| name | writing-for-agents |
| description | Improve documents that agents consume, including skills, AGENTS.md, CLAUDE.md, prompts, and linked instructions. Use when creating or editing agent-facing guidance whose triggers, hierarchy, completion criteria, or wording need to be clearer. |
Writing for agents
Owns: writing clarity for agent-consumed documents.
This is a writing reference. skill-creator remains the scaffold owner. This skill does not create package structure,
choose installation locations, or register a skill. When the
document is a skill, use SKILL-MECHANICS.md for invocation and
frontmatter choices, then return package validation to skill-creator.
Context pointers
A Context pointer names material outside the current context and states when to load it.
A skill description and an AGENTS.md link are both pointers. Sharpen the pointer before
inlining its target.
- Front-load the term that should trigger retrieval.
- Give each distinct branch one trigger; collapse synonymous restatements.
- Name what the target contains and the condition for reaching it.
- Remove identity already obvious from the target or surrounding document.
The two loads
- Context load is always-loaded text that spends tokens and attention on every turn.
- Cognitive load is what a person must remember exists and choose deliberately.
Move branch-specific detail behind a pointer to reduce Context load. Preserve Cognitive
load where a human decision is valuable; remove it where selection is mechanical.
Information hierarchy
Put material on the highest tier that needs it, and no higher:
- In-file step — an action needed in the current sequence.
- In-file reference — a definition or rule consulted during that sequence.
- Disclosed reference — branch-specific material reached through a Context pointer.
Co-locate a concept's definition, rules, and caveats under one heading. Split by branch or
sequence when the main path becomes obscured; do not split merely to shorten a file.
Steps and completion criteria
End every step with a Completion criterion that is both checkable and exhaustive.
Sharpen a vague bound before adding another process step. The criterion should name the
observable state that proves completion and the full set that must be accounted for.
Weak: “Improve the instructions.”
Strong: “Every pointer names its trigger branch, every local Markdown target resolves, and
no sentence duplicates an owned rule.”
Later visible steps can pull work toward premature completion. If a clear criterion still
does not prevent that behavior, disclose the later sequence across a real context boundary.
Leading words
A Leading word is a compact, reusable concept that focuses execution and retrieval.
Prefer an established term such as Seam, RED, or frontier; define a new term only when
the existing vocabulary cannot carry the distinction.
Repeat the word where the concept must fire, not its full definition. State the positive target behavior first.
Keep a prohibition only for a necessary guardrail, paired with the
behavior that should replace it.
Prune
- Keep each meaning in a Single source of truth and point to it elsewhere.
- Treat commands, configuration, and directory layout as discoverable environment state;
document only the convention, reason, or gotcha that lookup does not reveal.
- Remove stale, irrelevant, and default no-op sentences instead of polishing them.
- Re-read every pointer after pruning: a live target behind a weak pointer is still hidden.
Invocation examples
| Request | Owner |
|---|
| “Rewrite this AGENTS.md pointer so the right branch fires.” | Use this reference. |
| “Make these completion criteria observable and exhaustive.” | Use this reference. |
| “Prune duplicated guidance from this existing skill.” | Use this reference, then validate the existing package with its scaffold owner. |
| “Create a new reusable skill package and install it.” | Use skill-creator; consult this reference only for the words agents consume. |
Finish when the document's pointers, hierarchy, criteria, leading words, and ownership are
explicit; its links resolve; and each remaining line changes behavior or preserves a needed
guardrail.