bookforge-credibility-evidence-selector（Bookforge 可信度证据选择器）

Credibility Evidence Selector

When to Use

You have a claim a skeptical audience might reject, and a bag of possible proof points (stats, quotes, stories, demos, details). You need to decide which evidence to lead with — and which to cut. This skill ranks evidence across six named credibility sources and applies the Sinatra Test as a pass/fail filter to kill weak chains before they reach the audience.

Run this skill when the user is about to ship a claim without thinking hard about *why* the audience would believe it. Run it BEFORE writing the final copy, not after — the choice of evidence type reshapes the whole paragraph.

Preconditions to verify before starting:

• You know the specific claim. If missing, ask: "What is the exact one-sentence claim you need the audience to believe?"
• You have an evidence inventory (even a rough list). If missing, ask: "What proof points do you have — stats, stories, customers, experts, demos, details?"
• You know the audience's default skepticism. If missing, ask: "Who is judging this, and what is the one thing they'd say to dismiss it?"

Do NOT use this skill for:

• Fabricating evidence. This skill ranks what exists; it never invents sources.
• Designing experiments or user studies.
• Extracting the core message (use core-message-extractor first — credibility is applied to a core, not in place of one).

Context & Input Gathering

Required Context (must have — ask if missing)

• The claim: the exact sentence the audience must believe.

-> Check prompt for: a quoted claim, a pitch paragraph, a headline, a research finding. -> Check environment for: claim.md, draft.md, pitch.md, core-message.md. -> If still missing, ask: "What is the one-sentence claim — the thing the audience must believe?"

• Evidence inventory: the candidate proof points available.

-> Check prompt for: lists of customers, stats, quotes, case studies, demos, testimonials. -> Check environment for: evidence.md, proof-points.md, case-studies/, testimonials.md, customer-stories/, a data/ dir. -> If still missing, ask: "What evidence do you already have? Dump a rough list — customers, numbers, experts, stories, demos, details — I will rank them."

• Audience skepticism profile: who judges the claim, and their default objection.

-> Check prompt for: role, domain, familiarity, prior objections. -> If missing, ask: "Who will read this, and what would they say to dismiss the claim?"

Observable Context (gather from environment)

• Prior credibility drafts: existing "why trust us" copy, testimonial pages, case study PDFs.

-> Look for: about.md, trust.md, social-proof.md, testimonials.md, case-studies/. -> If present: mine for candidate evidence; treat as raw material, not final answer.

• Brand voice constraints: legal review rules, claims guidelines.

-> Look for: brand-guide.md, legal/claims.md. -> If present: note which evidence types are disallowed (e.g., "no unverified customer quotes").

Default Assumptions

• Medium: short-form text (one paragraph, one slide, one email). State this assumption if used.
• Audience default: mildly skeptical professional, not a hostile expert. Upgrade to hostile if the user says so.
• Evidence claims are truthful — the skill trusts the user's inventory and does not fact-check.

Sufficiency Threshold

• SUFFICIENT: claim + evidence inventory + audience known.
• PROCEED WITH DEFAULTS: claim + evidence known, audience inferable from context (state the inference).
• MUST ASK: claim is not a single sentence, OR evidence inventory is completely empty (cannot rank nothing).

Process

Step 1: Lock the Claim to a Single Sentence

ACTION: Write the claim as one sentence with a clear subject, verb, and measurable predicate. Reject hedges ("might," "could help," "in some cases"). Save it under a ## Claim heading in credibility-plan.md.

WHY: Credibility attaches to a specific assertion, not a vibe. If the claim is "we help teams ship faster," no evidence can be strong or weak — the claim is too fuzzy to judge. Locking the sentence forces the later steps to pick evidence that actually answers the specific assertion. Fuzzy claims are where weak evidence hides; tightening the claim is half the work of ranking the proof.

IF the claim is still multi-part ("we are fast, cheap, and reliable") -> split into one file per claim and run this skill once per claim. ELSE -> proceed.

Step 2: Inventory Evidence Into the Six Categories

ACTION: Take every proof point the user has and sort it into one of six credibility categories. Some evidence fits multiple — list it under the strongest fit only. Write the result as a ## Evidence Inventory section with six sub-headings.

