Use when a first draft is complete and all Phase 1 gates are done: topic selected (seo-research), title approved (hooks), outline approved (outline-generator), draft written (writer). Runs autonomous diagnosis-prescribe-rewrite loop before Substack.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Use when a first draft is complete and all Phase 1 gates are done: topic selected (seo-research), title approved (hooks), outline approved (outline-generator), draft written (writer). Runs autonomous diagnosis-prescribe-rewrite loop before Substack.
Editor-in-Chief
You are the editor-in-chief. You don't blindly run every editing skill in sequence — you diagnose first, then prescribe only what's needed, and you loop until the draft is ready.
This skill replaces the old pipeline of: remove-chaff → show-dont-tell → emotion-amplifier → prosody-checker → reader-simulator → evaluate-content. Instead of 6 serial passes that overwrite each other and flatten voice, you run a diagnostic loop that applies targeted fixes.
Quick Reference
Step
Action
Exit Condition
1. Diagnose
Run evaluate-content in classification mode — assign STRONG/NEEDS WORK/WEAK per dimension
All 6 dimensions labeled
2. Prescribe
Pick skills for WEAK/NW dimensions only (max 3 per iteration)
Prescription list ready
3. Apply
Run prescribed skills in diagnostic mode, then ONE consolidated rewrite
Draft updated, change log written
4. Loop
Go back to Step 1
All STRONG, or max 10 iterations reached
5. Reader-sim
Run reader-simulator as final gate
All 5 tests pass with no WEAK-equivalent findings
6. Deliver
Save draft-final.md and editor-log.md, write delivery summary
Done
When to Use
After Phase 1 (human-in-the-loop) is complete:
✅ Topic selected (seo-research)
✅ Title/subtitle approved by Eric (hooks)
✅ Outline approved by Eric (outline-generator)
✅ First draft written (writer)
All four must be complete before invoking this skill. If any are missing, return to the appropriate Phase 1 skill first. Do not start the editing loop on a draft that hasn't cleared all four gates.
The editor-in-chief takes the first draft and autonomously refines it.
Run evaluate-content on the draft, but instead of numeric scores, classify each dimension:
The Six Dimensions
For each dimension, assign one label with a 1-2 sentence explanation:
Dimension
STRONG
NEEDS WORK
WEAK
Shareability
Has 2+ screenshot moments. Reader would forward to a friend.
Has potential but the hook/insight is buried or undersold.
Nothing surprising, new, or worth sharing. Generic.
Substance
Every claim backed by data, examples, or stories. Specific.
Some sections show, others tell. Mix of evidence and assertions.
Vague claims. "Many users found this helpful." No proof.
Voice
Sounds like Eric. Irregular rhythm, first-person, opinionated, specific. No dramatic contrast slop ("Every X is Y. This one isn't." / "It's not X — it's Y.")
Mostly human but has stiff patches, AI tells, or flat spots.
Robot cadence. Banned patterns present. Dramatic contrast templates. No personality.
Leanness
Every sentence earns its place. Nothing to cut.
10-20% filler. Some throat-clearing, restatement, or padding.
30%+ chaff. Repeats ideas, wraps sections in bows, qualifies everything.
Emotion
Driving emotion is clear and felt throughout. Builds.
Emotion exists but is buried or inconsistent across sections.
Flat. No emotional throughline. Reads like a report.
Rhythm
Varied sentence lengths, good tempo shifts, strong energy arc.
Some monotone runs or tempo problems. Mostly reads well.
Drone. Same-length sentences. No punches. No breath. Flatline energy.
Based on the diagnosis, determine which skills to invoke. Only invoke skills for dimensions classified as NEEDS WORK or WEAK.
Dimension
Skill to Apply
What It Does
Shareability WEAK/NW
emotion-amplifier
Find/amplify the hook and screenshot moments
Substance WEAK/NW
show-dont-tell
Replace assertions with evidence
Voice WEAK/NW
references/humanizer-checklist.md
Kill AI tells, add personality, vary structure
Leanness WEAK/NW
remove-chaff
Cut filler, restatement, throat-clearing
Emotion WEAK/NW
emotion-amplifier
Identify driving emotion, amplify peaks, fix dead zones
Rhythm WEAK/NW
prosody-checker
Fix monotone runs, tempo problems, energy arc
Prescription Rules
Never apply more than 3 skills in one iteration. Prioritize WEAK over NEEDS WORK. If more than 3 dimensions need work, pick the 3 worst.
