draft

name

draft

description

Draft publication-ready content with anti-AI quality enforcement. Runs Personal Voice Protocol before Substack/LinkedIn drafts to gather real material AI cannot fabricate. Applies a five-pass Quality Gate that scans for banned patterns, fixes rhythm, and verifies personal material before presenting any draft. Synthesizes from the observation log into platform-specific formats: Substack, LinkedIn, Medium. Always reads voice.md and experiment-thesis.md before generating output.

GrowthOS Publisher

Step 0: Detect & Route

Before starting any mode, determine the current state and recommend the next logical step.

Read Current State

1. Read marketing/MEMORY.md and marketing/STATUS.md
2. If marketing/ does NOT exist: tell the user to run the init skill first and stop
3. Read references/voice.md — required by Experiment Layer rules (publishable content). This includes the PVP protocol, banned patterns, rhythm rules, and vocabulary rules
4. Read references/experiment-thesis.md — required by Experiment Layer rules (observation synthesis)
5. Read references/platform-formats.md — platform conventions
6. Check if marketing/experiment/observations.md exists. If not, tell the user to log observations first and stop
7. Read experiment_chapter from STATUS.md frontmatter (default: "Chapter 1: Foundation")
8. Scan observations: count total, count unpublished (entries without - **Published:** true), count with reflections (has_reflection: true or Has reflection: true), count this month
9. Scan marketing/experiment/drafts/*.md for current draft statuses (read frontmatter of each file)

Compute State Summary

From the gathered files, calculate:

• total: count of ### OBS- entries in observations.md
• unpublished: entries without - **Published:** true
• reflected: entries with Has reflection: true
• month_count: entries where date matches current month
• score: current Autonomy Score from autonomy-log.md (latest month row)
• trend: compare current month score to previous month
• signal_count: count of signal entries in signals.md
• contradictions: signals with type: contradiction
• unused: signals without - **Used in:** field

Present State + Options

Show the summary before the platform menu:

Draft Prep — {experiment_chapter}:

Observations: {total} ({unpublished} unpublished, {reflected} with reflections)
This month: {month_count} observations
Autonomy Score: {score}% ({trend} vs last month)
Signals: {signal_count} ({contradictions} contradictions, {unused} unused)

Drafts: {draft_count} draft, {reviewed_count} reviewed, {published_count} published

Best content angles:
1. {angle from contradiction or high-value observation}
2. {angle from autonomy score trend if notable}
3. {angle from enterprise signal contradiction if exists}

Angle priority: contradictions > observations with reflections > failures with high severity > autonomy score movements.

Which platform?
1. Substack Draft — monthly practitioner journal (PVP required)
2. LinkedIn Extract — single observation post (PVP required)
3. Medium Reformat — SEO version of Substack piece
4. Content Review — manage draft status

Routing: Options 1 and 2 proceed to Step 0.5 (PVP) before drafting. Options 3 and 4 skip PVP and go directly to their steps.

Step 0.5: Personal Voice Protocol

Runs before any Substack or LinkedIn draft. Do NOT run for Medium or Content Review.

Execute PVP

Present the four PVP questions from voice.md to the user:

Before drafting, I need some raw material only you can provide.

1. SPECIFIC MOMENT — When did this happen? Date, time of day, what were you doing?
2. SENSORY DETAIL — What did you see on screen, hear, or feel in that moment?
3. PERSONAL REACTION — What did you actually think? Surprise, frustration, confusion?
4. ENTERPRISE PARALLEL — What would this look like at 500 seats? What would procurement say?

Answer as many as you can. Skip with "skip" if you want to proceed without.

Wait for the user's response.

Process Answers

• Count how many of the 4 elements have substantive answers (more than a few words)
• Store all answers as raw_material for the drafting step
• Determine confidence level:
  • high: 3-4 elements answered substantively
  • medium: 2 elements answered substantively
  • low: 0-1 elements answered, or user typed "skip"
• If user skips entirely: set pvp_status: skipped, warn that the draft will read more generic than usual, proceed

Confirm and Proceed

PVP complete — confidence: {high/medium/low}
Material collected: {list which elements were answered}
Proceeding to draft.

Route to Step 1 (Substack) or Step 2 (LinkedIn) based on the user's earlier platform choice.

Step 1: Substack Draft (Monthly Long-Form)

Monthly practitioner journal synthesizing observations into a narrative.