The Six Credibility Sources (from Chapter 4 CREDIBLE):

1. External Authority — a recognized expert, institution, or celebrity endorsement. Example: a Nobel laureate, an industry analyst, a peer-reviewed study. *Strength:* borrows the audience's existing trust. *Weakness:* cheap to fake and audiences are now skeptical of expert voices; "analyst-quoted" pitches blur together.
2. Antiauthority (Credible Victim / Lived Experience) — an ordinary person who lived the consequences. Not an expert; their biography IS the proof. Example: Pam Laffin — age 29, mother of two, started smoking at 10, emphysema at 24, failed lung transplant — used for anti-smoking messaging because no statistic lands like a face that earned the cost. *Strength:* inverts expert fatigue. *Weakness:* emotionally heavy; must be ethical and consensual.
3. Testable Credentials ("Try It Yourself") — the audience can verify the claim in real time, with their own senses, before committing. Example: Wendy's "Where's the beef?" (you can see the tiny patty). Snickers "satisfies" (you can feel it). Barry Marshall drinking a beaker of *H. pylori* bacteria, inducing gastritis, then curing himself — a self-experiment as a public, falsifiable demonstration that ulcers are bacterial. *Strength:* collapses the skeptic's objection into a single observable act. *Weakness:* only works when real-time verification is possible.
4. Vivid, Convincing Details — specific sensory or biographical detail that an inventor could not have bothered to fake. Example: the juror study where jurors believed a mother who testified her child had a "Goofy toothbrush" he loved to brush with — the trivial detail signaled an observer telling the truth, not a litigant pitching a case. *Strength:* non-obvious signal that the speaker has been close to the thing. *Weakness:* easy to confuse with purple prose; the detail must be load-bearing, not decorative.
5. Sinatra Test (One Overwhelming Hero Example) — a single reference so dominant that the audience cannot doubt you afterward. Named after "If you can make it there, you'll make it anywhere." Example: Safexpress, an Indian logistics firm, won enterprise contracts by saying "we delivered the final Harry Potter book to 6,000 Indian bookstores in one day, sealed, no leaks" — one sentence that obliterates the due-diligence conversation. Fort Knox security contractor = you're in the running for every security contract. *Strength:* replaces a wall of credentials with one unforgettable story. *Weakness:* only works when the hero case is genuinely the hardest case in the domain.
6. Statistics-as-Illustration (Human Scale) — a number converted into a concrete, sensory comparison so the audience can feel the magnitude. Example: explaining US/Soviet nuclear arsenals not as "thousands of warheads" but as "if a BB represents the Hiroshima bomb, the world's nuclear stockpile is a garbage can full of BBs — and one BB can level a city." Stats as raw numbers are forgettable; stats as illustration are sticky. *Strength:* anchors an abstract scale to a physical image. *Weakness:* only works if the illustration is honest — if the analogy flatters the number, the audience will catch you.

WHY: Sorting forces you to see that you usually do not have six types of evidence — you have two or three, and one slot is empty. The empty slots are signals: if you have zero testable credentials, ask whether you can manufacture one (a free trial, a public benchmark, a demo). The inventory is also the cut list; evidence that does not fit any category is usually rationalization, not proof.

Anti-pattern to flag: *credentials wall.* If the entire inventory lands under "External Authority," the user is leaning on a stack of expert quotes. Audiences skim credentials walls and remember nothing. Go find one testable credential or one Sinatra hero and lead with that instead.

Step 3: Rank by the Default Preference Order

ACTION: Apply the default preference order from strongest to weakest and rank the candidates. Write the result as a ## Ranked Evidence section.

Default preference order (Chapter 4 CREDIBLE):

1. Testable Credentials — the audience verifies with their own senses.
2. Sinatra Test Hero Example — one overwhelming case.
3. Vivid Convincing Details — truthful trivia that signals firsthand knowledge.
4. Antiauthority (Credible Victim) — lived experience beats expertise for behavior change.
5. Statistics-as-Illustration — numbers rescued by human-scale analogy.
6. External Authority — expert quote as last resort, not first move.

WHY: The ordering is not arbitrary — it tracks how audiences actually process claims. Testable credentials short-circuit skepticism ("I just saw it"); Sinatra heroes override it ("if they did THAT, I believe the rest"); vivid details bypass it ("no one would bother to invent that"); antiauthorities reframe it ("a doctor would lie; this mother wouldn't"); illustrated stats make it tangible; and authorities — cheapest to fake, most diluted by overuse — come last. Inverting this order is the most common credibility mistake: leading with a quote from a Gartner report when you have a customer who would pass the Sinatra Test.

