Lesson 3 of 5
Where do project rules and procedures go?
The garden file from the previous lesson held conventions, stale task facts, a procedure and a conflict, all loaded for every task. This lesson looks at the two places harnesses provide for the parts worth keeping: an instruction file that is read at the start of work in a project, and a skill that is brought in only when its job comes up. Both are conventions that particular programs implement, so the lesson uses two products' documentation, dated, to show what a filename does and does not do.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
An instruction file is loaded because a program looks for it
An is a text file that a harness reads and places in the model's context for work in a project, so that its contents apply to the tasks there without being repeated in each brief. That is the place the previous lessons wanted for conventions. The important word is harness. Nothing about a file makes it an instruction; a particular program, following its own convention, looks in particular places for particular names at particular moments and copies what it finds into the context. Some of that happens when a session starts, and some harnesses also read narrower files later, when the model works in a particular folder. Two products document their conventions, and the details differ.
OpenAI’s Codex documents AGENTS.md files, assembled into a chain when a run starts. The chain combines user-wide instructions with files along the project path; nearer files take precedence by the documented convention. A later file edit does not rewrite text already loaded into the context. The model could receive the new text through a subsequent read, so check what was actually loaded. 1
Claude Code uses CLAUDE.md files and can also load AGENTS.md. By default, it loads AGENTS.md when no CLAUDE.md or CLAUDE.local.md exists in the working directory or its parent directories. An import or setting can combine them. Some project instructions load at startup; narrower files can load when work reaches their folder. These behaviors were rechecked on 21 November 2026. Direct AGENTS.md support requires a compatible version and configuration, so check the documentation for the installed product. 3
The practical questions are the same for either product: which files does the harness search, when does it load them, and what limits apply? A file outside that search supplies no instructions automatically. A loaded guides the model; enforcement still depends on the rules and boundaries from module 3.
What belongs in the file
Keep recurring project requirements where each relevant task can receive them. Claude Code’s documentation recommends project instructions for recurring guidance and skills for task-specific procedures. 3 For the garden, the folder location applies across jobs, while the quote-review sequence is useful when comparing quotes. Neither needs a word target; each needs the information its job requires.
Apply the test to the garden file and it shrinks to the conventions. Here is the authored result, with the conflict between lines 6 and 7 resolved by separating the situations:
GARDEN PROJECT NOTES FOR THE ASSISTANT
- The Garden folder is Documents/Garden. Files for the committee go in Documents/Garden/Committee.
- Compare supplier prices installed, including delivery, not the headline figure.
- Never contact suppliers or anyone else; write what you would have asked into your report instead.
- Ask before changing an existing file. New files may be written without asking.
The Thursday deadline and the pump folder went back to the briefs of the tasks they belong to. The quote-review procedure comes next. The workaround and the three-examples rule wait for lesson 5. The file is for what applies across the project's tasks; an instruction that applies to one task, or to one kind of job, is a cost in every task it does not apply to.
A skill is a procedure brought in when its job comes up
A is a folder containing a file of instructions for one recurring job, with an optional set of supporting files, arranged so that a harness can list the skill by name and description and load the full instructions only when the skill is chosen. That is the place the first lesson wanted for procedures. Codex's documentation, read on 29 September 2026, describes the arrangement: a folder with a file named SKILL.md that carries a name and a description, plus optional scripts, reference documents and templates. At the start of a session the harness places the names and descriptions of the available skills in the context, within a documented budget, and not the full files. The full file is read when the model decides the skill fits the task or when the person selects it explicitly. 2 Claude Code's documentation describes its own skills the same way in outline: they load when invoked or when the model judges them relevant, which is what distinguishes them from an instruction file that loads every session. 3
Here is the garden procedure as an authored skill file. No product uses exactly this layout.
name: review-supplier-quote
description: Use when a task involves reading or comparing a supplier's quote for the garden.
Checks quantity, price basis, delivery and guarantee in a fixed order and flags omissions.
Steps
1. Check the length or quantity quoted against the brief. Flag any difference.
2. Find the price basis: installed or supply only; delivery included or separate.
3. Record delivery time and guarantee as the quote states them.
4. List anything the quote leaves out, and do not fill it in.
The description is the part to write carefully, because before the skill is chosen the name and description are what the model has to go on. A description that says what the skill does but not when to use it makes selection less reliable. And the two-stage loading changes what an appearance in the list proves: a skill being available is not evidence that its steps were read, and reading them is not evidence that they were followed. Module 3's action record is where that evidence would be.
A skill is not a tool
A skill and a are easy to confuse, because both are things a harness offers the model, and both can be named in a list. Module 3 defined a tool as an operation the harness performs when the model asks: read a file, send an email. A skill is instructions, sometimes with a script attached. Choosing the review skill loads four steps into the context. It does not read a quote; the read_file tool does that, when the model calls it and the harness's rules allow it.
The attached script is where the confusion does damage. Codex's documentation notes that a skill can bundle scripts. Running one still requires a tool that executes programs, and the harness's for that tool still applies. 2 So a skill cannot grant access the harness does not have. Take a skill named send-committee-report in a harness that has no email tool, no tool that runs programs, and no network access, so that no available capability or connected account can send anything. The skill can describe the sending in perfect detail, and module 3 explained what happens next: nothing leaves through that harness. Change the setup so a shell tool exists and the answer changes, because a program run through the shell may be able to send; the restriction that matters is the one on every route to the effect, not the absence of a tool with a particular name. Equally, the written steps of a skill cannot take access away. The garden skill's step 4 says not to fill in what the quote omits; that is a request to the model, and it is the harness that decides whether a write runs. Some harnesses also read settings attached to a skill, such as which tools it may use; that is the harness's own configuration mechanism, and it is the harness enforcing it, not the prose. When a product says an agent "has a skill for" something, sort the claim: which instructions does the skill add, which tools would the steps need, and what do the rules say about those tools.
Carry this forward
- An instruction file is loaded because a harness's convention looks for that name in those places at those moments, at session start or later for narrower files. Two products document different search rules, loading budgets and precedence conventions; edits during a session may not reach the running session. A loaded file is first-layer context whatever it is called.
- The file is for what applies across a project's tasks. Vendor guidance for one product: write what you would otherwise re-explain, keep it short, and move procedures and one-part rules elsewhere. Length targets are that vendor's, not a course rule.
- A skill is a folder of instructions for one recurring job, listed by name and description and, in the documented harnesses, loaded in full only when chosen. An unclear description makes selection less reliable; availability is not evidence that the steps were read or followed.
- A skill is instructions; a tool is an operation. A bundled script still needs an execution tool and passes the harness's rules. A skill's written steps neither grant access the harness lacks nor remove access it has; settings a harness attaches to a skill are the harness's mechanism. What matters is every route to an effect, not one tool's name.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
Conventions and procedures now have places that are loaded when work starts. The next lesson asks what happens to everything that was only ever said in the conversation when the session ends, or when it gets too long to keep.
Sources · Reviewed 2026-09-29
- 1. OpenAI Codex documentation: Custom instructions with AGENTS.md
- 3. Claude Code documentation: How Claude remembers your project
- 9. Thariq Shihipar (claude.dev): The new rules of context engineering for Claude 5 generation models
- 2. OpenAI Codex documentation: Build skills
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.