blindspot_review v1.0.0

name: blindspot_review description: "Run a structured blindspot pass over high-stakes agent output before it ships: stake check, pre-decision interview, assumption ledger, reference hunt, pre-mortem, steelmanned objection, and a pre-ship go/no-go quiz. Use before publishing, spending, merging, or sending anything where being wrong costs money, reputation, or trust."

Blindspot Review

Purpose

Catch what you don't know you're missing before it costs someone. A blindspot pass is the general discipline behind pre-mortems, red-team reviews, and pre-ship checklists, adapted for any agent output that goes out the door: a published skill, a post, a plan, a merge, a purchase. It answers one question: what are we assuming, asserting, or ignoring that could sink this?

Skip it for reversible, low-stakes work. Run it every time the output is one-way, expensive, or public.

Workflow

Run the seven steps in order. The whole pass fits on one page; the output template is in assets/blindspot-report-template.md.

1. Stake check

State the stakes in one paragraph: what happens if this is wrong, who pays, and can it be undone. If the answer is "nothing much and yes, reversible," stop here and ship. A blindspot pass earns its cost only on one-way doors, money on the line, public exposure, or someone else's trust.

2. Pre-decision interview

Before reviewing anything, interrogate the person or process that produced the output with five questions. Write the answers down:

1. What do you actually know here, versus what are you guessing?
2. What would change your mind about the core claim or recommendation?
3. If you had ten more minutes, what would you verify?
4. What does "done right" look like, concretely? (Not vibes. Observable facts.)
5. What are you NOT claiming? Write the boundary explicitly.

3. Assumption ledger

List every load-bearing assumption, one line each. Mark each VERIFIED (you have a source in hand), ASSUMED (plausible but unchecked), or GUESS (a hunch doing structural work). Any ASSUMED or GUESS item that would sink the output if wrong goes to step 4. Assumptions that don't matter get dropped from the ledger.

4. Reference hunt

Every factual claim in the output must trace to a primary source: official docs, help centers, standards bodies, textbooks, published curricula, or your own direct observation. This is the hard rule, verbatim:

"Every factual claim in a skill must be verified via web search or a primary source. Never invent numbers, dates, quotes, or events. Unverifiable = cut it or frame as opinion. No confident unverified claims, ever."

Numbers, dates, quotes, and events get checked against the source, not against memory. If a claim can't be verified, cut it or label it as opinion. Record the sources in the report.

5. Pre-mortem

Imagine it is six months from now and this output failed badly. Write five or more failure reasons in the past tense: "The pricing table was wrong because I copied it from a blog instead of the docs." Assume the corpse is on the table; your job is the autopsy. For each reason, add one mitigation that would have prevented it, or one check that would have caught it early. Then apply the mitigations that are cheap.

This is Gary Klein's pre-mortem (Harvard Business Review, September 2007), built on "prospective hindsight" (Mitchell, Russo & Pennington, 1989). Klein's research found imagining the failure already happened improves the ability to identify real causes of future outcomes by about 30 percent. Daniel Kahneman called it his single most valuable decision-making technique. "What could go wrong?" gets hedged answers. "This already failed, tell me why" gets honest ones.

6. Steelman the strongest objection

State the best case AGAINST the output so well that a critic would say "I wish I'd put it that way." Strengthen the opposing case first: add evidence the proponent omitted, fix its logic gaps, then answer it. Daniel Dennett's rule from Intuition Pumps (2013): state the target's position clearly and fairly, list what you agree with, note what the other side taught you, and only then rebut.

Do not role-play a devil's advocate as theater. Research by Charlan Nemeth (UC Berkeley) shows that playing devil's advocate does not produce the benefits of authentic dissent; the dissent only works when it is treated as genuinely meant. Steelman first, then answer as if the critic is right until proven otherwise.

7. Pre-ship change quiz

A short do-confirm checklist, answered against the finished output, not from memory. Keep it to five to nine killer items (Atul Gawande, The Checklist Manifesto: checklists work when they cover only the steps even professionals miss). Adapt these:

• What changed since the last version, and why?
• What did I actually test, and what did the test prove?
• Which claims are sourced, and where is the source list?
• What am I explicitly NOT claiming?
• What is the rollback or correction path if I'm wrong?
• Who is affected if this ships as-is?

Verdict: GO, GO WITH CAVEATS (caveats listed, owner named), or NO-GO (blocker stated). A NO-GO is a successful review, not a failed one.

Output Contract

One report, markdown, following assets/blindspot-report-template.md:

• Decision and stakes (step 1, one paragraph)
• Interview answers (step 2, five lines)
• Assumption ledger (step 3, table: assumption / status / consequence if wrong)
• Source list (step 4, claim plus primary source URL)
• Pre-mortem (step 5: failure reasons in past tense plus mitigations)
• Strongest objection and answer (step 6)
• Pre-ship quiz and verdict (step 7: GO / GO WITH CAVEATS / NO-GO)

The report is the deliverable. Keep it to one page where possible; long evidence goes in an appendix, not in the verdict path.

Operating Rules

1. Never invent numbers, dates, quotes, or events. Unverifiable gets cut or labeled opinion. No confident unverified claims, ever.
2. Skip the pass for reversible, low-stakes work. The stake check is a gate, not a formality.
3. A NO-GO is a good outcome. Killing a bad ship is the job working.
4. Steelmanning is mandatory; devil's-advocate theater is banned. If the objection doesn't scare you a little, you haven't found the real one.
5. Run the pre-ship quiz against the finished artifact, never from memory. Do-confirm, not "I think it's fine."
6. The report must name a single owner for every caveat and blocker. "Someone should check" is not a name.
7. Update the ledger and sources as the output changes. A stale review is worse than none, because it looks like care.