Override rules (apply when defaults don't fit):

• IF the audience is a domain of credentialed peers (scientists reviewing research, doctors reading a trial) -> External Authority moves up to #1 or #2, because the audience's own credential norms require it.
• IF the claim is a behavior-change ask aimed at the people who resist expert messaging (teens on smoking, developers on technical debt, users on security hygiene) -> Antiauthority moves to #1.
• IF the claim is about scale / magnitude / cost / risk that audiences cannot feel intuitively -> Statistics-as-Illustration moves up (nuclear-warheads-as-BBs was the only way to make that scale real).

Step 4: Run the Sinatra Test as a Pass/Fail Gate

ACTION: Take the strongest candidate from Step 3 (especially if it's a case study, customer example, or hero story) and run the Sinatra Test as a binary pass/fail. Write the result as a ## Sinatra Test section with verdict, reasoning, and — if failed — a replacement plan.

The Sinatra Test, stated formally:

An example passes the Sinatra Test when ONE example alone is enough to establish credibility in a given domain — because the example represents the *hardest case* the domain can throw at you.

The three questions:

1. Is this the hardest case? If the audience heard only this one example and nothing else, would they concede the full category? (Fort Knox security = yes, any security contract follows. Harry Potter delivery for 6,000 bookstores in one day, sealed = yes, any logistics contract follows.)
2. Is the hard part legible? Can the audience *see* why it was hard without a paragraph of explanation? Sinatra examples are self-evidently difficult — "Fort Knox" and "Harry Potter launch day" need no setup. If you must explain why the case was hard, it is not Sinatra material.
3. Is it verifiable? Can the audience check the claim if they wanted to? A Sinatra hero that cannot survive a two-minute Google search is worse than no Sinatra at all — it becomes a credibility *liability* on discovery.

WHY: The Sinatra Test exists because one dominant example is cheaper and stickier than a capabilities deck, but a *weak* hero example is worse than a deck — the audience senses the cherry-picking and punishes you. The three questions are the diagnostic: hardest case, legibly hard, checkable. Any failure means replace, don't patch. The Marshall ulcer story passes because drinking a beaker of bacteria is unambiguous; a vendor's "we helped a Fortune 500 save 15%" fails because it is neither the hardest case nor self-evidently hard.

IF the candidate passes all three questions -> mark PASS, make it the lead evidence. ELSE IF it fails one question -> mark FAIL, cut it from the lead, and either find a stronger example or fall back to the next category in the ranking (vivid details or testable credentials). ELSE IF no candidate passes -> do NOT fabricate one. Accept that the claim has no Sinatra-grade proof and move to multi-source credibility (testable demo + vivid detail + illustrated stat).

Anti-pattern to flag: *Sinatra inflation.* The user wants their best customer to pass the Sinatra Test and is willing to argue for it. If you have to argue, it fails — the test is whether the audience recognizes the hardness instantly. Three extra sentences of framing is a flunk.

Step 5: Check for the "Sound Bite / Credentials Wall" Failure Modes

ACTION: Review the ranked evidence against two named failure modes from the book and note any hits in a ## Diagnostic Notes section.

Failure mode A — The Credentials Wall. Symptom: the evidence is a list of logos, titles, and expert quotes with no single piece of proof an audience can remember. Consequence: audience forgets everything and defaults to their prior belief. Fix: replace the wall with one Sinatra hero or one testable credential. If you cannot, your claim is stronger than your evidence — shrink the claim.

Failure mode B — The Raw Statistic. Symptom: a scary or impressive number without a human-scale translation ("we prevent 3.2M attacks per day"). Consequence: the number washes past the audience. Fix: apply the nuclear-warheads-as-BBs move — convert to a sensory comparison. "3.2M attacks per day = 37 attacks every second — one every time you blink."

WHY: Credibility is a negative-space problem. The right evidence makes the claim stick, but the *wrong* evidence actively weakens it — a credentials wall signals insecurity, and a raw statistic signals that the speaker never tested their own number on a human. Running these checks after ranking catches cases where the strongest evidence in the inventory still triggers a failure mode, and the fix is not ranking-higher but *rewriting* the evidence into a stronger form.

Step 6: Write the Output Artifact

ACTION: Assemble the final credibility-plan.md with these sections, in order:

# Credibility Plan

## Claim
<one-sentence claim>

## Audience Skepticism
- Who: <role / segment>
- Default objection: <the one thing they'd say to dismiss this>

## Evidence Inventory (Six Sources)
### External Authority
- <evidence> — <one-line note>
### Antiauthority (Credible Victim)
- ...
### Testable Credentials
- ...
### Vivid Convincing Details
- ...
### Sinatra Test Hero Example
- ...
### Statistics-as-Illustration
- ...

