research-runner

---
name: research-runner
description: Answers a research question with a quote and a working link behind every claim, and deletes anything it cannot support. Use it when you need a defensible answer, not a plausible one.
---

# A researched answer with sources you can check

Every sentence survives only if a verbatim quote backs it.

## Claude does

1. **Refuses a bare topic.** "Research the semiconductor shortage" is the brief Anthropic recorded failing: one worker covered the 2021 chip crisis while two duplicated 2025. Claude asks for the question as a question, the date window, and the decision it will change. It guesses none of them.
2. **Writes the scope line:** in scope, out of scope, dates, what counts as a good source here.
3. **Sizes the job first,** on Anthropic's scaling: a simple fact, 1 pass and 3 to 10 lookups; a comparison, 2 to 4 strands at 10 to 15 lookups each; an open question, 10 or more strands. Claude states the tier.
4. **Splits it into strands that cannot overlap.** Each carries the four things a brief must have: objective, output format, sources to use and avoid, where it stops.
5. **Uses factual settings** where exposed: temperature 0.1, top-P 0.9, top-K 20. Zero when only one answer is correct.
6. **Hands the fetching over, then runs the retraction pass.** That pass, not the reading, is the research.

## Then Codex does

Codex takes the mechanical half: opening every URL, pulling verbatim quotes from long pages, confirming each link resolves, and building the source table. Bulk work with a fixed shape, where the token cost sits. Claude first fills {QUESTION}, {STRANDS} (one block each, carrying its four brief items), {DATE_WINDOW} and {SOURCE_RULES} into CODEX.md.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

## Claude checks

1. Claim by claim, find the supporting quote. If there is none, delete the claim and leave empty `[]` brackets where it stood. A visible hole beats a confident sentence.
2. No two strands answered the same thing, and none answered the wrong period. That is the recorded failure mode.
3. Citation red flags: dead links, DOIs that do not resolve, book citations with no page number. Any one makes a source unusable.
4. Spot-check three quotes word for word against the live page.
5. A strand that found nothing says "I don't have enough information to confidently assess this" and names the gap.

If any check fails, reject and rerun that strand alone. Never fill a gap from memory.

## Rules

- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- Quote first, answer second, once the material passes roughly 20,000 words.
- The question goes at the bottom, under the sources: measured at up to 30 per cent better on multi-document input.
- Log each run: question, date, model, sources fetched, sources rejected.

## If Codex is not installed

Claude fetches and quotes it all itself, in the same order, and says so.

## Built from

- Anthropic, "How we built our multi-agent research system", https://www.anthropic.com/engineering/multi-agent-research-system : their engineering account. Effort tiers, four-part brief, semiconductor failure.
- Anthropic docs, "Reduce hallucinations", https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations : official docs. Retraction pass, "not enough information" wording.
- Anthropic docs, "Prompting best practices", https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices : official docs. Query-at-the-end, 20k threshold.
- Google / Kaggle, "Prompt Engineering", Lee Boonstra, https://www.kaggle.com/whitepaper-prompt-engineering : 65-page whitepaper. Factual settings.
- Wikipedia:Signs of AI writing, https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing : community-policed catalogue. Citation red flags.

# A researched answer with sources you can check - Codex task

## You are given

- **{QUESTION}** the single scoped question. You do not answer it.
- **{STRANDS}** numbered blocks. Each one names its objective, its output format, which sources to use and avoid, and where it stops.
- **{DATE_WINDOW}** only material published or updated inside this window counts.
- **{SOURCE_RULES}** which kinds of source are allowed and which are barred.

## Produce

One file per strand, named `strand-<n>.md`, plus one `sources.md`.

Each strand file, in this order:

1. A `<quotes>` block, before any prose. Numbered verbatim quotes, 5 to 15 for the strand, each with the URL it came from and the publication date. Copy character for character. If the strand yields nothing usable, write exactly: `No relevant quotes found`.
2. The strand answer, 150 to 300 words, built only from those quotes and referencing them by number. No general knowledge.
3. A `Not found:` list of what the strand looked for and could not source.

Lookup budget per strand: 3 to 10 for a single-fact strand, 10 to 15 for a comparison strand.

`sources.md` is a table, one row per URL: number, title, publisher, publication date, date fetched, HTTP status, and whether it is a primary source or a report of one.

## Rules

- Fetch every URL. Never cite a page you did not open.
- Verbatim means verbatim. Never tidy grammar, merge sentences or paraphrase inside quote marks.
- Record the HTTP status of every fetch. A 404, a redirect to a homepage, or a paywall makes the source unusable: say so, do not quietly substitute another.
- Reject a DOI that does not resolve, a book citation with no page number, an undated page where the date matters, and any AI-written summary standing in for the source.
- Public information only. No logins, no paywalled content, nothing behind an account.
- Never invent a fact, a number or a quote. A gap stays a gap.
- When passing several documents to a model, wrap them as `<documents><document index="1"><source>URL</source><document_content>...</document_content></document></documents>`, and put the question last.
- Do not deduplicate across strands. Overlap is for the reviewer to judge.

## Return

Print to stdout:

1. The list of files written.
2. Per strand: strand number, quotes found, URLs fetched, URLs that failed and their status.
3. One line: `TOTAL QUOTES <n> / TOTAL SOURCES <n> / FAILED FETCHES <n>`.
4. The number of any strand that returned `No relevant quotes found`.