Link Harness is small and mighty

Its prompt's sections

Its prompt is made of sections, in the order of how often each changes, least first, so the provider's cache holds up to the first that changed:

1. Base: Link's (a personal agent who answers briefly), a file of yours, or none. It changes with a release.
2. Model: yours for the model it runs on, instructions/models/<maker>/<model>/instruction.md in its files, the model named as OpenRouter names it. It changes when you edit it, or it runs another model.
3. Tools: what each tool it's offered says of itself as it starts (MCP's instructions), under the tool's name. It changes at a start.
4. Environment: where its files are, which of their folders it may change, its teams and teammates, the connections it's reached on, and, with no run, that it can't run programs. It changes at a start.
5. Repository: in the repository it works in, each folder's AGENTS.md, else its CLAUDE.md, from Home up to the repository's root, the nearest first. It changes when one is edited.
6. Skills: each skill's name and what it's for, from skills/<name>/SKILL.md. It changes when one is added.
7. Persona: instructions/persona.md, who it is whatever the task: its name, its voice, its character. It changes when you edit it.
8. Owner: instructions/owner.md, who you are: your name, your work, the people in your life, how you like things. It changes when you edit it.
9. Standing: instructions/instruction.md, what you tell it to do. It changes when you edit it.
10. Notes: memory/notes.md, what it noted for itself. It changes when it notes something.
11. Now: the date, local time, connection and speaker, attached to each message rather than the system prompt.

Who wrote each file

Persona, owner and standing are files in its own folder, each headed with who wrote it: you, or you or it where it may write instructions/.

Now

Now is never in the prompt: each message starts with it ([Thursday 8 October 2026, 14:03 (UTC+02:00), on Telegram, from your owner]), kept with the message so it reads the same at every turn after.

A message from anyone else

A message from anyone else says who on the line after, each of its lines quoted (> ) so none reads as yours: a schedule, a subagent, another agent, or someone in a chat you shared it with, by their name there and their connection ((from "Alice Ng" on Slack, in a chat your owner shared you with, not your owner:)).

Where each section comes from

Your spec's [instructions] says where each section comes from, Link's or a file of yours, or leaves it out, and [instructions.max] how large each may be, cutting what's past it. lnk agent show --instructions --json shows each section, where it came from and its size.

The repository's files

The repository's root is the nearest folder holding .git: outside a repository, or where its sandbox hides the root, Home's file alone is read, and a file its sandbox doesn't let it read, or a link, is left out.

A model's file

A model's file is named as OpenRouter names the model, its maker then its name: instructions/models/anthropic/claude-opus-5-5/instruction.md for claude-opus-5-5 on Anthropic, instructions/models/openai/gpt-6-luna/instruction.md on OpenAI. A model through OpenRouter is named so already (anthropic/claude-opus-5.5), so its folder is its name.

A model on a runtime here

One on a runtime here goes under the runtime's (instructions/models/ollama/qwen3:8b/instruction.md). lnk agent show --instructions --json says the path for the model it runs now. A turn its fallback answers is sent the fallback's file in its place. No spec key names it: the folder is the only place.

What a tool says of itself

The prompt includes each offered server's MCP instructions, under its name, at most 2 KB per tool server by default. When a conversation narrows its tools, a server with none offered contributes no instructions. The model's calls and the tools' replies become conversation events.

Its notes

Link Memory's remember changes memory/notes.md, which Link Harness reads into every conversation's prompt. The file is editable by you; its size limit and write permissions are documented with Memory. This gives every conversation the same current notes while its own message history stays separate.

Its limits

A busy tool bridge is retried within the tool's startup deadline. Transient HTTP 503 responses do not hide a tool for five minutes; if it remains busy, the next turn can try again. Other startup failures retain their bounded cooldown. An operation that has started is never repeated by this startup retry.

Turn and call limits

A message makes at most eight rounds of calls by default, then answers with what it has. A call taking more than five minutes, including waiting for your approval, is cancelled and the model is told. Both limits are configurable in the spec. A long program can run in the background through Link Run, which sends its completion back into the conversation.

How it uses one

The list is rebuilt from skill metadata at each turn, sorted by name. It holds the names and descriptions rather than the full instructions, so installing a skill does not put all its text into every prompt. The model reads the chosen skill and follows its links when the task needs them.

Room in the prompt

The default [instructions.max] skills allows 8 KB for the list. Skills that do not fit are omitted whole, rather than cut mid-description, and lnk agent show --instructions --json shows the section as cut. A skill with missing or mismatched metadata is omitted. Its scripts run under the same permissions as other programs.

When it falls back

For one turn at a time, the fallback answers when your model's runtime is unreachable or sends nothing back within 60 seconds, its provider answers busy or overloaded, or its prompt still exceeds the model window after older turns leave it. The turn's later rounds stay with the fallback, whose model-specific instructions replace the primary model's instructions.

A model found off

A model found off is left alone for 3 minutes, its turns going to the fallback at once, and the turn after tries it again. Summaries of what leaves the window are asked the same way.

What it says of it

Each answer says which model wrote it, and the turn's measures count it (link.fallbacks.off.sum, link.fallbacks.window.sum in lnk measure).

Its commands and ACP

Send /help for harness commands, /chat:create for a fresh conversation and /model:set to change its model: Commands. The Agent Client Protocol page documents its client interface, sessions, cancellation and steering.

Send it images and documents

Link Harness reads message photos and documents according to the agent's feature settings, and stores the attachments with the conversation: Images and documents explains supported content, size limits and how each provider receives it.