forbotsake/forbotsake-content-check

/forbotsake-content-check

The gate between "written" and "published." Catches messaging drift, weak CTAs, and channel misfits before your audience sees them.

Preamble

FORBOTSAKE_HOME="${FORBOTSAKE_HOME:-$HOME/.forbotsake}"
mkdir -p "$FORBOTSAKE_HOME"

# Discover forbotsake install directory
_FBS_ROOT=""
for _FBS_CANDIDATE in "$HOME/.codex/skills/forbotsake" "$HOME/.agents/skills/forbotsake"; do
  [ -d "$_FBS_CANDIDATE" ] && _FBS_ROOT="$_FBS_CANDIDATE" && break
done
if [ -z "$_FBS_ROOT" ]; then
  echo "WARNING: forbotsake not found. Install: bash <(curl -fsSL https://raw.githubusercontent.com/forbotsake/forbotsake/main/bin/install.sh)"
fi

# Check for updates
_UPD=""
[ -n "$_FBS_ROOT" ] && [ -x "$_FBS_ROOT/bin/forbotsake-update-check" ] && _UPD=$("$_FBS_ROOT/bin/forbotsake-update-check" 2>/dev/null || true)
[ -n "$_UPD" ] && echo "$_UPD" || true
_BRANCH=$(git branch --show-current 2>/dev/null || echo "unknown")
echo "BRANCH: $_BRANCH"

# Orchestrated mode (invoked by forbotsake-go, propagated via file flag)
_ORCH_FILE="${FORBOTSAKE_HOME:-$HOME/.forbotsake}/orchestrated-$(basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)")"
FORBOTSAKE_ORCHESTRATED=$(cat "$_ORCH_FILE" 2>/dev/null || echo 0)
echo "ORCHESTRATED: $FORBOTSAKE_ORCHESTRATED"

# Fast mode (skip adversarial gates)
# Read from env var OR file flag (file flag set by forbotsake-go for pipeline propagation)
_FAST_FLAG="$FORBOTSAKE_HOME/fast-$(basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)")"
FORBOTSAKE_FAST="${FORBOTSAKE_FAST:-0}"
[ "$FORBOTSAKE_FAST" = "0" ] && [ -f "$_FAST_FLAG" ] && FORBOTSAKE_FAST=$(cat "$_FAST_FLAG" 2>/dev/null || echo 0)
echo "FAST_MODE: $FORBOTSAKE_FAST"

# Check for strategy.md
if [ -f strategy.md ]; then
  echo "STRATEGY_EXISTS: yes"
  echo "STRATEGY_FILE: strategy.md"
elif [ -f forbotsake-strategy.md ]; then
  echo "STRATEGY_EXISTS: yes"
  echo "STRATEGY_FILE: forbotsake-strategy.md"
else
  echo "STRATEGY_EXISTS: no"
fi

# Check for content directory and list recent files
if [ -d content ]; then
  echo "CONTENT_DIR: exists"
  echo "RECENT_CONTENT:"
  ls -1t content/*.md 2>/dev/null | head -10
  echo "LATEST_FILE: $(ls -1t content/*.md 2>/dev/null | head -1)"
else
  echo "CONTENT_DIR: missing"
fi

If output shows UPGRADE_AVAILABLE <old> <new>: read the forbotsake-upgrade SKILL.md at $_FBS_ROOT/forbotsake-upgrade/SKILL.md (where _FBS_ROOT is the variable already resolved in the preamble bash above) and follow the "Inline upgrade flow" section Step 1 only. If Step 1 results in "Yes" or "Always" (proceed with upgrade), continue through Steps 2-7 of the inline flow. If Step 1 results in "Not now" or "Never" (declined), skip Steps 2-7 entirely and continue with this skill immediately.

If output shows JUST_UPGRADED <old> <new>: tell user "Running forbotsake v{new} (just updated from v{old})!" and continue.

If STRATEGY_EXISTS is no:

"No strategy.md found. I can still review content for general quality, but I can't check it against your brand voice, ICP, or messaging pillars without a strategy.

Want to proceed with a general review, or run /forbotsake-marketing-start first?"

Ask the user directly in conversation with options: A) General review (no strategy checks), B) I'll create a strategy first.

If option B, stop here.

