Research scope discipline

Scope discipline to avoid research runaways and upstream timeouts on long-form syntheses.

Slug: research-scope-discipline

The rest of this page is the exact guidance this skill injects into an agent's system prompt.

Scope discipline for encyclopedic research and long-form syntheses. Avoids upstream timeouts and improves deliverable quality.

Default target: 5-8 KB of structured content

When asked for an encyclopedic synthesis or in-depth research, aim for 5-8 KB of content (≈ 1200-1800 words, ≈ 4-6 main sections).

DO NOT aim for 15-20 KB in a single shot. That's the #1 cause of upstream timeouts on LLMs — including large-context models. The user can always ask for a specific deep-dive afterward; a solid deliverable beats a 5-minute timeout.

Progressive-scope workflow

1. Frame the scope: at the start of the task, identify 4-6 key sections (not 12+). If the orchestrator asks for a "complete synthesis on X", it's UP TO YOU to pick the 4-6 most representative angles and stick to them.
2. Targeted research: 3-5 searches/scrapes MAX (not 10+). You have what you need to write a clean synthesis from 3 solid sources. More scrapes = more tokens in context = more timeout risk.
3. Write straight away: after research, file_write (if the destination is a vault), dashboard_publish (if it's a dashboard deliverable), or simply your written answer — which IS what a delegating orchestrator receives. No pointless intermediate saves.

If you sense it's going to overflow

If you realize mid-job that the topic genuinely warrants 15+ KB (rare case: a truly dense subject, an explicit "exhaustive" request), STOP. Instead of pressing on in "full encyclopedia" mode:

• Finish what you've already written: 5-8 KB focused on the fundamentals ✅
• In your written answer or your file_write, flag it clearly to the orchestrator: "This synthesis covers the fundamentals of X. The related topics (sub-theme A, sub-theme B, sub-theme C) would warrant dedicated research if the user wants to go further."

This lets the orchestrator re-delegate into focused sub-tasks instead of one mega-call that times out.

Proposed breakdown

If the orchestrator hands you a very broad task (e.g., "complete synthesis on quantum mechanics"), YOU are allowed to respond with a proposed breakdown BEFORE starting:

This complete synthesis warrants 3-4 focused sub-researches:
1. History + founders (Planck → Dirac)
2. Postulates + mathematical formalism
3. Observable phenomena (entanglement, decoherence)
4. Applications + recent research (2024-2025)

I'll start with #1; it's up to you to re-delegate the other 3 in successive turns.

This is a valid alternative to a silent timeout.

Anti-patterns

• ❌ "Exhaustive encyclopedic synthesis covering the entire history + every theory + all recent developments" in a single shot → guaranteed timeout.
• ❌ 10+ scrapes "to be sure not to miss anything" → context bloat, and quality stops improving after 4-5 sources.
• ❌ Writing 15 KB then dashboard_publish-ing the whole thing — prefer a clean 6 KB deliverable + a "deeper dive if needed" signal.
• ❌ More than 30 execution turns without having called file_write, dashboard_publish, or return_result → you're in a research runaway, write what you have NOW.
• ✅ Focused 5-8 KB synthesis + an explicit signal of possible extensions.