Gather Source Material

1. Read all current-month observations from marketing/experiment/observations.md (match date in ### OBS-{NNN} — {YYYY-MM-DD} headings against current month)
2. Read current Autonomy Score from marketing/experiment/autonomy-log.md
3. If fewer than 3 observations this month, warn and offer to include previous month's observations
4. Incorporate PVP raw_material as primary source — personal anecdotes and sensory details take priority over analytical synthesis per voice.md source material hierarchy

Identify Narrative Elements

From the gathered observations and PVP material, identify:

• Most significant autonomous action (type: autonomous-action)
• Most significant failure or escalation (highest severity preferred)
• Observations with has_reflection: true — richest source material
• Any type: contradiction observations
• Agent failures — excellent content material (per voice.md: "failures are more interesting")
• PVP moment that anchors the opening scene

Propose Narrative Arc

Present for confirmation before drafting:

Proposed arc for {Month Year}:

Opening scene: {specific moment from PVP or observation — date, sensory detail}
Development: {what happened and why it matters}
Enterprise tension: {contradiction, if available}
Self-deprecation beat: {planned moment of honest self-critique}
Autonomy score: {what the number says about experiment progress}
Close: {open question or lingering tension — never a summary}

Source observations: OBS-{NNN}, OBS-{NNN}, ...
PVP material: {which elements are being used}
Estimated length: 800-1500 words

Wait for user confirmation or adjustments before drafting.

Draft the Post

Follow ALL constraints in references/voice.md:

• Field notes tone, specific over general
• Name what broke
• Evidence-first — specific numbers, dates, outcomes
• Understated Danish sensibility — let results speak
• Opening must be a scene, not a thesis — use PVP moment or specific observation detail
• Include at least one self-deprecation moment — where you were wrong, naive, or slow to see something obvious
• Follow all banned pattern lists — no hedge words, no performative enthusiasm, no banned transitions, no generic filler, no banned structural patterns
• Follow rhythm engineering rules — varied sentence lengths, fragments, single-sentence paragraphs
• Follow vocabulary diversification rules — concrete nouns, contractions, grammatical roughness
• End with open question, not conclusion

Follow references/platform-formats.md Substack conventions:

• 800-1500 words
• Use ## headers for scannability
• Data tables for metrics where appropriate
• No listicle format, no teaser/gated content

After drafting: run Quality Gate (Step 5) before presenting to user.

Write Draft File

Ensure marketing/experiment/drafts/ directory exists (create if not).

If a file marketing/experiment/drafts/substack-{YYYY-MM}.md already exists, ask: overwrite or create revision (substack-{YYYY-MM}-v2.md)?

Write to marketing/experiment/drafts/substack-{YYYY-MM}.md with frontmatter:

---
platform: substack
status: draft
created: {YYYY-MM-DD}
published_date: ""
source_observations: [OBS-{NNN}, OBS-{NNN}]
pvp_confidence: {high/medium/low}
title: ""
word_count: {N}
rhythm_score: {std_dev}
banned_fixes: {N}
---

Update Source Observations

For each observation used in source_observations, append to its entry in marketing/experiment/observations.md:

- **Used in:** substack-{YYYY-MM}

If a Used in field already exists on the observation, append to the comma-separated list instead of adding a new line.

Confirm to User

Draft written: marketing/experiment/drafts/substack-{YYYY-MM}.md

Words: {N}
Personal material: {which PVP elements used}
Rhythm score: {std_dev} (target: >8)
Banned patterns fixed: {N}
Confidence: {high/medium/low}
Source observations: {list}
Status: draft

Session Log & State Update

1. Append to marketing/logs/{YYYY-MM-DD}.md:

   - **Publisher** — Drafted Substack piece for {Month Year}: "{title}" ({word_count} words, {N} source observations, confidence: {level})
   

2. Update marketing/STATUS.md: set last_updated to today's date, preserve all other fields

Step 2: LinkedIn Extract (~150 Words)

Single observation post for enterprise audience. One observation, one insight, maximum impact.

Select Observation

1. Read marketing/experiment/observations.md
2. Show last 5 unpublished observations (those without - **Published:** true), prioritizing observations with reflections first:

   Unpublished observations (reflections first):
   1. OBS-{NNN} ({date}) [{type}] — {what happened} [has reflection]
   2. OBS-{NNN} ({date}) [{type}] — {what happened}
   ...
   

3. If only one has has_reflection: true, auto-suggest it
4. Let the user select one

Draft the Post

Follow ALL constraints in references/voice.md + references/platform-formats.md LinkedIn conventions:

• Under 200 words, under 1300 characters
• First line must be under 200 characters and contain a specific fact, date, or number
• Opens with specific fact or moment — NOT a question, NOT "I've been thinking"
• Incorporate PVP raw_material — lead with the personal moment
• 2-3 short paragraphs
• Ends with one implication, not a question — no engagement-bait CTAs
• No hashtag lists, no "Here's what I learned", no listicles
• Follow all banned pattern lists from voice.md
• Follow rhythm engineering rules from voice.md
• Follow vocabulary diversification rules from voice.md

After drafting: run Quality Gate (Step 5) before presenting to user.

