convert-bullshit-filter

# Convert Bullshit Filter - The Adversarial Coaching Pass

> *"Would I actually say this to someone I'm trying to help?"* - The Donahoe Method, Voice Rules

The Bullshit Filter is one question. It's also the single most important question in the Method.

This skill runs the Filter as an **adversarial coaching pass** - not a verdict. It refuses to grade copy as "passed" or "failed." Instead, it forces three concrete actions on every draft:

1. **Quote the 3 weakest lines** in the draft, even if the draft seems strong. There is always a worst sentence; this skill identifies it.
2. **Produce 3 alternative rewrites** of each weakest line, so the writer sees the actual menu of voicing choices instead of accepting an AI's first take.
3. **Diagnose the dominant voice failure pattern** in one paragraph (e.g., softeners, marketing-speak, abstract language, hedged claims, performative urgency).

The output is a **coaching artifact**. The user decides whether to ship after seeing the 3 weakest lines + 9 rewrites + the diagnosis. There is no PASSED, no FAILED, no boolean stamp. AI cannot grade its own work; this skill stops trying.

## When to Use

Trigger phrases: "bullshit filter", "run the filter", "donahoe filter", "would I say this to a friend", "read aloud test", "kill anything that sounds fake", "final voice gate", "last pass before shipping", "find the weakest lines", "rewrite the worst sentences", `/convert bullshit-filter <draft>`.

Use this skill as the **last coaching pass** before shipping. Run it AFTER:
- `convert-open` has produced the open
- `convert-three-locks` has architected the persuasion
- `convert-bullets`, `convert-proof`, `convert-close` have produced the body and close
- `convert-chute` has engineered momentum
- `convert-voice` has applied the Voice Rules

Do NOT use for:
- The full Voice Rules pass with sentence mechanics, banned words, structural moves (use `convert-voice` - Bullshit Filter is one rule among several in Voice; this is the *deeper standalone coaching* applied as a final pass)
- Method audit on existing live copy (use `convert-audit`)
- Writing from scratch (use the page-builders)
- Architecture or persuasion design (use `convert-three-locks`, `convert-four-questions`)

## Inputs

Required:
1. **The draft** - full asset or specific section. Run this on what the writer intends to ship.
2. **Read-target** - sales page / VSL spoken / email / ad / headline. Calibrates the coaching (a VSL script that reads stiff aloud needs harder rewrites than a sales page that reads stiff in print).

Optional:
- **Seller voice samples** - so the rewrites can match the seller's actual conversational voice
- **Scope** - full asset / specific section / specific paragraphs to audit
- `out_path` - caller-controlled output. Defaults via `build_report_path`.

## The Adversarial Mechanism

The standard Bullshit Filter asks: *"Would I actually say this to someone I'm trying to help?"*

This skill never lets the AI answer that question on the writer's behalf. Instead, it forces the AI to:

- **Find the 3 worst lines** even when the draft seems clean. The forced quota is the gate. AI grading its own work returns "all clear" too often; the quota stops that failure mode.
- **Write 3 different rewrites of each weakest line.** Three voicings forces the AI to demonstrate range instead of stamping approval. The writer sees options, picks one, and ships.
- **Name the dominant failure pattern in one paragraph.** This is the diagnosis the writer can carry into the next draft - not a stamp on this one.

### What "weakest line" means

A line is a candidate for "weakest" if any of the following is true:

- Sounds like writing, not talking ("It is essential to recognize that...")
- Sounds like marketing copy, not conversation ("Unlock your true potential...")
- Embarrassing to read aloud ("ARE YOU READY TO TRANSFORM YOUR LIFE?!")
- Treats the reader as a segment ("Many of our customers have found...")
- Abstract instead of specific ("significant revenue growth")
- Hedged ("Some have suggested that...", "Studies have shown..." without naming the study)
- Performative urgency or scarcity ("DON'T MISS OUT! ACT NOW!" when there's no real reason)
- Uses softeners ("Note that...", "Specifically...", "It's worth noting that...", "Importantly...")
- Uses LLM connectors ("Moreover...", "Furthermore...", "Additionally...", "In essence...")
- Uses showroom adjectives ("carefully crafted", "meticulously designed", "comprehensive", "robust")

If 0 lines feel weak, pick the 3 *least strong*. There is always a worst sentence in any draft. The skill finds it.

## Workflow

### Step 1 - Read the entire draft, end to end

At speaking pace. Don't fix anything yet. Mark candidate weak lines as you go.

### Step 2 - Force-rank the candidates

From all marked candidates, pick the **3 weakest lines**. Even if the draft seems clean, force-rank to 3. Ties broken by:
- Line that appears earliest in the draft (a weak line up top hurts more)
- Line in a load-bearing section (Open, close, P.S.)
- Line with the most layers of failure (e.g., a hedged + abstract + LLM-connector line beats a single-failure line)

### Step 3 - For each weakest line, produce 3 alternative rewrites

The 3 rewrites must demonstrate **range**, not refinement. Aim for:
- **Rewrite A** - closest to the original intent, but plain-spoken (the conservative rewrite)
- **Rewrite B** - punchier, more specific, more named-and-dated (the aggressive rewrite)
- **Rewrite C** - different angle entirely, possibly killing the line and routing the meaning elsewhere (the structural rewrite)

Each rewrite must answer YES to *"Would I actually say this to someone I'm trying to help?"* - but the three rewrites pick different things to say.

### Step 4 - Diagnose the dominant voice failure pattern

Read all 3 weakest lines together. Name the single most common voice failure pattern in one paragraph (3-5 sentences). Examples:

- *"Softeners - every weak line opens with 'Note