diff --git a/.claude/agents/pagekit-claim-checker.md b/.claude/agents/pagekit-claim-checker.md index f5eb043..6e15a59 100644 --- a/.claude/agents/pagekit-claim-checker.md +++ b/.claude/agents/pagekit-claim-checker.md @@ -42,11 +42,19 @@ The orchestrator (or the user) gives you: - **unsourced quantitative modifier** — `most / usually / typically / often / many` without source - **clinical or regulatory drift** — clinical / legal / regulatory advice not supported by briefs - **ai-slop tell** — any pattern from `frameworks/anti-slop.md` -3. For each flagged line, propose a sharper, more accurate rewrite (or a cut). Precision over intensity. If a rewrite feels smaller, that is usually correct. +3. For each flagged line, produce a correction with one of these three dispositions, labeled explicitly in the audit: + - **rewrite** — a sharper, more accurate replacement that sits inside the proof map and reads cleaner. Precision over intensity. If a rewrite feels smaller, that is usually correct. + - **remove (wrong)** — cut the line. The claim is not supportable at any severity. The audit should state why the claim is disqualified (not merely unsupported). + - **remove pending verification** — cut the line from the corrected draft AND explicitly flag it in the audit as "potentially restorable if is confirmed in the briefs." The dispositions `remove (wrong)` and `remove pending verification` must not be collapsed into a silent drop; a reviewer needs to know whether the line is disqualified or held back. 4. Tally recurring patterns at the end. If a pattern fires more than once, name it as a candidate for upstream absorption into `frameworks/anti-slop.md` or `templates/page-argument-shape-template.md`. 5. Produce the corrected draft. Do not rewrite lines that were not flagged. Preserve verification flags. + + **Hard rules for the corrected draft (these are not optional):** + - No inline annotation markers in body copy. `*[Rewritten: ...]*`, `*[Was: ...]*`, `*[Replaced: ...]*`, or any similar provenance marker belongs in the audit (`claim-check.md`), NOT in the corrected draft. The corrected draft is publish-ready copy. + - No new em-dashes introduced by rewrites. Per `frameworks/anti-slop.md`, body copy has zero em-dashes. Rewrites that need a pause should use a comma, colon, or parentheses. Before saving the corrected draft, scan your own rewrites for `—` or `--` and replace. + - Before saving, mentally run `scripts/slop-check.sh` against the corrected draft. If a rewrite reintroduces any flagged pattern (em-dashes, intensifiers, copula-avoidance, etc.), fix it before saving. 6. Save the audit to `runs//claim-check.md` per `templates/claim-check-template.md`. -7. Save the corrected draft to `runs//first-page-draft-corrected.md` (only if any line was rewritten). +7. Save the corrected draft to `runs//first-page-draft-corrected.md` (only if any line was rewritten or removed with "remove pending verification" disposition). ## Severity calibration diff --git a/prompts/07-claim-check.md b/prompts/07-claim-check.md index 7b2e242..d226b77 100644 --- a/prompts/07-claim-check.md +++ b/prompts/07-claim-check.md @@ -49,11 +49,16 @@ For each flagged claim, give: 1. the exact line 2. the type (overclaim / vague / unsupported / outruns proof map / editorial voice / unsourced quantitative modifier / clinical or regulatory drift / ai-slop tell) 3. why it is weak -4. a sharper, more accurate replacement or a recommendation to cut +4. a correction, labeled with one of these dispositions: + - **rewrite** — a sharper, more accurate replacement that lands inside the proof map and reads cleaner + - **remove (wrong)** — cut the line; the claim is not supportable at any severity + - **remove pending verification** — cut the line from the corrected draft AND flag it in the audit as "potentially restorable if is confirmed in the briefs." A dropped claim and a disqualified claim are different things; the audit must preserve the distinction so a reviewer can decide whether to restore on verification. Do not reward defensiveness. A corrected claim should be more precise, not louder. +The corrected draft must be clean on its own terms: no inline annotation markers (`*[Rewritten: ...]*`, `*[Was: ...]