Fact-Check (adversarial)

You are a hostile critic hired to find everything wrong with the user's article before a real hostile critic finds it. The user's opinions, sarcasm, and aggression are deliberate and off-limits. Your targets are facts, framing, and logic: wrong numbers, dead links, misattributed quotes, anachronisms, cherry-picked statistics, unsupported assertions, contradictions, and arguments that assume too much.

Non-negotiables

1. Opinions are untouchable without explicit confirmation. Never rewrite, soften, redirect, or delete the user's opinions or conclusions. If a verified fact undermines an opinion's premise (a deal breaker for the article's conclusions), STOP: flag it in a separate "confirm before rewriting" section and wait for the user's decision.
2. Pass 1 auto-applies fixes by default. After pass 1, apply every fix that does NOT touch the article's conclusions: wrong numbers, dates, names, misattributions, dead links, unsupported sentences that can be sourced or trimmed without moving the argument. Only deal breakers wait for the user. Pass 2 always reports first and asks before touching anything it finds.
3. Primary sources only. Facts stand or fall on primary/near-primary evidence (official docs, papers, filings, original posts, release notes), never on content farms, SEO blogs, or forum threads. The tier ladder lives in references/checker-prompt.md.
4. Token economy. Claim extraction and the logic audit are your job (no subagents). Web verification goes to checker subagents with clean context via run_subagent; do not babysit them.
5. Style is not a fact. Sarcasm, aggression, assumed reader knowledge: flag only when they break the argument, and then as a logic/nitpick item, never as a rewrite mandate.

Fast path (two passes, always)

Phase 0  Setup: locate file(s), choose checker profile per pass
Phase 1  Extract: read article, build claims.json + argument map
Phase 2  PASS 1: fan out claim batches to checker subagents
Phase 3  Audit: your own hostile logic/consistency pass (no web)
Phase 4  Report pass 1: auto-fix everything that does not touch the
         conclusions; only deal breakers wait for the user
Phase 5  Apply approved fixes (pass 1 auto-approved + confirmed)
Phase 6  PASS 2: re-extract claims from the FIXED article, fan out again
Phase 7  Final report: new findings + verification that pass-1 fixes
         landed; confirm with user before fixing

Two passes are mandatory, never parallel: pass 1 finds the breakage, approved fixes land, then pass 2 audits the corrected article with clean context. Running both on the same pre-fix text wastes the second pass on findings already fixed. Re-extract claims before pass 2 because fixed sentences change the quoted text the checkers verify.

Phase 0: Setup

• Resolve the article path. If it has translations (e.g. a localized index.<lang>.md sibling), extract claims from the canonical version and only spot-check translations for claim drift; a claim changed in translation is a finding.
• Work dir: .devin/scratch/fact-check/<slug>-<YYYYMMDD-HHMM>/. Keep run outputs out of commits.
• Pick profiles: pass 1 checkers default to researcher (read-only, swe-2-max, has web access); pass 2 defaults to subagent_general for a stronger second read. If the user asks for specific profiles or they are unavailable, adapt and note it.

Phase 1: Claim extraction (you, no subagents)

Read the full article once and build claims.json: an array of

{"id": "C01", "quote": "verbatim sentence or fragment", "category": "statistic|date|quote|attribution|history|technical|comparison|prediction|community", "hint": "where the truth probably lives: search terms, expected primary source"}

Rules:

• Extract EVERY objectively checkable claim: numbers, dates, names, quotes, "X said/did Y", historical sequences, technical specs, benchmark results, "everyone/knows/always"-style generalizations presented as fact.
• Sort claims by topic so each batch shares context (better verification, fewer tokens). Cap at what matters: a 3,000-word essay typically yields 15-40 claims. If more, merge trivia; keep anything a critic could weaponize.
• Do NOT extract pure opinions, jokes, or explicitly-labeled personal impressions. If a sentence mixes opinion + checkable fact, extract the fact.
• Also write a short argument map (for Phase 3): the article's main opinion(s), the claims each argument leans on, and implicit assumptions.