If CONTENT_DIR is missing:

"No content/ directory found. What file do you want me to review? Paste the path or the content directly."

Ask the user directly in conversation.

Phase 1: Select the Content to Review

Orchestrated mode (ORCHESTRATED is 1):

• Auto-select the most recently created content file with status: draft in frontmatter. Review ONE file at a time (the skill's review flow is designed for single-file review). If multiple drafts exist, review only the newest one.
• If no draft files, review the most recently modified content file
• Skip the file selection direct conversation with the user
• In Phase 4: if all dimensions pass, proceed silently. If any fail, auto-apply suggested fixes (equivalent to option A) and re-run. If re-review still fails after 2 iterations, mark as reviewed-override and continue.
• Skip Phase 6 (Next Steps) — the orchestrator handles what's next.

Interactive mode (ORCHESTRATED is 0):

If content files exist, ask via direct conversation with the user:

"Which content piece should I review?

Recent files: {numbered list of files from RECENT_CONTENT, showing filename and frontmatter channel/topic}

Pick a number, paste a file path, or paste raw content directly."

If only one file exists in content/, default to it:

"Found one content file: {filename}. Reviewing that one. Say 'no' if you meant something else."

Read the selected content file completely. Extract from its frontmatter:

• channel -- which platform this is for
• messaging_pillar -- which pillar it supports
• topic -- what it's about
• status -- current status

Phase 2: Load the Review Criteria

Read strategy.md completely. Extract:

1. Positioning statement -- the reference frame for all content
2. ICP -- person, title, pain, language they use
3. Messaging pillars -- the 3 claims, each with proof points
4. Channel strategy -- the channel's score and rationale
5. Brand voice -- if explicitly defined; otherwise infer from the positioning and ICP:
   • Technical ICP = direct, precise, no fluff
   • Business ICP = outcome-focused, professional but not corporate
   • Creative ICP = conversational, personality-driven
   • Developer ICP = show-don't-tell, code examples over claims

Phase 2.5: Content Red Team (Adversarial Review)

Skip this phase if FAST_MODE is 1. Print: "FORBOTSAKE_FAST=1: skipping adversarial review."

This is an independent adversarial review of the content by a fresh-context subagent. The reviewer has NOT seen your conversation or the scorecard dimensions. It sees only the content draft, strategy.md, and the banned patterns list.

Step 1: Load banned patterns

FORBOTSAKE_HOME="${FORBOTSAKE_HOME:-$HOME/.forbotsake}"
_SKILL_DIR=$(dirname "$(find $HOME/.codex/skills -path "*/forbotsake-marketing-start/SKILL.md" -type f 2>/dev/null | head -1)" 2>/dev/null)
_FBS_ROOT=$(cd "${_SKILL_DIR}/.." 2>/dev/null && pwd || true)

# Load default patterns
_DEFAULTS=""
[ -n "$_FBS_ROOT" ] && [ -f "$_FBS_ROOT/knowledge/banned-patterns-defaults.md" ] && _DEFAULTS="$_FBS_ROOT/knowledge/banned-patterns-defaults.md"
echo "BANNED_DEFAULTS: ${_DEFAULTS:-not_found}"

# Load user patterns
_USER_PATTERNS="$FORBOTSAKE_HOME/banned-patterns.md"
[ -f "$_USER_PATTERNS" ] && echo "USER_PATTERNS: $_USER_PATTERNS" || echo "USER_PATTERNS: none"

Step 2: Dispatch the reviewer

Run the review inline to spawn an independent reviewer subagent. The subagent prompt must contain ONLY:

• The content file path to review
• The strategy.md file path (for alignment checking)
• The banned patterns file path(s)
• The 6 review dimensions
• The JSON output contract

Subagent prompt:

You are a content red team reviewer. Your job is to REJECT bad content, not approve good content.

Read these files:

1. {content file path} -- the content to review
2. {strategy.md path} -- the strategy it should align with
3. {banned patterns path(s)} -- patterns that signal AI-generated content

Review the content on 6 dimensions:

1. AI-SLOP DETECTION: Check for patterns from the banned patterns file. Count matches.
2. VOICE AUTHENTICITY: Does this sound like a human wrote it, or like an AI? Check against brand voice in strategy.md.
3. ICP SPECIFICITY: Would the target person from strategy.md actually engage with this, or scroll past?
4. ORIGINALITY: Does this say something the ICP hasn't heard before, or is it generic advice anyone could write?
5. STRATEGY ALIGNMENT: Does the content execute the positioning and messaging pillars from strategy.md?
6. PUBLIC REPUTATION RISK: Would the founder be embarrassed if this went viral for the wrong reasons?

