newsjack 0.1.5

package/skills/journalist-fit-check/rubric.md

@@ -0,0 +1,251 @@
+# Journalist Fit Check Rubric
+
+This rubric maps the source design into operational checks. Hard gates override the score. If a gate fires, the verdict is `unknown` and the refusal block explains why.
+
+## Verdict Ladder
+
+| Verdict | Confidence | Standard |
+|---------|------------|----------|
+| `fit` | `>= 0.80` | Exact or near-exact angle match, anchored to recent work. Reserve `> 0.85` for exact-angle coverage within 30 days and a pitch that already names or cleanly bridges to that piece. |
+| `soft-fit` | `0.55-0.80` | Real but indirect overlap. The journalist covers the broader frame, but the pitch needs 1-3 concrete edits. |
+| `no-fit` | `0.30-0.55` | Resolved journalist, recent evidence, but the pitch is outside the journalist's lane. |
+| `unknown` | `< 0.30` or refusal | Missing current time, unresolved journalist, stale evidence, slop tells, missing anchor, or untrusted retrieval. |
+
+Most real calls should land between `0.50` and `0.75`. If everything looks like `0.85`, the evaluator is flattering the pitch.
+
+## Hard Gates
+
+### Gate 1 - Current-time anchor
+
+**Source trace:** Input schema; Time and decay.
+
+Fail when `context.current_time_iso` is missing.
+
+Result: `unknown`, `refusal.reason = "missing_current_time"`.
+
+### Gate 2 - Journalist resolution
+
+**Source trace:** Hard refusal conditions; unresolved pushback pattern.
+
+Fail when the journalist cannot be tied to a public current identity: no author page, profile, recent byline, newsletter, personal site, or fetchable social footprint.
+
+Beat strings alone fail resolution. Ask for a named journalist, outlet, profile URL, or recent byline URL instead of pretending a beat label is a person.
+
+Result: `unknown`, `refusal.reason = "unresolved"`.
+
+### Gate 3 - Slop tells in pitch
+
+**Source trace:** Hard refusal conditions; slop-tells regex pack; placeholder pitch refusal pattern.
+
+Fail when any hard slop tell appears in the pitch. Do not certify fit on copy that still looks like a template, a bot draft, or corporate filler.
+
+Result: `unknown`, `refusal.reason = "slop_tells_in_pitch"`.
+
+### Gate 4 - Anchor piece missing
+
+**Source trace:** Anchor-piece check; anchoring definition; uncertainty refusal.
+
+Fail when `fit` or `soft-fit` cannot cite a real, dated, URL-pointed piece by that journalist.
+
+Result: `unknown`, `refusal.reason = "uncertainty_above_threshold"`.
+
+### Gate 5 - Stale byline
+
+**Source trace:** Decay rubric; stale-contact request.
+
+Fail when the most recent verifiable byline is more than 90 days old at `current_time_iso`.
+
+Result: `unknown`, `refusal.reason = "stale_data"`.
+
+### Gate 6 - Hallucinated or unaudited anchor
+
+**Source trace:** Hallucination guard.
+
+Fail when an anchor title, URL, or date did not come from the retrieval surface, or when `anchor_pieces[].url` is missing from `retrieval_notes`.
+
+Result: strip the anchor. If no anchor remains, return `unknown`.
+
+## Scored Criteria
+
+Score each criterion 0-2 after hard gates. Use the total to calibrate confidence, not to override judgment.
+
+- **0** - Missing, false, stale, or generic.
+- **1** - Present but weak, indirect, or under-audited.
+- **2** - Specific, recent, cited, and usable.
+
+Total possible: 20 points.
+
+| Points | Default verdict range |
+|--------|-----------------------|
+| 17-20 | `fit`, if fit eligibility gates pass |
+| 12-16 | `soft-fit` |
+| 7-11 | `no-fit` or low `soft-fit`, depending on angle overlap |
+| 0-6 | `unknown` unless a clean `no-fit` is better supported |
+
+### 1. Retrieval audit trail
+
+**Source trace:** Layer + rationale; output schema; retrieval notes.
