multithink — Claude Code skill: evidence-based review
The multithink Claude Code skill adjudicates independent reads of one artifact or question using cited evidence. It preserves the original reads and records disputed findings.
MIT 130 installs ★ 1,470 updated 2026-10-25
npx skills add LilMGenius/paperthin --skill multithink -a claude-codeWhat multithink does
multithink checks the evidence behind collected independent reads, then labels each claim as supported, contested, unsupported, or insufficient. Agreement counts do not affect those labels. The report retains each original and tracks where findings came from, including reasons for rejecting citations.
The skill does not recruit reviewers or modify the artifact under review. The caller may run one exchange limited to disputed citations; the process stops after that round. To preserve source bytes, run node scripts/segments.mjs pack <report> <round0-file>.... The script writes a report and an index containing source paths, labels, byte offsets and lengths, plus SHA-256 hashes. If the original bytes cannot be obtained, request them and leave preservation unverified.
What it helps with
- compare independent reads of one artifact
- classify claims using supporting and opposing citations
- preserve original reads and their provenance
- record evidence-driven changes after an exchange
Example prompts for the agent
- Review these collected reads and classify each finding using only checkable evidence.
- Compare the responses to this question, preserve the originals, and produce an auditable report.
- Check the disputed citations and record which findings change because of new evidence.
How to install multithink
1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill multithink -a claude-code2Manually: put the files in place
~/.claude/skills/multithink/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/multithink/
1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill multithink -a codex2Manually: put the files in place
.agents/skills/multithink/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill multithink -a cursor2Manually: put the files in place
.cursor/skills/multithink/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install LilMGenius/paperthin skills/mesh/multithink2Manually: put the files in place
.github/skills/multithink/SKILL.md1Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill multithink -a gemini-cli2Platform CLI
gemini skills install https://github.com/LilMGenius/paperthin --path skills/mesh/multithinkManually: put the files in place
.devin/skills/multithink/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add LilMGenius/paperthin --skill multithink -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about multithink
How do I install multithink in Claude Code?
Run npx skills add LilMGenius/paperthin --skill multithink -a claude-code. The skill file is at ~/.claude/skills/multithink/SKILL.md.
What does multithink do?
It evaluates claims from existing independent reads against cited evidence, rather than counting agreement. It keeps the originals and records the reasoning in an auditable report.
What does multithink need to run?
It needs completed independent reads with their provenance, and the ability to run node scripts/segments.mjs to preserve their original bytes. It does not start reviewers itself.
Source text
SKILL.md · MIT
--- name: multithink disable-model-invocation: true description: "Adjudicate already-collected independent reads of one artifact or question into findings classified by cited evidence alone, never by how many agreed, preserving every original, with one optional caller-run exchange that circulates only disputed citations and records each evidence-caused revision." --- Receive independent reads, check their evidence, and preserve the audit trail behind each finding. ## Goal Return a flat set of final finding records plus a full auditable report containing every round-0 original verbatim. Classes describe evidence support, never a vote or a guarantee of truth. ## Workflow 1. Frame the artifact or question and receive independent round-0 reads with their provenance. Do not launch reviewers. Request missing inputs rather than inventing independent reads. Record the target version, available coverage, and spend limit for any optional exchange. 2. Preserve each round-0 original byte for byte before interpretation. From this skill's folder run `node scripts/segments.mjs pack <report> <round0-file>...`: it copies each original's raw bytes into its own delimited report segment, with labels outside the segment and no change to encoding, whitespace or line endings, and writes `<report>.index.json` listing each original's source path, label, byte offset, byte length and sha256. Exit codes: 0 packed, 2 a usage error or an unreadable original. Do not summarize a segment or overwrite it after exchange, and write the rest of the report after the segments so their recorded offsets hold. If raw bytes are unavailable, request the originals and leave preservation unverified. 3. Extract each distinct assertion or abstention into a record using the five fields below. State the finding and its class alongside the record. Keep opposing positions separately visible and retain each origin when merging duplicates. Missing information is explicit, never invented; an incomplete record stays a candidate and cannot be promoted to supported. 4. Check the cited material at its location and follow the reproduction where safe and read-only. Identify which cited support survives and which cited counter-support survives. Reject a citation that does not support the stated position, and record the reason. Apply the assertion-then-evidence rule below to each position, including unsourced assertions and abstentions. 