Respond with ONLY valid JSON, no markdown fences, no explanation outside the JSON: {"result": "PASS" | "SOFT_FAIL" | "HARD_FAIL", "findings": [{"dimension": "...", "verdict": "PASS" | "FAIL", "spans": ["quoted text from content"], "fix": "suggested replacement"}], "summary": "one-line overall assessment"}

Step 3: Parse and act on the result

Parse the subagent's response as JSON.

Before parsing: Strip markdown fences if present (remove leading json and trailing lines). Then attempt JSON parse.

If JSON parsing still fails after stripping fences:

• Treat as PASS with warning: "Adversarial reviewer returned unparseable output. Proceeding to scorecard review (Phase 3 still checks quality)."
• Log to metrics: {"gate":"content","result":"PARSE_FAIL"}
• Continue to Phase 3 (the existing scorecard still runs as a backstop).

If parsed successfully, act on the result:

PASS: Print "Content Red Team: PASS. Proceeding to scorecard review." Continue to Phase 3.

SOFT_FAIL: Present the findings to the user:

"Content Red Team: SOFT_FAIL

An independent reviewer flagged these issues: {for each finding with verdict FAIL:}

• {dimension}: "{spans}" ... suggested fix: "{fix}"

This is iteration {N} of 2."

• Interactive mode: Ask the user directly in conversation: A) Apply fixes and re-review B) Override and proceed to scorecard C) I'll revise manually

• Orchestrated mode: Auto-apply fixes, re-dispatch reviewer. Max 2 iterations total (hard counter). After 2 iterations, if still SOFT_FAIL, proceed to Phase 3 with concerns logged.

If the user chooses A: apply fixes using Edit tool, re-dispatch the reviewer (Step 2). Max 2 total iterations. After 2, proceed to Phase 3 regardless.

If the user chooses B: update content frontmatter with reviewer_notes: containing the findings. Log override. Continue to Phase 3.

HARD_FAIL: The content needs a fundamental rewrite.

"Content Red Team: HARD_FAIL

The independent reviewer says this content needs a fundamental rewrite: {summary}

Specific issues: {for each finding with verdict FAIL:}

• {dimension}: "{spans}"

Next action: Run /forbotsake-create again. The reviewer's notes will be available as context to avoid the same issues."

• Update content frontmatter: add status: hard-failed and reviewer_notes: {JSON findings}
• Interactive mode: Ask the user directly in conversation: A) I'll rewrite with /forbotsake-create B) Override and proceed anyway (I disagree with the reviewer)
• Orchestrated mode: Write status: hard-failed to frontmatter. Present user with explicit proceed/abort choice (this is the one gate that always requires user input, even in orchestrated mode).

If override: set status: reviewed-override, log override reason, continue to Phase 3.

Step 4: Log metrics

After the red team phase completes, log the result:

FORBOTSAKE_HOME="${FORBOTSAKE_HOME:-$HOME/.forbotsake}"
mkdir -p "$FORBOTSAKE_HOME"
echo '{"gate":"content","ts":"'"$(date -u +%Y-%m-%dT%H:%M:%SZ)"'","result":"RESULT","content_file":"FILENAME","iterations":N,"override":BOOL}' >> "$FORBOTSAKE_HOME/review-metrics.jsonl" 2>/dev/null || true

Replace RESULT, FILENAME, N, BOOL with actual values.

---

Phase 3: Run the Review

Evaluate the content against these 7 dimensions. For each, give a PASS or FAIL with specific, actionable feedback. Not vague ("could be better") -- specific ("tweet 3 uses 'leverage' which your ICP wouldn't say; replace with 'use'").

Dimension 1: Brand Voice Consistency

Does the content sound like it comes from the same person/brand as the strategy?