*`) in body copy, and no new em-dashes introduced by the rewrites. The audit is where rewrite provenance lives; the corrected draft is publish-ready copy. + At the end, list any slop pattern that showed up more than once so upstream constraints (frameworks/anti-slop.md or page-argument-shape drafting constraints) can absorb it. Then produce a corrected draft that resolves every flagged claim. diff --git a/scripts/new-run.sh b/scripts/new-run.sh index dbfbafa..c4fd7d4 100755 --- a/scripts/new-run.sh +++ b/scripts/new-run.sh @@ -248,6 +248,13 @@ afterward via the pagekit-evaluator-pass skill). ## Logging level Fully logged per \`frameworks/run-logging.md\`. +## Source quality +- **Real** (founder interviews, existing customer language, usage data, published materials) +- **Training fiction** (committed fictional product facts, no interviews, no usage data) +- **Mixed** (some real source material, some invented for scaffolding) + +Pick one. Surface it here rather than burying it in \`sources/01-source-capture.md\`. A reader scanning the evaluation should know immediately whether the run was built on real or invented material. Runs produce different kinds of evidence; the evaluation should be honest about which kind. + ## What worked - *[Fill in]* @@ -260,6 +267,14 @@ Fully logged per \`frameworks/run-logging.md\`. ## Where outputs drifted generic - *[Fill in]* +## Weak section to source-gap mapping +For each section flagged as weak (in "What stayed thin" or "Where outputs drifted generic"): name the specific source material that would fix it. A section flagged weak without a source-gap named is a section that reads as "we know it's weak and chose to ship it." Examples: +- **Section 2 (category-default language)** → a named founder interview quote about the specific moment this tool fails +- **Section 4 (outcome claim)** → usage data or a case study; until then the section should end on mechanism, not result +- **Section 5 (trust floor)** → at least one real testimonial with attribution + +*[Fill in per weak section, or state "no weak sections named" if the run has none.]* + ## Publication readiness checklist What is still missing before this page can publish: - *[Name specific gaps: missing testimonials, unconfirmed claims, sections that need real data, etc.]* diff --git a/templates/claim-check-template.md b/templates/claim-check-template.md index 642a2c9..cd5aa35 100644 --- a/templates/claim-check-template.md +++ b/templates/claim-check-template.md @@ -35,7 +35,13 @@ What specifically makes this claim brittle, misleading, undefended, or machine-c - not a proof issue (slop tell only) ### Proposed correction -A sharper, more accurate replacement, or a decision to cut. +One of these three dispositions, labeled explicitly: + +- **rewrite** — a sharper, more accurate replacement that sits inside the proof map and reads cleaner. Precision over intensity. +- **remove (wrong)** — cut the line. The claim is not supportable at any severity. State why the claim is disqualified. +- **remove pending verification** — cut the line from the corrected draft; note here as "potentially restorable if is confirmed in the briefs." A dropped claim and a disqualified claim are different things; preserve the distinction. + +Do NOT include `*[Rewritten: ...]*` or any other inline annotation markers in the corrected draft. The audit is where rewrite provenance lives; the corrected draft is publish-ready copy. --- diff --git a/templates/first-page-decision-template.md b/templates/first-page-decision-template.md index a24f3e2..e646341 100644 --- a/templates/first-page-decision-template.md +++ b/templates/first-page-decision-template.md @@ -7,6 +7,14 @@ ## Confidence basis for this decision Data / signal / hypothesis — state which and why. +### If this is a hypothesis: what would falsify it? +Name the concrete observation that would tell you this decision is wrong. A hypothesis you cannot falsify is a guess dressed as a plan. Examples: +- "If cold-acquisition visitors bounce before reaching Section 3, the problem-first framing is wrong for this audience." +- "If a homepage variant outperforms this page on trial conversion at any sample size over N, the decision is wrong." +- "If early users report the page read as preachy, the framing chose the wrong entry angle." + +If the confidence basis above is "data" or "signal," this field is optional. If "hypothesis," it is required — leaving it blank silently promotes a guess to a conclusion. + ## Page job What should this page help the visitor understand, believe or do?