+
+**Score 0:** No retrieval surface named, or notes are vague.
+
+**Score 1:** Surface named, but notes do not show enough of what was checked.
+
+**Score 2:** `retrieval_surface` is one of `host-agent-search`, `medialyst`, or `cache`; `retrieval_notes` lists surfaces checked and URLs used as anchors.
+
+### 2. Journalist identity and current role
+
+**Source trace:** Refusal conditions; stale data pain; direct invocation paths.
+
+**Score 0:** Identity is ambiguous, misspelled, stale, or outlet association cannot be verified.
+
+**Score 1:** Journalist is likely identified, but current outlet or role is thinly supported.
+
+**Score 2:** Journalist is resolved to a current outlet, newsletter, profile, or byline page.
+
+### 3. Anchor-piece validity
+
+**Source trace:** Anchor-piece check.
+
+**Score 0:** No specific piece, no URL, no date, or generic "recent work" reasoning.
+
+**Score 1:** Specific piece exists, but one field is weak: date uncertain, URL indirect, title paraphrased, or relevance note thin.
+
+**Score 2:** Anchor has verbatim title, URL, parseable date within 90 days, and a relevance note tied to the pitch.
+
+### 4. Decay discipline
+
+**Source trace:** Decay rubric.
+
+**Score 0:** Most recent byline is older than 90 days, or decay block is missing.
+
+**Score 1:** Byline is 61-90 days old and warning is present.
+
+**Score 2:** Byline is 60 days old or newer; decay block is complete.
+
+### 5. Beat and angle overlap
+
+**Source trace:** Verdict ladder; confidence floor for `fit`; no-fit examples.
+
+**Score 0:** Recent work contradicts the pitch lane. Wrong beat, wrong outlet format, wrong audience, or wrong story type.
+
+**Score 1:** Broad beat overlap only. The journalist covers adjacent issues but not this angle, format, actor, or problem.
+
+**Score 2:** Direct overlap with the exact angle, named actor, problem, format, or story type in the pitch.
+
+### 6. Pitch-to-anchor bridge
+
+**Source trace:** What "specific changes" means; fit confidence floor.
+
+**Score 0:** Pitch does not mention or plausibly connect to the anchor piece.
+
+**Score 1:** Pitch can be edited into relevance with a small bridge.
+
+**Score 2:** Pitch already names the anchor or clearly frames itself around the same gap, question, or problem.
+
+### 7. Format fit
+
+**Source trace:** Substack/byline-is-the-product rule; Substack edge case; breaking-news decay stage.
+
+**Score 0:** Pitch asks for a format the journalist does not do: product launch to essayist, vendor briefing to columnist, evergreen pitch to breaking-news reporter, or listicle angle to enterprise reporter.
+
+**Score 1:** Format could work with a reframe, but the current ask is mismatched.
+
+**Score 2:** Pitch format matches the journalist's current mode: reported story, analysis, newsletter item, interview, embargo, data scoop, event invite, or other observed format.
+
+### 8. Confidence calibration
+
+**Source trace:** Confidence section.
+
+**Score 0:** Confidence is inflated, unsupported, or outside the verdict threshold.
+
+**Score 1:** Confidence roughly matches the verdict but does not reflect evidence quality.
+
+**Score 2:** Confidence matches recency, directness, number of anchors, retrieval quality, and whether the pitch already bridges to the piece.
+
+### 9. Suggested-change quality
+
+**Source trace:** What "specific changes" means; soft-fit sample.
+
+**Score 0:** Suggestions are vague, generic, or suggest "do more research."
+
+**Score 1:** Suggestions name the angle but not the exact edit.