5. Optionally prepare one exchange round for disputed evidence within the stated spend limit. Circulate only the disputed citation, location and reproduction. Never circulate another reviewer's prose, confidence, positions, position counts, or whole output. The caller returns independently obtained responses; do not launch reviewers. Preserve exact exchange payloads and responses in the report. Reclassify by the same rule and record each changed finding's prior class, final class, and the specific evidence causing the change. Prose-only revisions cannot change a class. 6. Stop when a round supplies no new independent evidence; agreement is not a stop condition. Repeated assertions or repeated citations to the same support are not new independent evidence. There is at most one exchange round: after it, stop even if new evidence arrived, state that the cap was reached, and leave unresolved findings unresolved. Without exchange, report the round-0 classifications and that no exchange occurred. 7. Emit the flat final records and full auditable report. The report includes candidates, each class's evidence rationale, panel-split provenance, evidence rejection reasons, exchange payloads and revisions if any, the stop reason, and all round-0 originals verbatim. Flag unresolved questions and the evidence needed to settle them without repairing the target. ### Finding record Every finding carries these five fields; the finding statement, class and provenance annotation are metadata, not substitutes for any field. | Field | Contents | |---|---| | source | The external document identifier or target artifact supplying support, with a checkable citation. A mind's assertion is not evidence. State explicitly when no cited support survives. | | location | The exact file and line, passage, or specific claim the finding concerns. | | reproduction | The command, passage lookup, or query another reader can repeat to inspect the evidence. Record an unavailable or unsafe reproduction as missing, not successful. | | counterevidence | What could defeat the finding, what was checked, and cited counter-support found; distinguish looked and found none from did not look. | | independence | How support arrived: external material, a re-derivation from the target, or assertion only. Record shared origins and unknown dependencies. Record model, persona and lens spread as a coverage fact, never as a weight or numeric independence score; reads the caller split by failure mode are adjudicated like any others. | ### Assertion-then-evidence rule Apply these four steps in order, per position, using only surviving cited evidence: 1. no position asserted or only cannot-determine = insufficient 2. asserted with cited support and cited counter-support = contested 3. asserted with cited support and no cited counter-support = supported 4. asserted with no surviving cited support = unsupported An explicit cannot-determine answer is abstention, not an assertion. An asserted but unsourced position is unsupported, never insufficient. Contested retains its support component: it is supported and opposed by cited evidence, not merely unsettled because readers disagree. ## Rules - Read-only: inspect the target and cited evidence; do not edit the artifact, execute mutating reproductions, or implement fixes. Producing the requested records and report is the output operation. - User invocation controls the expense of a wide review. Receive completed reads; do not independently commission a wider run. - Aggregate existing reads with at most one constrained evidence exchange. Do not perform a fresh blind cold read, split an artifact into lenses by distinct failure mode, or choose how many lenses to use. No other skill needs to be installed. - `panel-split` records who held which position as provenance only. It never enters the class. Keep it in the report and separate from evidence-based class rationales. - No raw count, distinct-origin count, or majority is a class input, confidence, weight, rank or tie breaker. Model spread describes coverage, not independent correctness; different origins can share errors. Label a single-model run as coverage only. - A sourced minority stays supported against unsourced opposition, with panel-split annotated. Report the opposing unsourced assertion as unsupported on its own evidence. If both positions have surviving cited support, preserve both sides and classify the disputed claim as contested. - The report never says a finding is true because the panel agreed. Agreement neither validates evidence nor supplies missing support. A filled record alone does not establish a source's validity. ## Verification 1. Every final finding has source, location, reproduction, counterevidence and independence; missing required information remains an explicit candidate, not a supported finding. 2. Reapply the four-step rule to each finding's assertion and surviving citations alone. Confirm unsourced assertions are unsupported, abstentions insufficient, and two-sided cited evidence contested with its support retained. 3. Inspect every class rationale and revision: no count, majority, origin spread or panel-split enters a class decision, confidence, weight, rank or tie breaker. Confirm sourced minority support survives unsourced opposition. 4. From this skill's folder run `node scripts/segments.mjs verify <report> <report>.index.json <round0-file>...` with the received round-0 originals in pack order. It re-extracts every segment by its recorded byte offset and length and checks its sha256 against the index and its bytes against the original. Exit codes: 0 every segment byte-identical, 1 a mismatch, named by segment, 2 a usage error or an unreadable input. Require exit 0, not visual similarity. Without the originals it checks digests only; report any unavailable comparison as unverified. 5. Inspect the recorded outbound exchange payloads: only disputed citation, location and reproduction were circulated, at most once. Every class change names the evidence that caused it. If no exchange occurred, say so without claiming an exchange check passed. 6. Confirm both outputs are present, the stop reason follows the evidence rule or exchange cap, unresolved findings remain visible, and the target was not modified.
© LilMGenius · MIT · LilMGenius/paperthin