
Check your understanding
- Can a skill replace an authenticated integration? Answer: no; it can guide use of an integration but cannot supply its implementation or authority.
- Why load detailed references on demand? Answer: to keep working context relevant while preserving access to complete information.
- What should a compressed observation retain? Answer: the source identifier, key facts, uncertainty, and any time or version needed to judge freshness.
Sources
- AGENTS.md — agent-focused project instructions.
- MCP architecture — integration participants and capabilities.
- Agent Skills overview — format and progressive disclosure.
- Angie Jones: Did Skills Kill MCP? — procedures and integrations.
- Angie Jones: 3 Design Principles for Creating Agent Skills — explicit constraints and deterministic responsibility.