Writing Directed Reading Instructions

How to write an instructions note that a sub-agent can execute with clean context. Directed reading — "read document X through the lens of goal Y" — is the first case, but the principles apply to any delegatable work.

Core principle: frontload everything

The caller has context the sub-agent won't. The caller's job is to do the thinking — resolve references, select what's relevant, make decisions — and write it all down. The sub-agent's job is to execute, not to figure out what you meant.

A good instructions note is self-contained. A bad one has parameters to fill in.

What to include

The document(s) to read. Full paths. If there are multiple, say why each is included and what to look for in each one.

The goal, fully articulated. Not "analyze this" — say what you're looking for. If the goal comes from another note, don't just point to it — extract the relevant parts and include them. The sub-agent shouldn't have to read a whole note to find the one paragraph that matters.

Relevant context from connected notes. Run /connect or search yourself before writing instructions. Include the connections that bear on the goal — a sentence each, with paths. Don't make the sub-agent discover what you already know.

Where to focus. If you know the relevant sections, say so. "Pay attention to sections 3.1-3.3 and the prompt templates in Appendix B" saves the sub-agent from reading the whole document with equal attention.

Output location and format. Where to save the report, what to name it. If you want a specific structure, say so. If you don't, say "the goal determines the structure."

Working copy instructions. If the sub-agent will run /connect, tell it to work on a copy and delete it after. Don't leave this to the sub-agent's judgment.

Example

# Instructions: A-MEM learning operations analysis

Read kb/sources/a-mem-agentic-memory-for-llm-agents.md

The question: how does A-MEM learn? What are the concrete operations
it performs, and are they all automatic? Focus on sections 3.1-3.3
(note construction, link generation, memory evolution) and the prompt
templates in Appendix B — the action vocabulary in B.3 is especially
relevant.

Connected notes that bear on this:
- kb/notes/automating-kb-learning-is-an-open-problem.md — frames
  automated re-organization as "boiling cauldron mutations." A-MEM's
  memory evolution is an empirical instance of this.
- kb/notes/claw-learning-loops-must-improve-action-capacity-not-just-retrieval.md — argues that
  retrieval-only benchmarks miss most of what learning means. A-MEM's
  evaluations are retrieval-only.

Write a report that inventories the operations and assesses what's
missing. The goal determines the structure — no fixed template.

Save as: kb/sources/a-mem-agentic-memory-for-llm-agents.ingest.report-learning-operations.md

What NOT to include

• The search history that led you to these documents. The sub-agent doesn't need to know what you rejected.
• Hedging about whether the documents are relevant. You decided they are — commit to it.
• Instructions to "search for more sources." That's a different task; don't mix gathering and analysis.