Atomize

@../_shared/storytelling-frameworks.md Read when: selecting a narrative framework in Step 5 (or by the Step 5 subagent, per-format) @../_shared/persuasion-principles.md Read when: selecting persuasion principles in Step 5 (or by the Step 5 subagent, per-format)

Distribution Engine (Atomization) Skill

Takes one core message — either a brief.md campaign handoff or a manually described idea — and produces properly-adapted drafts for every named format in a single run, keeping the core claim and proof points identical across all of them while letting structure, length, and tone vary per format's own conventions.

Key principle (from research): fix the facts, vary the frame. The specific claim, statistics, named proof points, product/entity names, and quoted figures must stay byte-identical across every atomized piece; structure, length, hook, opening, tone, CTA, and formatting must be rewritten per platform to match its native conventions. This skill enforces that separation to prevent "copy-paste repurposing" — a truncated blog dumped onto LinkedIn — which audiences and algorithms both punish.

Step 0: Recall learnings

If .claude/learnings.md exists, read it silently, [cc-content:*] tags plus any cross-plugin entries relevant to distribution/repurposing. Never announce.

Step 1: Load context

Read the context table from all loaded CLAUDE.md files:

grep -A 200 '## Context files' CLAUDE.md 2>/dev/null || echo "(no context table)"

CLAUDE.md files may exist at multiple hierarchy levels (workspace root, project root, sub-directory). The harness already loads all applicable ones into your context window. If multiple ## Context files tables exist, rows from more specific CLAUDE.md files take precedence over less specific ones.

If no context table is found in any loaded CLAUDE.md, ask once:

"I don't see any context files registered. Would you like to: (a) Pause and run /content-onboarding to set up context (b) Continue without project context (output will be generic)"

Stop if (a); note "generating without project context" and continue if (b).

If a context table exists, read every file listed in the File column.

After loading, assess what each file covers by reading its Summary entry. Map the loaded files to these content needs:

When multiple files plausibly cover the same need, pick the one whose Summary best fits this specific task. For example, if one brand-voice file is "casual — employer branding" and another is "formal — corporate communications", and this is for LinkedIn promotion, load the more casual one and note the choice.

Coverage gaps — flag these two:

If no loaded file plausibly covers brand voice, ask once:

"I don't see any writing style or brand voice context. Is this intentional, or should I pause while you run /content-onboarding?"

• Intentional: note the gap; label output ⚠ DEGRADED OUTPUT — no brand voice context
• Pause: direct the owner to onboarding and stop.

Apply the same ask for organization background if no loaded file covers it.

For absent audience or language: note silently and continue.

Step 2: Resolve the format list and the core message source

Check for brief.md in the working directory first:

ls brief.md 2>/dev/null && echo "found" || echo "missing"

• Found: read it. If it names a channel mix (as campaign-concept's and gtm-plan's output does), propose that list as the target formats and ask the owner to confirm or adjust before continuing.
• Missing: ask: "Which formats do you want this in? (e.g. LinkedIn post, newsletter, blog post) I'll draft each from the same core message." Then ask for the core message/idea directly, same prompt long-tail-copy uses today.

Step 3: Get the content idea, audience, and length target

Ask for the content idea, audience, and any core proof points once, not once per format. The core message, key value proposition, and proof points must stay identical across every format produced in this run — only structure, length, and tone vary per format's own conventions (handled in Steps 4–5 below: Step 4 routes each format, Step 5 drafts each Tier-B/C format — normally in its own isolated subagent, in parallel, or inline when only one format needs drafting).

If the owner has not said what the text should be about, ask:

"What should this text say? Describe the content idea or key message, or paste raw notes and I'll shape them into the finished ."

Then settle anything still open that materially shapes the draft:

• Output language: write the finished text in the language specified by loaded context — even when an inspiration article is in another language (an English source does not make a German-default project's text English). Only if no output language is in context, fall back to the language of the owner's request.
• Audience: if the loaded audience context has more than one persona, infer which one(s) this text targets and confirm in one line. If audience context is absent and the owner hasn't said, ask only if it would change the draft meaningfully.
• Core proof points & constraints: if the owner names specific facts, statistics, or claims that must appear in every format, note them explicitly — this is the "substance layer" that stays byte-identical across all formats per the research.

Step 4–5: Route each format, then draft in parallel

For each format in the confirmed format list, first resolve routing (Step 4). Formats that land in Tier B or Tier C then get drafted concurrently, one isolated subagent per format (Step 5) — not one after another in this session's context.