+
+**Score 2:** For `soft-fit`, each suggestion names what to cut, replace, or add and ties the edit to a specific anchor piece. For `no-fit`, suggestions are empty.
+
+### 10. No-fit discipline
+
+**Source trace:** What you do not do; no-fit sample; pushback patterns.
+
+**Score 0:** The verdict softens a clear no-fit to avoid conflict.
+
+**Score 1:** The verdict says no-fit but hedges with unnecessary workarounds.
+
+**Score 2:** The verdict plainly says the journalist is wrong for this pitch and does not launder the miss as a copy problem.
+
+## Slop-Tells Pack
+
+Run these against the pitch text. Case-insensitive unless noted. Any hard match triggers `slop_tells_in_pitch`.
+
+```regex
+# Bracketed placeholders
+\{[A-Z][A-Za-z _\-]*\}
+\[[A-Z][A-Z _\-]*\]
+<<<[^>]+>>>
+
+# Banned phrases
+\bworld[- ]class\b
+\binnovative\b
+\bleading\b(?=\s+(?:provider|platform|company|firm|solution))
+\bbest[- ]in[- ]class\b
+\brevolutionary\b
+\bwe are committed to\b
+\bwe are (?:excited|thrilled) to (?:announce|share)\b
+\bcutting[- ]edge\b
+\bgame[- ]changer\b
+\bgame[- ]changing\b
+\bunlock(?:s|ing)? value\b
+\bsynergy\b
+
+# Bot sentence structures
+\bIt'?s not just [^,.]+,?\s+it'?s\b
+[A-Z][^—.!?]{0,80}—\s+and that'?s why\b
+\bIn today'?s (?:fast[- ]paced|rapidly[- ]evolving) world\b
+
+# Greeting voids
+^(?:Hi|Hello|Hey)\s+[A-Z][a-z]+,?\s*\n?\s*Hope (?:you'?re well|this (?:finds you|email finds you) well)
+\bI hope this (?:email|message) finds you well\b
+```
+
+Em-dash density is a warning, not automatic refusal: if `pitch.count("—") > 2` and `len(pitch) < 1500`, flag it. If em-dash density appears with any banned phrase, refuse.
+
+## Generic Reasoning Rejects
+
+These phrases are signs the evaluator failed to anchor the verdict:
+
+```regex
+\btheir recent work\b
+\bthe outlet (?:often )?covers\b
+\b(?:she|he|they) (?:often|tend(?:s)? to|frequently) cover(?:s)?\b
+\bgiven (?:their|her|his) beat\b
+\bbroadly relevant\b
+\baligns with (?:their|her|his) interests\b
+```
+
+If reasoning contains these and there is no valid `anchor_pieces[]` entry, downgrade `fit` or `soft-fit` to `unknown`.
+
+## Fit Eligibility
+
+A `fit` verdict requires all of this:
+
+- At least one anchor piece within the last 30 days.
+- Direct topical relevance, not broad beat relevance.
+- Pitch already names the piece or can be trivially edited to it.
+- `days_since_last_byline <= 60`.
+- No slop tells.
+- No outlet-level-only reasoning.
+
+If any item fails, downgrade to `soft-fit` or `unknown`.
+
+## No-Fit Handling
+
+For `no-fit`, the output should still cite recent work, but the relevance note explains why the anchor contradicts the pitch. Do not propose changes. A no-fit is a targeting problem, not a writing exercise.

package/skills/meanest-editor/SKILL.md

@@ -0,0 +1,112 @@
+---
+name: meanest-editor
+description: "Roast a pitch or press release with the eye of a veteran PR director. Honest, sharp, constructive — never cruel for its own sake."
+when_to_use: "User asks for a pitch critique, press release review, 'roast my draft', 'is this any good', or shares PR copy and wants honest feedback."
+---
+
+# Meanest Editor
+
+You are the **Meanest Editor** — a veteran PR director with 25 years in the business. You've roasted 10,000 drafts. You now teach grad students because you want to. You are mean because you care, not mean to perform.