Write Draft File

Write to marketing/experiment/drafts/linkedin-{YYYY-MM-DD}.md with frontmatter:

---
platform: linkedin
status: draft
created: {YYYY-MM-DD}
published_date: ""
source_observations: [OBS-{NNN}]
pvp_confidence: {high/medium/low}
title: ""
word_count: {N}
char_count: {N}
rhythm_score: {std_dev}
banned_fixes: {N}
---

Update Source Observation

Append to the source observation's entry in marketing/experiment/observations.md:

- **Used in:** linkedin-{YYYY-MM-DD}

If a Used in field already exists, append to the comma-separated list.

Confirm to User

Draft written: marketing/experiment/drafts/linkedin-{YYYY-MM-DD}.md

Words: {N}
Characters: {N}/1300
Personal material: {which PVP elements used}
Rhythm score: {std_dev} (target: >8)
Banned patterns fixed: {N}
Confidence: {high/medium/low}
Source: OBS-{NNN}
Status: draft

Session Log & State Update

1. Append to marketing/logs/{YYYY-MM-DD}.md:

   - **Publisher** — Drafted LinkedIn post from OBS-{NNN}: {word_count} words, {char_count} characters, confidence: {level}
   

2. Update marketing/STATUS.md: set last_updated to today's date, preserve all other fields

Step 3: Medium Reformat

SEO-accessible version of a Substack piece for wider reach. PVP is NOT required — source material comes from the existing Substack piece.

Select Source

1. List available Substack drafts from marketing/experiment/drafts/substack-*.md (read frontmatter of each)
2. Display:

   Available Substack drafts:
   1. substack-{YYYY-MM}.md — "{title}" ({status}, {word_count} words)
   ...
   

3. Auto-suggest if only one exists
4. Let the user select

Reformat

Read the selected Substack draft and reformat:

• New title optimized for search intent (accessible to non-insiders, avoid jargon)
• Add subtitle (Medium supports natively)
• Add 2-sentence intro explaining context for unfamiliar readers
• Restructure sections that assume experiment familiarity
• Add "About this experiment" footnote (1-2 sentences)
• Include canonical URL note for SEO
• Follow references/platform-formats.md Medium conventions
• Target 600-1200 words (shorter than Substack version)

After reformatting: run Quality Gate (Step 5) — reformatting can introduce banned patterns.

Write Draft File

Write to marketing/experiment/drafts/medium-{YYYY-MM}.md with frontmatter:

---
platform: medium
status: draft
created: {YYYY-MM-DD}
published_date: ""
source_observations: [OBS-{NNN}, OBS-{NNN}]
based_on: substack-{YYYY-MM}
canonical_url: ""
title: ""
word_count: {N}
rhythm_score: {std_dev}
banned_fixes: {N}
---

Confirm to User

Medium reformat: marketing/experiment/drafts/medium-{YYYY-MM}.md

Based on: substack-{YYYY-MM}
Title: "{SEO title}"
Words: {N}
Rhythm score: {std_dev} (target: >8)
Banned patterns fixed: {N}
Remember to set canonical_url before publishing on Medium.

Session Log & State Update

1. Append to marketing/logs/{YYYY-MM-DD}.md:

   - **Publisher** — Reformatted substack-{YYYY-MM} for Medium: "{title}" ({word_count} words)
   

2. Update marketing/STATUS.md: set last_updated to today's date, preserve all other fields

Step 4: Content Review

Manage draft status across all platforms. No content generation — Quality Gate does not apply.

List All Drafts

1. Scan marketing/experiment/drafts/*.md, read frontmatter of each file
2. Display table:

| # | File | Platform | Title | Status | Created | Published |
|---|------|----------|-------|--------|---------|-----------|
| 1 | substack-2026-03.md | Substack | "{title}" | draft | 2026-03-10 | — |
| 2 | linkedin-2026-03-12.md | LinkedIn | "{title}" | reviewed | 2026-03-12 | — |

If no drafts exist, tell the user and suggest starting with Step 1 or Step 2.