## Ranked Evidence
1. <category> — <evidence> — <why it leads>
2. ...

## Sinatra Test
- Candidate: <the hero example tested>
- Hardest case? <yes/no + reasoning>
- Hard part legible? <yes/no + reasoning>
- Verifiable? <yes/no + reasoning>
- VERDICT: PASS | FAIL
- IF FAIL: <replacement plan>

## Diagnostic Notes
- Credentials wall? <yes/no + fix>
- Raw statistic? <yes/no + illustration>

## Recommendation
Lead with: <the one piece of evidence to put first>
Support with: <1-2 backup pieces>
Cut: <evidence that weakens the claim and should be removed>

## Assumptions
- <any inferred audience/claim/evidence assumption the user should confirm>

WHY: The artifact separates *ranking* (what I have) from *recommendation* (what to lead with) from *cut list* (what to remove). Users resist cuts; seeing them attached to specific failure modes ("this is a credentials-wall entry") makes the cut negotiable point-by-point rather than an all-or-nothing argument. The Sinatra Test section is kept verbatim (three questions, verdict, replacement) because it is the single most defensible step — a downstream reviewer who disagrees with the lead evidence has to disagree with a named, checkable test.

Inputs

• The one-sentence claim that needs credibility.
• The evidence inventory (rough dump is fine — customers, stats, experts, stories, demos, details).
• The audience skepticism profile (who judges; their default objection).
• Optional: medium constraints (character limit, slide vs paragraph vs email), legal/brand guidelines on allowed evidence.

Outputs

• credibility-plan.md — the artifact defined in Step 6, containing:

- The locked claim. - Evidence sorted into the six categories. - Ranked evidence with an explicit lead. - A Sinatra Test block with pass/fail and reasoning. - Diagnostic notes on credentials-wall and raw-statistic failure modes. - A cut list of evidence that should be removed. - An assumptions block for anything inferred.

Key Principles

• One overwhelming example beats a wall of credentials. — The Sinatra Test exists because audiences remember one hero case and forget ten expert quotes. When you have a genuine hardest-case reference, lead with it and cut the quote stack; when you don't, do not manufacture one — fall back to testable credentials or vivid details.
• Testable credentials collapse skepticism into a single observable act. — "Where's the beef?" works because the audience verifies with their own eyes. Barry Marshall drinking *H. pylori* works because the demonstration is public and falsifiable. When you can give the audience a way to check the claim themselves, nothing else in the evidence stack matters nearly as much.
• Antiauthority beats authority for behavior change. — The people most resistant to expert messaging — teens on smoking, developers on technical debt, users on security — are often best reached by someone who lived the consequences, not someone who studied them. Pam Laffin's biography outperformed any Surgeon General report because her body was the proof.
• Raw statistics are forgettable; illustrated statistics are sticky. — A number alone ("3.2M attacks per day") washes past. The same number translated to human scale ("37 every second — one every time you blink") lands. The nuclear-warheads-as-BBs move is not decoration; it is how abstract magnitude becomes emotionally real.
• Vivid details signal firsthand observation. — The juror who believed a mother because she mentioned a "Goofy toothbrush" did not decide on the toothbrush — she decided on the fact that no liar would bother inventing that detail. Load-bearing trivia is a credibility signal; decorative trivia is just prose.
• Evidence ordering is a trust hierarchy, not a taste preference. — The default preference order (testable > Sinatra > details > antiauthority > illustrated stats > authority) tracks how skepticism actually resolves. Inverting it — leading with a credentials wall — is the most common and most damaging credibility mistake in this book.
• When the evidence fails, shrink the claim. — If no evidence category produces a defensible lead, the claim is stronger than the proof supports. The fix is not to pad with weaker evidence; it is to tighten the claim until the available evidence is genuinely sufficient.