Step 4: Identify the format and route correctly (per format)

First, settle on the requested output format for this iteration. (If the owner has not named each clearly, they should have in Step 2's format list.)

Once you know the format, work through three tiers in order:

Tier A — A dedicated skill exists for this format → route away

Some formats have their own dedicated skill that applies format-specific best practices this skill does not. Using this skill for them produces weaker results.

How to check: look at the skills available in this session (your available-skills list). The cc-content plugin ships dedicated skills for some formats — for example blog articles (blog-article) and LinkedIn posts (linkedin-post) — and the project may have added its own dedicated skills for other formats. Do not rely on a hardcoded list; judge from the skills actually available whether any one's purpose squarely covers the requested format.

If a dedicated skill covers the format:

Tell the owner: "The format has a dedicated skill — /<skill-name> — which applies format-specific best practices. I'll route this format there and draft the remaining formats here. (You can always run /<skill-name> standalone later if you want to adjust the version independently.)"

Then stop drafting this format here; it's handled by the dedicated skill. Note which formats are routed away for the final summary in Step 6.

Tier B — No dedicated skill, but a format-guideline file is registered → flag it

From the context files you loaded in Step 1, check whether any file's Summary covers best practices, structure, or length for the requested format (e.g. a whitepaper guideline, e-mail best-practices, newsletter guideline, webinar guideline). If one exists, treat it as authoritative for this format and note its path — the Step 5 subagent for this format will read it directly, so it doesn't need to be loaded here. Note: "Format rules: <file>."

Tier C — Neither → use your own best practices

If no dedicated skill and no registered guideline cover the format, note: "No project guideline for ; using general best practices." The Step 5 subagent for this format will draft using its own knowledge of current best practices for that format's length, structure, and conventions.

Step 5: Dispatch drafting subagents (parallel, per format)

Every format assigned Tier B or Tier C in Step 4 needs a full pass — framework selection, persuasion-principle selection, draft, self-edit — but none of those formats depend on each other's output; only the fixed core message from Step 3 is shared. Rather than running that pass sequentially for each format in this session's context, dispatch one subagent per Tier-B/C format, all in parallel, and let each work in its own isolated context loaded with only what its format needs.

If zero formats landed in Tier B/C, skip Step 5 entirely — there's nothing to draft, and Step 6 will just present the routing summary.

If exactly one format landed in Tier B/C, skip dispatch and do the pass directly in this session instead. If it's Tier B, Read the guideline file noted in Step 4 first — that load was deliberately deferred out of Step 4 on the assumption a Step 5 subagent would do it, so the inline path has to do it explicitly here instead. Then apply the drafting task below (framework selection, persuasion selection, draft pass 1, self-edit pass 2 — plus the channel-formatting rule if this format strips formatting) to yourself, the same as it's written for a subagent prompt, rather than dispatching one — parallelization overhead isn't worth it for a single format. Since there's no subagent to supply the batch-learnings note the return contract below asks for, jot down that same one-line observation (a correction you had to make, or how well this format took the core message — or nothing, if there's nothing notable) yourself for Step 7 to pick up.

Otherwise, call the Agent tool once per Tier-B/C format, all in a single message (parallel tool calls — see the Agent tool's guidance on this). Fan-out scales with the number of Tier-B/C formats from Step 4, not the raw format count from Step 2 (formats routed to Tier A never reach this dispatch) — fine for the handful of formats a typical campaign leaves in Tier B/C. If that count is large (as a rough guide, more than ~6), this is a heads-up, not a stop-and-wait: mention the fan-out size to the owner, then proceed with dispatching — don't block on a reply. The formats were already confirmed in Step 2, so this is informational (they can say "fewer formats" for next time) rather than a gate on this run. Each dispatch call gets:

