CONCEPT Cited by 1 source
Less prompting is more for analytical agents¶
A design lesson for LLM agents whose task is analytical reasoning over a known data substrate (writing SQL, navigating a metadata catalog, deciding which of several investigation paths to take): prescriptive multi-step system prompts make the agent worse. High-level guidance plus free model agency outperforms micromanagement.
Canonical wiki framing¶
From the Cloudflare Skipper / Town Lake launch post:
"Less prompting is more. Early versions of Skipper had elaborate, prescriptive system prompts: 'First, use
search_datasets. Then, useget_entity_details. Then, uselist_schema_fieldsif needed...' Quality went down. The model is good at reasoning about analytical workflows; it doesn't need to be micromanaged. We replaced the prescriptive prompts with high-level guidance and let the model pick its own path. Results got better."
This is one of four explicit design lessons named in the post, alongside "tool overlap is poison", "code captures meaning", and "memory matters."
Why prescriptive prompts hurt¶
Three structural reasons (extracted, not all stated explicitly in the post):
- Brittleness across question types — a fixed sequence like "first search, then describe, then list" fits some question shapes (open-ended exploration) but not others (point lookup, comparison query, follow-up refinement). The model is forced down the wrong path on the wrong shape.
- Tool-call sequencing as model-strength leakage — modern LLMs are good at deciding the next-best tool call given the current state. Prescribing the sequence overrides the model's strength rather than complementing it.
- Token cost and noise — a multi-step prescriptive system prompt eats context, gets re-tokenised on every turn, and acts as noise the model has to attend to.
What replaces prescriptive prompts¶
The contract becomes high-level guidance + good tools:
- High-level guidance — short, declarative statements about
what the agent does and what's available. Skipper's curated
data-model pages are the canonical example: "Prefer tables
tagged 'curated'. Avoid
scratch_r2and tables tagged 'internal'. Search with data model terms (e.g., 'billing product revenue') not natural language." - Good tools — minimal, non-overlapping (tool overlap is poison), well-named, with clear schemas. The tool surface itself is the encoding of the workflow space.
The model picks its own path through the tool surface, given the guidance.
When prescriptive prompts are still useful¶
The post does not enumerate exceptions, but the implicit boundary is task domain:
- Analytical / reasoning tasks (SQL generation, investigation, navigation) — less prompting is more.
- Structured-output tasks (extraction with strict schema, classification with fixed labels) — prescriptive prompts may still help by anchoring output shape.
The Skipper context is firmly in the first category.
Composes with .meta.json and curated data-model pages¶
The lesson is the prompt-side complement to the code-as-context lesson. The agent's intelligence comes from:
- Rich grounded context at retrieval time (5 layers — see concepts/layered-grounded-context-for-data-agent).
- Minimal, well-named tools that the model picks freely.
- Short high-level guidance that anchors policy without prescribing sequence.
What's not in the prompt: the tool sequence, the table-by-table behaviour, the per-question heuristics. Those are either left to the model or moved into the tool surface itself.
Sibling concepts¶
- concepts/tool-overlap-poisons-agent-accuracy — the parallel Skipper design lesson on the tool side.
- concepts/code-as-context-for-data-agents — the parallel lesson on context substrate.
- concepts/agent-self-correction-loop — what the agent does when given freedom; closed-loop reasoning composes well with high-level-guidance prompting.
Seen in¶
- sources/2026-05-28-cloudflare-how-we-built-cloudflares-data-platform-and-an-ai-agent-on-top-of-it
— canonical wiki source. The "first, use
search_datasets. Then, useget_entity_details..." anti-pattern quoted verbatim.
Related¶
- systems/cloudflare-skipper — canonical wiki agent.
- concepts/tool-overlap-poisons-agent-accuracy — sibling design lesson.
- concepts/code-as-context-for-data-agents — sibling design lesson.
- concepts/agent-self-correction-loop — the closed-loop reasoning that benefits from this prompting approach.