Examples

Scenario: B2B security startup with a credentials wall Trigger: Founder says "our homepage lists 14 analyst quotes and 40 customer logos but nobody books a demo. How do I make the security claim more credible?" Process: (1) Lock claim: "Our system blocks every known credential-stuffing attack pattern." (2) Inventory — 14 entries under External Authority, 2 under Vivid Details (an incident-response write-up mentioning a specific attacker TTP), 1 under Sinatra (they run security for a top-3 global bank's consumer login). (3) Rank: the bank reference is Sinatra-grade; testable credentials missing; authority dominates inventory (credentials wall). (4) Sinatra Test on the bank reference — Hardest case? Yes (top-3 global bank, consumer login = highest-volume attack surface). Legibly hard? Yes (audience instantly recognizes). Verifiable? Partial — the bank will not be named publicly. VERDICT: PASS with anonymization ("a top-3 global bank, consumer login, 400M accounts, zero successful credential-stuffing incidents in 18 months"). (5) Diagnostic flags credentials-wall (14 analyst quotes). Fix: cut to 3. (6) Recommendation: lead with the anonymized bank Sinatra hero; back with two incident-response details; cut 11 of 14 analyst quotes. Output: credibility-plan.md with the bank Sinatra as lead, the quote wall shrunk, and a note that the founder should push for a testable credential (open benchmark or public red-team) to upgrade the credibility stack further.

Scenario: Nonprofit fundraising email, abstract statistics Trigger: Fundraiser has a draft saying "3.2 million children are affected by food insecurity each year" and asks "why isn't this landing?" Process: (1) Lock claim: "Child food insecurity is a scale emergency in our region this year." (2) Inventory — 1 raw statistic (the 3.2M number), 1 External Authority (a public-health study), 0 Testable, 0 Sinatra, 0 Antiauthority, 0 Vivid Details. (3) Rank: only two entries, both weak; the statistic is a Raw Statistic failure mode and needs illustration. (4) Sinatra Test: no hero candidate exists. Do NOT fabricate. Instead: convert the statistic. "3.2M children = every elementary-school seat in the five largest school districts in the country, empty at breakfast." (5) Diagnostic: Raw Statistic HIT. Fix: the BB-translation above. Also: add one Antiauthority — a single named family from the region, their story, their specific morning (vivid-details pairing). (6) Recommendation: lead with the illustrated statistic AND the one-family Antiauthority story side-by-side (Mother Teresa effect — identifiable victim paired with illustrated scale). Cut the authority-study citation to a footnote. Output: credibility-plan.md recommending a two-piece lead (illustrated stat + identifiable family) with a note that the fundraiser should recruit an antiauthority spokesperson BEFORE next campaign.

Scenario: Academic publishing a counterintuitive research finding Trigger: Researcher has a finding that contradicts 20 years of prior consensus. Asks "how do I make this credible to reviewers who will dismiss it on sight?" Process: (1) Lock claim: "Finding X contradicts prior consensus Y because of mechanism Z." (2) Inventory — External Authority is dominant (peer-reviewed replication); Testable Credentials possible (open data + reproducible pipeline); one Antiauthority candidate (a respected prior defender of Y who reversed position); no Sinatra hero. (3) Rank: override default — audience is credentialed peers, so External Authority upgrades to #1. But lead with the Testable Credentials (open data) because a reviewer who can reproduce the result in an hour is more convinced than one reading a methods section. Antiauthority (the reversed defender) comes next — a Barry-Marshall-pattern Sinatra hero by analogy: someone who publicly reversed position is the "hardest case" this audience responds to. (4) Sinatra Test on the reversed defender: Hardest case? Yes (they were a public critic). Legibly hard? Yes (the reversal is documented). Verifiable? Yes. VERDICT: PASS. (5) Diagnostic: no credentials wall, no raw-stat failure. (6) Recommendation: lead with the open-data link (testable), follow with the reversed-defender statement (Sinatra-pattern antiauthority), support with the peer-reviewed replication. Cut nothing. Output: credibility-plan.md with a testable-credential lead, a named Sinatra-pattern antiauthority, and a note that the open dataset should be ready BEFORE submission — the credibility plan depends on it.

References

• For long-form worked examples of each of the six credibility sources, see credibility-sources-catalog.md
• For the Sinatra Test worksheet and replacement decision tree, see sinatra-test-worksheet.md

License

This skill is licensed under CC-BY-SA-4.0. Source: BookForge — *Made to Stick: Why Some Ideas Survive and Others Die* by Chip Heath and Dan Heath.

Related BookForge Skills

This skill is standalone (a Level 0 foundation skill in the Made to Stick skill set). It pairs well with core-message-extractor (run first to lock the claim) and with downstream Concrete, Emotional, and Story skills (which apply stickiness techniques once credibility is established).

Browse more BookForge skills: bookforge-skills