• subagent_type: "general-purpose"
• run_in_background: false — Step 6 needs every result before it can present anything, so there's nothing to gain from backgrounding these.
• description: e.g. "Draft <format> atomization"
• prompt: a self-contained brief (the subagent has no access to this conversation) containing:
  1. The content idea / core message exactly as gathered in Step 3.
  2. The substance layer, verbatim: the specific claim(s), statistics, named proof points, product/entity names, and quoted figures that must appear byte-identical in the output. State explicitly that these must be re-framed, never altered.
  3. The confirmed output language and audience persona(s) from Step 3.
  4. The single target format this subagent is drafting — nothing else.
  5. File paths for the subagent to Read itself:
     • the brand-voice and organization-background context file(s) identified in Step 1 (project-relative, same as the context table).
     • the two shared reference files — but resolve ${CLAUDE_SKILL_DIR} to an absolute path yourself first (e.g. run echo "$CLAUDE_SKILL_DIR") and embed that resolved absolute path in the prompt, as <resolved-path>/../_shared/storytelling-frameworks.md and <resolved-path>/../_shared/persuasion-principles.md. A dispatched subagent has no CLAUDE_SKILL_DIR of its own to expand, and a hardcoded plugins/cc-content/skills/_shared/... guess only resolves inside this dev checkout, not a normal plugin install.
     • for Tier B, the format-guideline file path noted in Step 4. For Tier C, state plainly that no guideline file exists and it should use its own best-practices knowledge for this format's length, structure, and conventions.
  6. An explicit tool-scope constraint:
     • state plainly that the subagent's job is read-and-draft only — it must Read only the exact file paths listed above, and must not use any other tool (no web access, no fetching URLs, no shell commands, no writing files).
     • this skill's own allowed-tools (Read, Write, Bash, Agent) has no web access at all, and its Bash use is limited to the narrow ls/grep checks in Steps 1–2. A dispatched subagent needs neither shell nor web access to do its drafting job, so scope it tighter than even this skill's own allowance, not merely no wider than it.
     • a general-purpose subagent defaults to a much broader tool surface than that; since the content idea and any brief.md this prompt draws on may contain untrusted or externally-sourced text, don't let a subagent's effective tool access exceed what its actual job requires.
     • this is a best-effort mitigation, not a hard boundary: the Agent tool has no parameter to actually restrict which tools a dispatched subagent can call, so a sufficiently effective prompt injection in the source content could still get a subagent to disregard this instruction. Don't treat this line as a guarantee when auditing the skill's security posture — if the source content is untrusted enough that this matters, that's a reason to review it before running /atomize, not something this instruction can enforce on its own.
  7. The full task, in order:
     • Select a storytelling framework per storytelling-frameworks.md's selection process; apply it as the structural spine where it fits; let SPIN (Situation → Problem → Implication → Need-Payoff) shape the argument's progression where that fits too.
     • Select 1–3 persuasion principles per persuasion-principles.md's selection process, plus a pre-suasive opener strategy; layer them into the prose, not as labeled callouts.
     • Draft pass 1: write the text in the output language as a creative, expressive finished piece for this format, holding to: keep the substance layer fixed and adapt everything else (hook, structure, length, tone, CTA, formatting) to this platform's native conventions; connect to the organization subtly (no overt self-promotion); offer original value beyond the core message's bare facts, tailored to this format's audience; use the loaded brand voice; honor the Tier-B guideline or general best practices for this format's length/structure/conventions; aim at the confirmed persona(s); and if this format is a social-media post or any channel that strips formatting, no bold/italics/Markdown — replace bullet points with fitting emojis, keep the body plain.
     • Self-edit pass 2: read the pass-1 draft sentence by sentence and fix, with minimal intervention: factual errors or ambiguous phrasing (cut rather than heavily rewrite if a fix would require it); unsupported claims about the organization not traceable to the loaded context (soften, rephrase, or remove); hype and over-dramatization; style drift from brand voice or the format guideline (including in sentences added while editing); em-dash-set-off parenthetical insertions ("–"/"—"), rewritten as clean sentences — a frequent AI tell; and a final check that the substance layer is still present and unchanged (abbreviating a statistic to fit platform constraints is fine as long as the number and meaning survive). If a Tier-B guideline specifies an optimal length, use the full range rather than coming in thin.
  8. The exact return contract — the subagent should return only:
     • the finished text (nothing else mixed into it).
     • the storytelling framework it used.
     • the persuasion principles it used, plus its opener strategy.
     • the format-rules source (guideline path or "general best practices").
     • the length (count + unit) and, if the Tier-B guideline specifies an optimal length range, that range too (else "n/a" — the main session has no other way to know it, since it never reads the guideline itself on this path).
     • one line for the batch learnings step — a correction it had to make, or an observation about how well this format took the core message — or "none".