Available Actions

Present available actions:

• Mark as reviewed — select a draft, update its frontmatter status: reviewed
• Mark as published — select a draft, then:
  1. Update frontmatter: status: published, set published_date to today's date
  2. For each observation ID in source_observations: find the observation entry in marketing/experiment/observations.md and append - **Published:** true (skip if the observation already has this field)

Session Log & State Update

1. Append to marketing/logs/{YYYY-MM-DD}.md:

   - **Publisher** — Content review: marked {filename} as {new_status}
   

2. Update marketing/STATUS.md: set last_updated to today's date, preserve all other fields

Step 5: Quality Gate

Runs internally before presenting ANY draft (Steps 1, 2, 3). This is not a user-facing step — it executes automatically after drafting and before the confirmation output. Fix problems in place; do not present unfixed drafts.

Pass 1: Banned Pattern Scan

Scan the full draft text against every list in voice.md "Banned Patterns" section:

• Hedge words: somewhat, relatively, arguably, it seems, perhaps, might suggest, could potentially, tends to
• Performative enthusiasm: exciting, incredible, game-changing, thrilled, amazing, remarkable, groundbreaking, revolutionary, delighted, proud to share
• Transitions: Therefore, However, Moreover, Furthermore, In conclusion, Notably, Interestingly, Additionally, Consequently, It's worth noting, It should be noted
• Generic filler: in today's rapidly evolving landscape, at the end of the day, moving forward, leveraging, utilize, synergy, paradigm shift, deep dive, unpack
• Structural patterns: bold-colon in body paragraphs, triple parallelism, em dash overuse (>1 per 300 words), neat-bow endings

For each match: rewrite the offending sentence to eliminate the pattern while preserving meaning. Increment banned_fixes counter.

Pass 2: Rhythm Analysis

Calculate sentence lengths (word count per sentence) across the entire draft.

• Compute standard deviation of sentence lengths → this is the rhythm_score
• If std dev < 8: break up monotone runs by splitting long sentences or combining short ones
• Check for 3+ consecutive sentences within 5 words of same length → vary them
• Verify fragment quota: at least 1 fragment per 300 words → add if missing
• Verify long sentence quota: at least 1 sentence >30 words per 500 words → extend one if missing
• Verify at least 1 single-sentence paragraph exists → create one if missing

Pass 3: Paragraph Structure

• Scan for [claim][explanation][example][conclusion] paragraph pattern → restructure offending paragraphs
• Verify at least one paragraph opens with a concrete detail, not a topic sentence → rewrite one opening if needed
• Check final paragraph: must NOT summarize or restate thesis → rewrite ending if it wraps up too neatly

Pass 4: Personal Material Verification

• Count how many PVP elements (specific moment, sensory detail, personal reaction, enterprise parallel) appear in the draft text
• If fewer than 2 PVP elements are present in the actual text: flag to user with a warning, do not silently proceed
• If PVP was skipped: include a note that the draft may read generic

Pass 5: Opening Line Test

• First sentence of the draft must contain at least one of: a date, a proper name, a place, a specific action, or a number
• If the opening is generic or abstract: auto-rewrite it to include a concrete anchor from the source material

Quality Gate Output

After all passes complete, record these metrics for the confirmation output:

• rhythm_score: sentence length standard deviation
• banned_fixes: total banned patterns found and fixed
• confidence: high (PVP 3-4 elements + rhythm >8) / medium (2 elements or rhythm 6-8) / low (PVP skipped or <2 elements)

These metrics appear in the draft frontmatter and the user-facing confirmation.

Step 6: Revision Support

When the user requests changes to an existing draft:

1. Apply the requested edits
2. Re-run the full Quality Gate (Step 5) on the revised text
3. Track what changed: new banned_fixes count, updated rhythm_score
4. Never re-introduce banned patterns while making revisions
5. Present the same confirmation format with updated metrics

Reference Table

Mode	References Read	User Files Read	User Files Written
Step 0	voice.md, experiment-thesis.md, platform-formats.md	MEMORY.md, STATUS.md, observations.md, drafts/*.md	—
Step 0.5	voice.md (PVP section)	—	—
Step 1	—	observations.md, autonomy-log.md	drafts/substack-{YYYY-MM}.md, observations.md, STATUS.md, logs/{today}.md
Step 2	—	observations.md	drafts/linkedin-{YYYY-MM-DD}.md, observations.md, STATUS.md, logs/{today}.md
Step 3	—	drafts/substack-*.md	drafts/medium-{YYYY-MM}.md, STATUS.md, logs/{today}.md
Step 4	—	drafts/*.md, observations.md	drafts/*.md, observations.md, STATUS.md, logs/{today}.md
Step 5	voice.md (banned patterns, rhythm, vocabulary)	draft in progress	draft in progress (fixes applied in place)
Step 6	voice.md (all constraints)	existing draft	revised draft, STATUS.md, logs/{today}.md