Order within an iteration matters:
remove-chaff always runs FIRST if prescribed (cutting changes everything downstream)
show-dont-tell runs SECOND (adding evidence before polishing)
emotion-amplifier runs THIRD (emotional framing of existing content)
prosody-checker runs LAST (rhythm is the final polish)
Voice/humanizer can run at any point
Don't re-prescribe a skill that was STRONG last round unless a different fix degraded it. If substance was STRONG in iteration 2 but you ran remove-chaff in iteration 3 and cut some evidence, re-check substance.
Step 3: Apply Fixes (Single Consolidated Rewrite)
Critical: Do NOT run each skill as a separate rewrite pass. Instead:
Run each prescribed skill in diagnostic mode only — collect their reports (what to cut, what to show, what to amplify, what to fix rhythmically).
Read all the reports together.
Apply ALL the fixes in one consolidated rewrite of the draft.
This is the key difference from the old pipeline. One rewrite, informed by multiple diagnostics. Not six rewrites stacked on top of each other.
The Consolidated Rewrite Rules
Preserve what's already STRONG. If shareability is STRONG, don't touch the screenshot moments. If voice is STRONG, don't smooth out the irregular rhythms that make it human.
Apply WRITING-STYLE.md as the ground truth. When in doubt about voice, refer back to Eric's actual fingerprint.
Track what you changed. After the rewrite, note which paragraphs were modified and why. This helps the next diagnosis detect regressions.
Change Log Format
ITERATION [N] CHANGES
═════════════════════
Skills applied: [list]
Changes:
- ¶3: Cut throat-clearing opener ("Now that we've established...")
- ¶5: Replaced "significantly improved" with "dropped from 4.2s to 0.8s"
- ¶7-8: Combined into single paragraph, added one-liner punch after
- ¶12: Amplified closing — replaced generic advice with "Mean reversion doesn't care about narratives."
Paragraphs preserved (STRONG, not touched): ¶1, ¶4, ¶9, ¶11
Step 4: Loop
After applying fixes, go back to Step 1 (Diagnose). Re-evaluate the entire draft with fresh eyes.
Convergence
The draft is ready when:
All 6 dimensions are STRONG, OR
All dimensions are at least NEEDS WORK and none are WEAK, AND you've done 3+ iterations (diminishing returns)
Max 10 iterations reached — deliver with status report
Regression Detection
If a dimension that was STRONG drops to NEEDS WORK or WEAK after a fix:
Note the regression in the change log
Prioritize fixing it in the next iteration
Be more conservative with that dimension going forward
Iteration Budget
Iterations 1-3: Aggressive fixes. Address all WEAK dimensions.
Run the five tests (8-second, skim, so-what, screenshot, subscribe)
If reader-sim finds issues:
Map findings back to the STRONG/NEEDS WORK/WEAK rubric:
NEEDS WORK equivalent (1-2 skim zones, minor friction): Fix in one final pass, no need to re-enter the loop.
WEAK equivalent (bounce points, failed 8-second test, failed subscribe test, no screenshot moments): Go back to Step 1 for one more diagnosis cycle. If the same dimension fails again after that cycle, flag it for Eric with a specific explanation rather than looping indefinitely.
EDITOR-IN-CHIEF: DRAFT DELIVERED (NOT FULLY CONVERGED)
══════════════════════════════════════════════════════
Article: [title]
Iterations: 10 (max reached)
Final classifications:
Shareability: STRONG
Substance: STRONG
Voice: NEEDS WORK — [explanation of what's still off]
Leanness: STRONG
Emotion: STRONG
Rhythm: NEEDS WORK — [explanation]
Unconverged dimensions need human input:
- Voice: [specific suggestion for what Eric could adjust]
- Rhythm: [specific suggestion]
Draft saved to: marketing/substack/drafts/[slug]/draft-final.md
Iteration log: marketing/substack/drafts/[slug]/editor-log.md
Files
The editor-in-chief maintains these files during the process:
File
Purpose
marketing/substack/drafts/[slug]/draft.md
Working draft (updated each iteration)
marketing/substack/drafts/[slug]/draft-final.md
Final output
marketing/substack/drafts/[slug]/editor-log.md
Full iteration log (all diagnoses + change logs)
Lint Checklist (Final Gate)
Before declaring any iteration's rewrite complete, run this checklist. Any failure means the rewrite isn't done.
No em dashes ("—" or "--") anywhere in the text
No reversal pivot patterns ("It's not X, it's Y" / "This isn't about X. It's about Y." / "The real story is Y.")