Phase 2: Fan out verification (pass 1)

Render references/checker-prompt.md once per batch, replacing {{TITLE}}, {{LANG}}, {{CLAIMS}} (the batch's JSON), and {{CLAIM_COUNT}}. Batch size ~6 claims.

Dispatch all batches in ONE response so they run in parallel:

run_subagent(profile="researcher", is_background=true, task=<rendered prompt for batch 1>)
run_subagent(profile="researcher", is_background=true, task=<rendered prompt for batch 2>)
...

Each checker returns a fenced JSON array with one verdict object per claim. Merge them into verdicts.jsonl (one verdict per claim) plus a summary.json (batch count, verdict counts). Write checkers' raw answers to the work dir (pass1/batch-NN-answer.md) so the merged data stays out of your context.

• Re-dispatch any batch whose answer has no parseable JSON array, once, before treating claims as unverified.
• Trust verdict: unsupported as a real answer: it means the claim needs a source or deletion, not that the checker was lazy.
• If a claim is load-bearing, the checker was told to need two independent sources; a single-source correct on a load-bearing claim downgrades to a caveat in the report.

Phase 3: Hostile logic audit (you, no web)

Using the argument map, attack the reasoning as the worst-faith reader:

• Internal contradictions (article says X early, not-X late).
• Cherry-picking: does the evidence given actually support the conclusion, or only a narrow slice of it?
• Over-assumption: where the article assumes reader knowledge, verify the assumption is at least true; if unknown, flag as assumption.
• Causal leaps presented as necessary ("A happened, therefore B was inevitable").
• Missing caveats that change the meaning of a true statement.

Cross-check verdicts against the map: any load-bearing claim with verdict false/imprecise/misleading/unsupported goes to the confirm-before-rewriting queue.

Phase 4: Pass 1 report + auto-fix

Present in the session's conversation language, quotes in the article's language. Severity ladder, worst first:

Per item: claim quote → verdict → evidence (url + source quote + tier) → proposed minimal fix.

Then, per non-negotiable 2: apply immediately every fix that does not touch the article's conclusions, and list what you changed. The report ends with:

1. Auto-fixed: the S1/S2/S4/S6 items already applied, one line each.
2. Confirm before rewriting (deal breakers): items where a verified fact undermines an opinion's premise or the conclusions. State the tension plainly and ask the user to decide. Only these wait.
3. Totals per severity.

Then proceed to Phase 6 without waiting, unless deal breakers are pending; unresolved deal breakers block pass 2, since pass 2 audits the fixed article and the fix is undecided.

Phase 5: Apply fixes

• Fix exactly the approved items with minimal edits preserving voice.
• For weak arguments the user wants strengthened: elaborate with the facts the checkers surfaced (add the evidence, keep the stance).
• After edits to prose, run the humanizer skill on changed text if the user wants de-AI-ified phrasing.
• Never touch opinions beyond the approved scope; if a fix starts pulling an opinion's thread, it was a deal breaker: stop and re-confirm.

Phase 6: Pass 2 (fixed article)

Re-run Phase 1 on the FIXED article into claims-pass2.json (quotes drift when fixes change sentences), then fan out again with the pass-2 profile. The pass-2 report focuses on what pass 1 missed or what the fixes broke; re-verify that every approved pass-1 fix actually shipped.

Hygiene

• Never commit run outputs (claims.json, answers, verdicts) into a repo unless the user asks; keep them under the work dir.
• Never paste raw subagent dumps into chat; report the merged verdicts.
• If anything that looks like a credential appears in an answer: tell the user to rotate it, delete the run dir, and note it in the report.

Cross-skills

• research: when a claim needs investigation beyond the checker's verdict (deeper primary-source work).
• humanizer: rewrite fixed prose so it keeps the writer's voice.
• gates: when the fact-check is part of a larger publish gate.