Deliberate search

Deliberate Search

Treat research as a deliberate, budgeted loop, not a single pass: open several lines of inquiry, judge the evidence on each, drop dead ends, and synthesize across sources. The structure borrows from ReAct (reason → act → observe), Tree-of-Thoughts (branch, evaluate, prune), and LATS (select · evaluate · reflect with memory), adapted to lightweight, prompt-level research. Borrow the discipline, not the heavy machinery: keep the tree shallow (2–4 branches), the budget small, and lean on evaluation and reflection rather than literal tree-search math. Web research has no clean pass/fail reward, so the value comes from judging evidence and reflecting on dead ends — not from formal selection/backprop.

Tools — harness-agnostic

This skill needs two capabilities from whatever agent runs it. Map them to the host's built-in tools; never require an MCP or external dependency:

• SEARCH(query) — run a web query. Claude Code: WebSearch · Codex: native web search · other harnesses: their equivalent search tool.
• FETCH(url) — open and read a page. Claude Code: WebFetch · others: the equivalent fetch/read tool. Prefer it for primary sources. If the host has no fetch tool, work from search snippets, make queries more targeted, and note that you could not open the primary text.
• FINISH — stop searching and synthesize.

The search_filters in each category profile are plain query operators (site:, filetype:), so they apply identically no matter which tool backs SEARCH.

The loop

1. Frame & decompose — restate the user's real decision. Split it into 2–4 sub-questions (branches). Pick the best-fit category and load only that profile from references/category-profiles.yaml.
2. Plan the frontier — for each branch draft 1–2 queries, scoped with the category's search_filters. Rank branches by expected payoff toward the decision.
3. Loop (budget: a few rounds; stop early when confident):
   • Select the highest-value open branch.
   • Act — run a small SEARCH batch; FETCH only the most promising results (prefer primary sources). Apply search_filters on the first pass, then drop or broaden them once a branch turns weak or repetitive.
   • Evaluate — score the evidence on the category's evaluate_on plus cross-source agreement. Is the sub-question actually answered?
   • Reflect & update — append findings to the scratchpad. If a branch dead-ends or sources conflict, write a one-line reflection, prune or backtrack, and reallocate the remaining budget. Open a new sub-question if a gap appears.
4. Stop — see Stopping.
5. Synthesize — organize around the user's decision, across branches (not one source at a time). Surface agreement, conflicts, dates, and confidence; include links.

Keep a lightweight scratchpad as working memory across the loop:

Findings:    claim — source — confidence
Conflicts:   X says A; Y says B; stronger = …
Reflections: failed query pattern → switch to …
Frontier:    open branches · budget left

A ReAct-style trace per step is optional — surface it only when it helps the user follow along:

Thought:  gap or hypothesis
Search:   query
Observe:  what the top sources actually said
Reflect:  closed? next gap? dead end → reflection

Stopping

One global rule (categories carry no stop criterion of their own):

• Close a branch when its sub-question is confirmed by a source that fits the category's evaluate_on priors and — for contestable facts — corroborated by ≥2 independent sources; or when fresh queries stop yielding anything new.
• Stop the loop when the user's decision is answered with corroborated primary evidence, or the budget/returns are exhausted. Note any branch left unresolved.

Categories

Pick one; load only that profile.

• it — software engineering, cloud, security, infrastructure, developer tools, standards.
• ai — models, benchmarks, AI products, papers, provider capabilities, prompting, agents.
• shopping — products, pricing, specs, alternatives, buying advice, market comparison.
• general — broad current-events-adjacent research, business, policy, mixed-source explainers.
• Not listed? Infer a temporary category from custom_template, state the assumption in one sentence, and continue.

Each profile carries three fields, one per loop stage:

• description → category selection
• search_filters → scopes Act (recommended first-pass operators; soft, not a whitelist)
• evaluate_on → the rubric for Evaluate

Synthesis rules

• Organize around the user's decision, not the order of sources.
• Separate fact, interpretation, and uncertainty; highlight what multiple sources agree on.
• Do not average conflicting sources — explain the conflict and which source is stronger.
• Flag weak evidence: stale pages, affiliate-heavy content, missing primary confirmation.
• Keep the answer concise unless the user asks for a deep dive.

Output shape

Use when it fits:

1. Direct answer or recommendation
2. Key evidence across sources
3. Tradeoffs or disagreements
4. Short conclusion
5. Source links

References

• references/category-profiles.yaml