Wait for every dispatched subagent to finish before continuing to Step 6. Treat either of the following as a failed draft, and re-dispatch that one format once:

• the result is missing the substance-layer facts, or contradicts them — tell the retry prompt specifically which fact was dropped or altered last time, so it doesn't reproduce the same mistake; or
• the dispatch returned no usable result at all (the subagent errored out or returned nothing) — retry with the same prompt, no diagnosis needed since there's no prior draft to compare against.

Re-dispatching (not drafting inline) is the retry for both cases — it keeps the isolation benefit and is what "redo once" means here. The only time to skip straight to drafting inline, without a retry, is if dispatch itself looks broken across the whole batch (e.g. every subagent in this run is erroring out, not just one) — that's a sign the retry won't fare any better. If a single-format retry still fails, stop retrying: draft that one format directly yourself, following the single-format fallback above (including the guideline-file read and the channel-formatting rule). Add the optional Note: footer line described in Step 6 to flag that it needed a manual fix.

Step 6: Present all drafted formats

Present each drafted format (i.e. every format that was not routed to a dedicated skill in Step 4) in its own delimited block using the format below, in sequence, using the text and metadata each Step 5 subagent returned, or that you drafted directly for any format that went through one of Step 5's inline paths (the single-format fallback, or the exhausted-retry manual fix within an otherwise subagent-dispatched batch). After all blocks, add a one-line routing summary.

Present the finished text in a clearly delimited block so the owner can copy it cleanly:

─────────────────────────────────────────────
<format> draft
─────────────────────────────────────────────
<the finished text only — no preamble, no commentary inside the block>
─────────────────────────────────────────────
Format rules: <Tier-B guideline path | "general best practices">
Framework: <name> · Persuasion: <list>
Length: <count> <unit> (target: <range or "n/a">)
Note: <manual fix after a failed subagent retry — omit this line otherwise>
─────────────────────────────────────────────

The block must contain only the deliverable text — no introduction, no explanation. Put any notes in the footer rows, not inside the text. The Note: row is optional — include it only for a format that went through Step 5's exhausted-retry path.

Channel formatting is drafted in, not applied here: Step 5's draft-pass-1 instructions are the source of truth for the no-bold/no-italics/emoji-bullets rule on formats that strip formatting — that's where it actually gets applied, on every path (subagent-dispatched, single-format fallback, and manual-fix fallback alike). Before presenting, spot-check that the rule was actually followed for any such format; if it wasn't, fix it here rather than re-deriving the rule from scratch.

If the output is degraded (brand voice or organization context missing), prepend:

⚠ DEGRADED OUTPUT — generated without: <list of missing context>

Routing summary: after all format blocks, add one line listing every format and where it ended up, e.g.: "Routed: LinkedIn → /linkedin-post; Drafted here: newsletter, Twitter/X thread."

Step 7: Feedback

Run this feedback step once for the whole batch of drafted formats, not once per format — the same tag and qualification rules apply to the batch as a whole.

Auto-store phase. Before asking for feedback, review this run, including the one-line notes each Step 5 subagent returned for the batch learnings step. For each qualifying observation, append one tagged line to .claude/learnings.md (create with the standard header if missing):

[cc-content:atomize] <concise observation> — <YYYY-MM-DD>

Qualifies: content preferences or constraints not already in any loaded context file or CLAUDE.md; corrections the owner made to the output; project-specific facts that would change future output (e.g. "repurposing always targets these exact five formats"); accepted/rejected deviations from best practices; observations about which formats atomize well from the core message (relates to research findings on reliable vs. unreliable format pairs).

Does not qualify: standard behavior applied without deviation; facts already in context files or CLAUDE.md; anything derivable by re-reading context files; facts semantically equivalent to an existing .claude/learnings.md entry under any plugin tag — when in doubt, skip; redundancy is worse than a missed entry.

Check for the file before appending:

ls .claude/learnings.md 2>/dev/null && echo "exists" || echo "missing"

Standard header when creating the file:

# Learnings

Corrections and feedback collected during content sessions.
Entries are tagged by skill and dated.

---

Explicit feedback. After the auto-store phase, ask:

"Did this batch of formats meet expectations? Any corrections or notes for future atomization runs — or press Enter to finish."

• If the owner provides a correction: append it as a tagged entry using the same format and qualification criteria above. Confirm: "✓ N learning(s) saved to .claude/learnings.md."
• If the owner confirms or skips: if any entries were auto-stored, confirm "✓ N learning(s) auto-saved to .claude/learnings.md." Then exit. If nothing was stored, exit directly.