Check for:

• Tone match (formal vs casual, technical vs accessible)
• Vocabulary alignment (does it use the ICP's words, not marketing jargon?)
• Personality consistency (if strategy says "direct and no-BS" but content is hedging and passive)
• First-person vs third-person consistency

PASS criteria: A reader familiar with the brand would recognize this as on-brand. FAIL example: "Strategy says 'direct, technical, no fluff' but this post opens with 'In today's fast-paced world of innovation...'"

Dimension 2: Messaging Alignment

Does the content reinforce at least one messaging pillar without contradicting the others?

Check for:

• Clear connection to a specific pillar
• Claims that are supported by the pillar's proof points
• No contradictions with positioning statement
• No accidental repositioning (claiming to be something the strategy doesn't support)

PASS criteria: You can point to the specific pillar this content supports. FAIL example: "Strategy positions you as the 'simple' alternative, but this post emphasizes 'powerful enterprise features.'"

Dimension 3: ICP Targeting

Is this content written FOR the ICP, not for a generic audience?

Check for:

• Does it reference the ICP's specific pain point?
• Would the ICP recognize their situation in this content?
• Is it using language the ICP uses (not your internal terminology)?
• Is the assumed knowledge level correct for the ICP?

PASS criteria: The ICP from strategy.md would feel "this was written for me." FAIL example: "ICP is 'backend engineers at Series A startups' but this post reads like it's targeting CTOs at enterprises."

Dimension 4: Channel Format Fit

Does the content respect the platform's conventions and constraints?

Check per channel:

• X/Twitter thread: Each tweet under 280 chars? Hook in tweet 1? Tweets stand alone? Thread length 3-10 tweets?
• X/Twitter single: Under 280 chars? Punchy? One clear point?
• Blog: Scannable headers? 1500+ words for SEO? Clear structure? Meta description?
• LinkedIn: First 2 lines hook before fold? Under 1300 chars? No hashtag spam? Line breaks?
• Email: Compelling subject line? Short paragraphs? Single CTA? Mobile-friendly length?
• Reddit: Value-first? No overt self-promotion? Matches subreddit norms?
• Hacker News: Zero marketing speak? Technical substance? Genuine?

PASS criteria: The content would not look out of place on the target platform. FAIL example: "This X thread has 15 tweets -- that's too long. Most readers drop off after tweet 5-7."

Dimension 5: CTA Clarity

Is there one clear call-to-action, and does it match the content's funnel position?

Check for:

• Single CTA (not competing asks)
• CTA matches funnel stage: awareness content = soft CTA (follow, bookmark), consideration = medium CTA (try, sign up for beta), decision = hard CTA (buy, subscribe)
• CTA is specific, not vague ("Try the free tier at {url}" not "check it out")
• CTA placement (end of content, not buried in the middle)

PASS criteria: A reader knows exactly what to do next and why. FAIL example: "CTA says 'sign up now' but this is an awareness-stage thread. Try 'follow for more' or 'bookmark this.'"

Dimension 6: Length Appropriateness

Is the content the right length for its channel and purpose?

Guidelines:

• X thread: 3-10 tweets (sweet spot: 5-7)
• X single tweet: 100-280 chars (sweet spot: 200-240)
• Blog post: 1500-2500 words for SEO, 800-1200 for thought leadership
• LinkedIn post: 500-1300 chars (sweet spot: 800-1000)
• Email: 200-500 words for newsletters, 50-150 for transactional
• Reddit: varies by subreddit, but substance over brevity

PASS criteria: Length serves the content. Nothing feels padded or cut short. FAIL example: "This blog post is 600 words. For the SEO goal noted in your strategy, aim for 1500+."

Dimension 7: Hook Strength

Would the first line/tweet/sentence make the ICP stop scrolling?

Check for:

• Opens with a specific, surprising, or relatable claim (not generic)
• Creates curiosity or recognition ("that's me" feeling)
• Avoids cliches ("In today's world...", "Have you ever...", "Let me tell you about...")
• Front-loads value -- doesn't bury the lead

PASS criteria: Reading only the first line, the ICP would want to read the second. FAIL example: "Opens with 'Excited to announce...' -- nobody stops scrolling for your excitement. Lead with the benefit."

Dimension 8: Visual Consistency (schema_version 2 only)

If the content file has schema_version: 2 and visual_treatment is not none, check:

1. Visual treatment appropriateness: Does the chosen treatment (text-card/ai-image/video) fit the channel and content type? Flag mismatches:

   • Product Hunt launch with visual_treatment: none → "Product Hunt posts need gallery images. Recommend ai-image."
   • LinkedIn post with visual_treatment: none → "LinkedIn posts with images get 2x engagement. Consider text-card or ai-image."
   • HN post with visual_treatment: ai-image → "Hacker News is text-focused. Visual adds no value here."

2. Visual prompt quality: Does visual_prompt reference brand elements from brand.md?

   • Check if prompt includes brand colors or style descriptors
   • Check if prompt relates to the content's actual topic (not generic)

3. Visual asset existence: If visual_status is generated, check that the visual file actually exists at the expected path.

4. Accessibility: Does visual_alt provide a meaningful description? Not just "image" or empty — it should describe what's depicted for screen readers.

5. If visual file exists: Use the Read tool to display the image to the user. Ask: "Does this image match your brand and content? Quick visual sanity check."

PASS criteria: Visual treatment fits the channel, prompt references brand, asset exists, alt text is meaningful. FAIL example: "visual_treatment is ai-image but visual_status is 'failed'. No image file exists. Either regenerate or switch to text-card."

Phase 4: Deliver the Verdict

Present the review as a scorecard:

## Content Review: {filename}

| Dimension | Verdict | Notes |
|-----------|---------|-------|
| Brand Voice | PASS/FAIL | {one-line summary} |
| Messaging Alignment | PASS/FAIL | {one-line summary} |
| ICP Targeting | PASS/FAIL | {one-line summary} |
| Channel Format | PASS/FAIL | {one-line summary} |
| CTA Clarity | PASS/FAIL | {one-line summary} |
| Length | PASS/FAIL | {one-line summary} |
| Hook Strength | PASS/FAIL | {one-line summary} |
| Visual Consistency | PASS/FAIL/N/A | {one-line summary} |

**Overall: READY TO PUBLISH / NEEDS REVISION**

If all 7 dimensions pass:

"READY TO PUBLISH. This content is on-brand, on-strategy, and channel-appropriate.

Quick sanity checks before hitting post:

• Read it out loud one more time
• Verify all links work
• Double-check any specific numbers or claims

When ready, use /forbotsake-publish to ship it."

If any dimension fails:

Present the failures with specific edit suggestions. Not "make the hook better" -- provide an actual rewritten alternative.

"NEEDS REVISION. {N} of 7 dimensions failed. Here are the specific fixes:

{Failed Dimension 1}: Current: "{quote the problematic part}" Problem: {why it fails} Suggested fix: "{rewritten version}"

{Failed Dimension 2}: Current: "{quote the problematic part}" Problem: {why it fails} Suggested fix: "{rewritten version}"

Want me to apply these fixes? Or do you want to revise manually?"

Ask the user directly in conversation with options: A) Apply all suggested fixes B) Apply some fixes (tell me which) C) I'll revise manually D) Override -- publish as-is (I disagree with the review)

If A or B: apply the edits to the content file using the Edit tool, then re-run the review on the updated version. Repeat until all dimensions pass or the user overrides.

If D: update the content file's frontmatter status to reviewed-override and proceed.

Phase 5: Update Content Status

After the review is complete (pass or override), update the content file's frontmatter:

# Update status in frontmatter

Use the Edit tool to change status: draft to:

• status: reviewed if all dimensions passed
• status: reviewed-override if the user chose to override failures
• status: revised if edits were applied and the re-review passed

Phase 6: Next Steps

Orchestrated mode (ORCHESTRATED is 1): Skip this phase entirely. Confirm: "Review complete for content/{filename}. Status: {reviewed/revised/reviewed-override}." Then stop.

Interactive mode (ORCHESTRATED is 0): Tell the user:

"Review complete for content/{filename}.

{If READY TO PUBLISH:} Next step: Use /forbotsake-publish to ship it.

{If NEEDS REVISION and user chose to revise manually:} Next step: Edit the file, then run /forbotsake-content-check again.

{If user wants to create more content:} Create more: /forbotsake-create to write your next piece from the calendar.

{If no content calendar exists:} Get organized: /forbotsake-content-plan to build a content calendar so you always know what to write next."