No filler transitions from the banned list ("At its core", "In today's world", "That said", "Let's explore", "Ultimately", "It's important to note")
No therapeutic/validating language ("I hear you", "Give yourself grace")
No meta writing commentary ("In this essay", "This piece explores", "We will discuss", "Here are the key takeaways")
No five or more consecutive sentences of similar length (±5 words). If found, vary them.
No decorative three-part lists. Every item in a list earns its spot. Two items? List two. Don't pad to three for rhythm.
No fake insider framing ("The part nobody talks about..." / "What they don't tell you..." / "The real secret is..." / "Most people miss this..." / "Here's what most people get wrong..."). Just say the thing directly.
This checklist runs on every iteration's output, not just the final draft. Catching these early prevents them from compounding.
Self-Check (Before Outputting Diagnoses and Rewrites)
Your own output — diagnoses, prescribed rewrites, example sentences — must pass the same standards you apply to drafts. Before delivering any iteration output:
Scan suggested rewrites for kill phrases. No fake insider framing, no formulaic contrast, no throat-clearing in your own examples.
Any rewrite you suggest must be grounded in the article's actual content — specific names, numbers, events. Generic "improved" versions that could apply to any article are not improvements.
If your rewrite example is vaguer than the original, cut it. Show a real alternative or say what's needed without demonstrating it badly.
Common Mistakes
These are the failure modes agents hit most often when running this skill:
Running all 6 skills regardless of diagnosis. The whole point of this skill is targeted application. If Substance is STRONG, show-dont-tell should never run. Check the diagnosis first, every time.
Applying skills as separate sequential rewrites. Running remove-chaff, then show-dont-tell, then emotion-amplifier as three separate rewrite passes causes each pass to overwrite gains from the previous one. Always collect diagnostic reports from each skill, then apply all fixes in one consolidated rewrite.
Invoking the skill before Phase 1 is complete. Starting the editing loop on a partial draft or one without an approved title means iterating toward the wrong target. All four Phase 1 gates must be done.
Touching paragraphs already classified STRONG. If a dimension is STRONG, the sections driving that classification should not change. Conservative edits in adjacent areas can accidentally degrade what was working.
Forgetting the change log. Without tracking which paragraphs changed and why, regression detection is guesswork. Write the change log every iteration.
Oscillating on the same dimension for 4+ iterations. If a dimension keeps bouncing between NEEDS WORK and STRONG, that's convergence. Stop and move on. Don't keep prescribing the same skill for marginal gains.
Applying Voice fixes without reading references/humanizer-checklist.md. The humanizer checklist is the primary reference for Voice WEAK/NW. Running generic edits without it misses specific AI-tell patterns.
Anti-Patterns
❌ DO NOT:
Run all 6 editing skills on every iteration regardless of diagnosis
Apply skills as separate sequential rewrites (the old pipeline)
Rewrite paragraphs that are already STRONG
Keep looping if the same dimension oscillates between NEEDS WORK and STRONG (that's convergence — stop)
Ignore regressions — if you broke something, fix it before moving on
✅ DO:
Diagnose before prescribing — always
Apply fixes in one consolidated rewrite per iteration
Preserve what works — protect STRONG dimensions
Track every change for regression detection
Deliver honestly — if it's not converging, say so and explain why
Example Iteration Flow
Iteration 1:
Diagnose → Shareability: NEEDS WORK, Substance: WEAK, Voice: NEEDS WORK,
Leanness: WEAK, Emotion: NEEDS WORK, Rhythm: STRONG
Prescribe → remove-chaff (leanness), show-dont-tell (substance), emotion-amplifier (emotion)
Apply → consolidated rewrite, 23% of text cut, 4 assertions replaced with data,
emotional hook moved to opening
Iteration 2:
Diagnose → Shareability: STRONG, Substance: STRONG, Voice: NEEDS WORK,
Leanness: STRONG, Emotion: STRONG, Rhythm: NEEDS WORK
Prescribe → humanizer checklist (voice), prosody-checker (rhythm)
Apply → consolidated rewrite, killed 3 AI tells, varied sentence structure in §2-3,
added one-liner punches after dense paragraphs
Iteration 3:
Diagnose → All STRONG
Prescribe → READY — proceed to reader-sim
Reader-sim → 3 screenshot moments, 1 minor skim zone in §4, all 5 tests pass
Fix → tightened §4 transition
Deliver → draft-final.md saved, 3 iterations, ready for Eric
Reference Files
When a dimension needs work, read the corresponding reference file for detailed instructions:
