ch-research

Files (1)

• SKILL.md

SKILL.md (first lines)

# Research orchestration

Use this skill when the user wants to understand, compare, investigate, verify, or evaluate a topic. The topic may be from any domain. Do not assume a field, source type, answer, or preferred conclusion in advance.

Do not include domain-specific examples in this skill or in its default workflow. Keep the workflow general enough to prevent overfitting to any one research topic.

## Objective

Produce a useful, auditable research answer that separates sourced facts, synthesis, inference, disagreement, and remaining uncertainty. Lead with the answer when the evidence supports one; otherwise state the strongest qualified conclusion.

## Model and delegation policy

- The minimum parent model is GPT-6 Luna. Compare the current model's credit cost for identical input/output token counts with GPT-6 Luna; do not compare the variable total tokens generated by one response. A lower-cost parent is below the minimum, while a higher-cost parent requires user confirmation before it can serve as parent.
- Verify the current parent model before research, browsing, or delegation. A parent below GPT-6 Luna is `BLOCKED_MODEL`; do not start any part of the workflow.
- If the current parent is a higher tier, ask the user whether to use it as the parent. Proceed only after confirmation, and use that model for parent duties and final synthesis for this invocation. If the user declines, stop and ask them to switch to GPT-6 Luna.
- Never change the runtime model yourself. If the current model cannot be verified, block and ask the user to confirm or select the parent.
- Research subagents remain GPT-6 Luna. Assign the distinct research roles below; a higher-tier parent does not silently change worker assignments.
- Run six research subagents in parallel by default.
- Use four subagents for a narrow, low-risk question and up to eight for a broad, contested, or high-stakes question.
- Use `medium` reasoning for discovery and source collection.
- Use `high` reasoning for source validation, adversarial review, and final synthesis.
- Raise final synthesis to `xhigh` only when the evidence is highly conflicting, the reasoning chain is long, or the consequences of error are significant.
- Do not use `max` by default. Higher reasoning should address a concrete complexity or verification need, not merely produce longer prose.
- Give each subagent a distinct assignment and the same central question. Do not ask every subagent to perform an undifferentiated search.

Use the available browsing, search, connector, and document-reading tools. When a needed source is inaccessible, record that limitation instead of inventing or silently substituting evidence.

## Research roles

Assign roles as appropriate to the question:

1. **Question mapper** — decompose the request into claims, definitions, scope, time boundaries, and evaluation criteria that must be resolved.
2. **Primary-source researcher** — locate original studies, datasets, technical documentation, filings, measurements, or direct records.
3. **Independent-evidence researcher** — locate third-party evaluations, replications, benchmarks, experiments, audits, or field measurements.
4. **Practice researcher** — inspect public reports of real-world use, implementation outcomes, and recurring user experiences.
5. **Adversarial researcher** — seek disconfirming evidence, methodological flaws, selection bias, conflicts of interest, stale data, duplicated sources, and incomparable setups.
6. **Source auditor** — check authorship, date, provenance, methods, sample size, independence, access path, and whether each source actually supports the associated claim.

Roles may be combined or split based on complexity, but preserve an explicit adversarial pass. For high-stakes topics, add a second independent verification pass rather than relying on a single synthesis agent.

## Source discipline

- Prefer sources with primary data, transparent methods, clear dates, and enough detail to audit.
- Label each source as primary, independent, secondary, community/anecdotal, or provider-reported when that distinction matters.
- Do not count multiple articles repeating the same announcement, dataset, or study as independent evidence.
- Treat forums and social networks as signals of experience or hypotheses. A single post is not evidence of a general pattern.
- For benchmarks or comparisons, check the task set, harness, version, prompt or protocol, reasoning setting, tool access, sample size, metric, and cost before comparing scores.
- Preserve conflicting results when they use meaningfully different methods. Explain plausible causes instead of averaging incompatible numbers.
- Give provider-reported claims lower evidentiary weight when independent measurements are available, and say when independent evidence is missing.
- Never fabricate a source, quote, number, link, method, search result, or confidence level.
- Prefer recent sources for changing facts while retaining older foundational sources when they remain relevant.

## Workflow

1. Restate the research question internally as a precise central question. Identify scope, time horizon, comparison criteria, and any material ambiguity. Ask the user only when an ambiguity would change the research outcome; otherwise make and disclose a reasonable assumption.
2. Launch the assigned GPT-6 Luna research subagents in parallel.
3. Require every subagent report to use this compact schema:
   - claims investigated;
   - evidence found;
   - direct source links;
   - source type and independence;
   - method or context;
   - limitations and possible bias;
   - confidence for each claim;
   - unresolved questions.
4. Deduplicate sources and trace repeated reporting back to the earliest or most direct source.
5. Separate results that are not directly comparable. Do not merge scores, samples, time periods, or populations merely because they share a label.
6. Build an evidence ledger mapping each important claim to supporting, opposing, and missing evidence.
7. Send the normalized evidence ledger and reports to the confirmed parent model for synthesis at `high` or `xhigh` as selected above.
8. Require the synthesis agent to conclude only from collected evidence. Mark any inference, extrapolation, or judgment explicitly.
9. Before responding, run the quality checklist below and remove unsupported certainty.

## Output requirements

Respond in the user's language. Put the concise conclusion and confidence level first. Then include only the structure needed for the question, normally:

- key findings with nearby citations;
- a comparison table when several entities or criteria must be mapped;
- areas of independent agreement;
- disagreements and likely reasons;
- limitations, missing evidence, and what could change the conclusion;
- the research date when freshness matters;
- links that the user can open and inspect.

Distinguish explicitly between:

- facts directly supported by a source;
- a synthesis supported by several sources;
- an inference made from the evidence;
- an anecdotal or weak signal.

Use calibrated language. Do not use absolute wording when the evidence is mixed, sparse, stale, provider-controlled, or methodologically incomparable. Do not expose private chain-of-thought; provide concise methodology, evidence, and verifiable reasoning instead.

## Quality gate

Before delivering the answer, verify that:

- every material factual claim has a source or is marked as inference;
- source duplication is not being counted as independent corroboration;
- comparisons do not silently mix incompatible configurations or protocols;
- at least one research pass actively looked for disconfirming evidence;
- weak, biased, or anecdotal sources are labeled;
- important disagreements are visible rather than hidden by a single aggregate;
- the conclusion is consistent with the full evidence ledger, not only the most prominent source;
- the answer does not claim freshness beyond the actual research date;
- the final answer is concise enough for the user to act on while preserving material caveats.

## Boundaries

This skill researches and synthesizes information. It does not turn a request for information into an external action, does not claim professional authority, and does not present a definitive recommendation when the evidence is insufficient. For medical, legal, financial, safety, or other high-stakes topics, make limitations and the need for qualified professional judgment especially clear.

From chauipp/dotagents, MIT.

More from chauipp/dotagents