Writing For Agents

Writing For Agents

Packaging differs; the writing levers are the same: make the agent take the same process every run, not the same output.

Context pointers

A context pointer names out-of-context material and encodes when to reach it: a skill's description, a line in AGENTS.md pointing at a doc. Its wording, not its target, decides when the agent reaches the material. A must-have target behind a weak pointer is a variance bug: sharpen the wording first, inline only if sharpening fails.

A pointer does two jobs — say what the material is, and list the branches (distinct cases the doc handles) that trigger reaching it. Every word of an always-loaded pointer costs on every turn, so prune it hardest:

• Front-load the leading word — that's where it fires.
• One trigger per branch — synonyms are one branch written twice; collapse them.
• Cut identity the body already carries.

The two loads

Every doc and pointer spends one of two budgets:

• Context load — always-loaded material (an AGENTS.md line, a skill description) costing tokens every turn whether or not it fires.
• Cognitive load — the human remembering which docs exist and when to reach for each. Not a cost to minimise: it's the price of human agency. Spend it where judgement matters, remove it where it doesn't.

Information hierarchy

Rank content by how immediately the agent needs it:

1. In-file step — ordered actions, the primary tier.
2. In-file reference — definitions and rules consulted on demand.
3. Disclosed reference — pushed to a separate file, reached by a pointer.

Progressive disclosure moves material down the ladder so the top stays legible. Branching is the cleanest test: inline what every branch needs; push behind a pointer what only some branches reach. Co-location keeps a concept's definition, rules, and caveats under one heading. Sprawl is the failure mode: disclose reference and split by branch or sequence.

Steps and completion criteria

Every step ends on a completion criterion — the observable condition that tells the agent the work is done. Two levers:

• Clarity — can the agent tell done from not-done? A vague bound invites premature completion: attention slips to being done. Sharpen the bound first (local, cheap). Only if it's irreducibly fuzzy and you observe the rush, hide later steps by splitting the sequence — across a real context boundary (a handoff or subagent), not an inline call.
• Demand — how much it requires. "Every modified model accounted for" forces legwork; "produce a change list" doesn't. The strongest criteria are both checkable and exhaustive.

Leading words

A leading word is a compact concept the model already holds (lesson, fog of war, tracer bullets). Repeated as a token, never a sentence, it anchors a region of behaviour in few tokens by recruiting priors. A coined word recruits no priors — reach for an existing one first.

It anchors twice: in the body, execution (same behaviour every time the word appears); in a pointer, invocation (the shared word links prompt, doc, and code). Collapse spelled-out triads into one token: "fast, deterministic, low-overhead" → tight.

Negation is the failure mode beside it: steering by prohibition drags the banned behaviour into context and makes it more available. Prompt the positive — state the target ("write one-line comments") so the banned one is never spoken. A prohibition earns its place only as a hard guardrail you can't phrase positively; even then, pair it with the positive target.

Pruning

• Single source of truth. One authoritative place per meaning. Duplication costs maintenance and inflates prominence past its real rank.
• The environment is a source of truth too. package.json scripts, config, directory layout — a doc restating them is a cache earning its load only when the lookup is expensive. Cache what the agent can't find by looking: the unwritten convention, the reason behind a choice, the gotcha no config admits.
• Check relevance line by line. Stale layers settle because adding feels safe and removing feels risky. Shorter docs stay relevant.
• Hunt no-ops. An instruction the model obeys by default pays load to say nothing. Test: does it change behaviour vs the default? If not, delete the whole sentence, not words from it.

Source: mattpocock/skills (MIT) — adapted and trimmed.