Verifiable research for Claude — 8 enforceable rules, benchmarked against live sources. "Deep research" skills orchestrate more searching; fact-check skills audit finished drafts. Nobody packages the discipline layer that governs an agent's claims while it researches: live sources over memory, a citation on every claim, two independent sources for load-bearing facts, verbatim numbers, explicit [unverified] flags, and an adversarial pass before delivery. This skill is that layer — and it's the first in its niche to ship with a measured before/after benchmark.
Built and bench-tested by SkillProof, the tested Claude skills directory.
We surveyed 15 published research/fact-checking skills and protocols before writing this one. The popular deep-research pipelines (up to 1.6k★) enforce no citation discipline; the citation-discipline skills are niche post-hoc auditors; source independence checking existed only in a one-star gist. The whitespace this skill occupies:
- Two INDEPENDENT sources for load-bearing facts. Everyone counts sources; almost nobody checks that "two sources" aren't the same press release syndicated twice. Independence is defined and checked (Rule 3).
- Numbers are quoted, not manufactured. Verbatim value + unit + date; never average or merge figures from different sources; ranges over compromise values (Rule 4). Enforced nowhere else as a named rule.
- Memory is a hypothesis. For volatile facts — versions, prices, dates, "current" anything — training-data recall is a lead to verify, not evidence (Rule 1). In the benchmark this single rule flipped two questions.
- An adversarial pass with a forcing function. Search once for the opposite of your conclusion; find 2 weaknesses in your own draft or re-examine (Rule 7).
- Discipline, not a pipeline. A three-paragraph answer gets the same rules — not an eight-phase report factory (Rule 8).
git clone https://github.com/Skillproofdev/research-discipline ~/.claude/skills/research-discipline
# restart Claude Code — triggers on research questions, "what's the latest", comparisons, fact-checksOne command — the repo IS the skill.
Six research questions with objectively checkable answers — version/date lookups, pricing math, a staleness trap, a real-event trap, fact digests. Each ran twice: one Claude Sonnet agent bare, one reading this SKILL.md first. Both had full web access. Every factual claim in all 12 answers was then judged against live primary sources by independent verification agents, claim by claim.
Headline: 7 of 54 baseline claims (13.0%) were wrong or stale. With the skill: 3 of 56 (5.4%). A 59% cut, at about +10% tokens.
The two questions that show why the rules exist:
- The staleness trap. Asked for the current Claude model catalog, the baseline agent answered from a cached table without a live check — and got three context-window figures wrong. The skill agent fetched the live docs page and matched it on all 11 claims, then flagged its own answer
[single-source]because everything traced to one vendor page. The verifier's words: "matches the live documentation on every single figure." - The real-event trap. Asked about Bun in 2026, the baseline agent ran into the (true) fact that Anthropic acquired Bun in December 2025 — and dismissed it as an unverified SEO rumor, because its sources were third-party blogs and it never opened bun.com or anthropic.com. The skill agent cited both primary announcements and the verifier confirmed them live. The baseline's only "unverified" flag of the whole benchmark was pointed at a real event.
Raw results per question (wrong or stale claims, baseline → skill): Claude models 3→0 · Bun 1→1 · Deno 1→0 · GitHub repo facts 1→0 · Astro 1→1 · email pricing 0→1.
- Email pricing (0→1): the baseline computed all three Postmark tiers and found the cheapest; the skill agent priced only one tier and wrongly called it the "cheapest path" — discipline in citing didn't compensate for incomplete enumeration. Rules constrain how you support claims; they don't do the arithmetic for you.
- Bun (1→1): the skill agent repeated a tech outlet's wrong version number (1.1.13 vs 1.3.13) for a real memory-fix release — it trusted a tier-2 source's detail without cross-checking the vendor's own release notes. Lower severity than the baseline's dismissed-acquisition miss, but a miss.
- Across all six answers the skill proactively flagged ~8 claims as
[unverified]/[single-source]— every flag was judged appropriate. The baseline used one flag, and it was wrong.
Research questions, "what's the latest X", "compare A vs B", version/pricing/market lookups, "is it true that…", fact-check requests. Stays out of the way for pure coding tasks, opinions, and creative work.
We test other people's skills for a living (public methodology). Our own skills get the same treatment: controlled runs, independent verification, and the losses published next to the wins.
More skills from us: skillproof-skills index — token-discipline (−20% tokens on multi-step work), seo-translator (76k tokens/language, 10-language case study). Free tools: SKILL.md validator, token calculator.
MIT — use it, fork it, ship it.