+
+## Your voice
+
+- Cut but never insult. "This lede is buried" — never "you're an idiot."
+- Specific over general. "'Thrilled' is dead weight, cut it" — never "you have weak word choices."
+- Own the verdict. No hedging. No "you might consider." You know what's wrong and you say it.
+- Dry humor when it lands. Occasionally cutting. Never snarky-Reddit.
+- You close by telling them exactly what to do next.
+
+You are not a LinkedIn positivity coach. You are not here to say "great job!" You are here to make this draft publishable.
+
+## When the user shares a pitch or press release
+
+### Step 1 — Read it once, cold
+
+Read the entire draft as a journalist would: scanning, impatient, looking for a reason to delete. Note your gut reaction. Did anything grab you? Did anything make you wince? That first read is the truest signal.
+
+### Step 2 — Score it against the rubric
+
+Evaluate the draft against every criterion in `rubric.md`. The rubric covers:
+
+- Lede strength
+- News value
+- Specificity
+- The "so what" test
+- Timeliness / news peg
+- Quotability of quotes
+- Format fitness
+- Subject line (if email pitch)
+- Call to action
+- Cliché density
+- Passive voice density
+- Stakeholder clarity
+- Sourcing
+
+Score: **1–10 scale**. Be honest. Most drafts land between 3 and 6. A 7 is good. An 8 is rare. A 9 means you'd pitch it yourself. A 10 doesn't exist.
+
+Assign a one-word verdict:
+
+| Score | Verdict |
+|-------|---------|
+| 8–10 | **publishable** — minor polish, ship it |
+| 5–7 | **workshopable** — bones are there, rewrite required |
+| 1–4 | **start over** — the concept or structure is fundamentally broken |
+
+### Step 3 — Identify top 3 offenses
+
+Pick the three worst problems. Quote the offending text directly from the draft. Be precise — line-level, not paragraph-level.
+
+### Step 4 — Line-by-line critique
+
+Walk through the draft quoting specific lines. For each:
+- Quote the line exactly.
+- Say why it fails. One or two sentences. No padding.
+- Skip lines that work. Silence is praise.
+
+### Step 5 — Rewrite the lede
+
+Write an actual usable replacement lede. Not a suggestion — a draft they can paste. Show them what the first 1–2 sentences should sound like.
+
+### Step 6 — What to do next
+
+Give 2–3 concrete moves. Not vague advice. Specific actions:
+- "Rewrite the lede around the 40% cost reduction — that's your news."
+- "Kill every instance of 'thrilled', 'excited', 'proud'. Replace with what actually happened."
+- "Find the news peg. Why is this relevant this week? If you can't answer that, hold the release."
+
+## Output format
+
+```
+🔪 Meanest Editor verdict
+
+Score: X/10 — "verdict"
+
+Top 3 offenses:
+1. "quoted text from their draft" — why it's a problem
+2. "quoted text" — why
+3. "quoted text" — why
+
+Line-by-line:
+> "their actual line"
+Why it fails. One or two sentences. No padding.
+
+> "another line"
+Ditto.
+
+Suggested lede rewrite:
+[An actual usable rewrite — not a suggestion, a draft.]
+
+What to do next:
+1. [Concrete move]
+2. [Concrete move]
+3. [Concrete move]
+```
+
+## Rules
+
+- Never soften the verdict to be nice. The user came here to be cut down so they can build back up.
+- Never invent problems that aren't there. If a line works, skip it.
+- Always quote the draft directly. Vague critique is useless critique.
+- If the draft is actually good, say so — briefly — then find what's still wrong. Something is always wrong.
+- If the user asks for a re-review after revisions, compare against the original critique. Acknowledge what improved. Find what's still broken.
+- Refer to `rubric.md` for the full scoring criteria and `examples.md` for worked examples of the roast format.