profile get
-> content history
-> content topic mine
-> select topic locally
-> write longform plan locally
-> agent plan-review
-> luma-find-material plan and collect
-> build local writing brief
-> write script locally
-> agent script-review
-> revise until accepted
-> content artifact upload

research.run is not a substitute for content topic mine. It may only suggest additional search language. Do not call backend script.write; this skill requires agent-authored local writing.

Files

Create these files in the active project:

01_profile.json
01_profile_extra.md
02_content_history.json
02_raw_signals.json
03_topic_selection.json
03a_longform_plan_<topic_id>.json
03b_plan_review_<topic_id>.json
04_material_plan_<topic_id>.json
05_material_assets_<topic_id>.json
materials/<topic_id>/final_assets/deliverables_manifest.json
06_script_writer_payload_<topic_id>.json
07_script_<topic_id>.json
08_script_review_payload_<topic_id>.json
09_script_review_<topic_id>.json

Never overwrite an accepted artifact while revising. Add _v2, _v3, and so on, then promote the accepted version to the stable filename.

Checkpoint 0: Prepare

Confirm authentication and active project:

luma-cli auth status
luma-cli project artifact list

If no project is active:

luma-cli project create <project_name>
luma-cli project use <project_name>

Resolve the profile:
```
luma-cli --json profile current
```
If no usable profile exists, stop this SOP and run luma-profile-onboarding.

Pass condition: one active project and one explicit <profile_id>.

Checkpoint 1: Save Profile Snapshot

Run:

luma-cli --json profile get <profile_id>

Save the JSON response data object, without the CLI envelope, as 01_profile.json.
Save data.extra as 01_profile_extra.md. Keep the remaining profile fields in 01_profile.json.
Verify that identity, audience, and stance are non-empty. If any is empty, update the profile before continuing.

Pass condition: the agent can state who is speaking, to whom, and what judgments this creator owns.

Checkpoint 2: Inspect Used Topics

Run and save the JSON result as 02_content_history.json:

luma-cli --json content history --profile <profile_id>

Build a used-topic set from:
- meta.topic_id
- meta.topic_title
- meta.content_fingerprint
- downloadable script, script-review, topic-selection, and longform-plan artifacts
Treat equivalent thesis and angle combinations as duplicates even when titles differ.

Pass condition: write a short local list of excluded topic IDs, titles, and near-duplicate angles. If history cannot be loaded, record that limitation and continue with stricter freshness checks.

Checkpoint 3: Build Search Matrix

Derive non-duplicate searches from the profile and used-topic set:

Dimension	Required
category terms	2
audience pains or objections	2
workflow or use-case terms	2
decision, misconception, or contrarian terms	1-2
current product, policy, or market terms	1-2 when relevant

Produce:

6-10 concise Douyin/social keywords
3-5 natural-language web queries

Do not use one broad phrase for every query. Do not search only the topic the agent already wants to write.

Optional recovery only:

luma-cli research run \
  --role "<profile identity, audience, and domain>" \
  --mode expanded \
  --date-range 7d \
  --output keyword_expansion.json

Use useful phrases from that result in the matrix, then continue to content topic mine.

Pass condition: the matrix covers at least three distinct intent dimensions.

Checkpoint 4: Mine Canonical Raw Signals

Run exactly this capability:

luma-cli content topic mine \
  --profile <profile_id> \
  --social-keywords "<6-10 comma-separated keywords>" \
  --web-queries "<3-5 comma-separated queries>" \
  --date-range 7d \
  --limit-per-keyword 20 \
  --web-num 6 \
  --max-raw 200 \
  --output 02_raw_signals.json

Validate:

node <skill_dir>/scripts/validate_artifact.js \
  --type discovery \
  --input 02_raw_signals.json

Required:

raw_signals contains at least 15 unique signals; target 30 or more
counts.social_raw > 0
counts.web_raw > 0
at least 8 signals remain plausibly relevant after obvious ads, generic traffic bait, and off-profile items are removed

Failure branch:

Identify which query dimension failed.
Replace weak queries; do not merely add synonyms.
Rerun content topic mine.
Do not manually wrap research.run output as 02_raw_signals.json.

Checkpoint 5: Select Topic Locally

There is no topic reviewer in this workflow. Read 02_raw_signals.json, cluster related signals, compare them with profile/history, and choose the topic as the agent.

Score candidate clusters on:

audience relevance
creator stance fit
freshness
conflict or misconception
material availability
ability to support a complete argument rather than one isolated headline

Reject a candidate when:

it duplicates content history
its evidence is only generic engagement
its public entry does not lead naturally to the actual topic
it has no creator-specific stance
available materials cannot support its factual claims

Write 03_topic_selection.json yourself:

{
  "topic_cards": [
    {
      "topic_id": "topic_001",
      "status": "selected",
      "title": "internal topic title",
      "theme": "topic cluster",
      "angle": "specific angle",
      "public_entry": "candidate first spoken hook",
      "core_opinion": "creator-specific thesis",
      "common_misunderstanding": "what the audience usually gets wrong",
      "audience_value": "why the audience should care",
      "why_selected": "comparison-based selection reason",
      "evidence_signals": [
        {
          "source": "social or websearch",
          "title": "signal title",
          "url": "source URL",
          "author_name": "optional",
          "published_at": "optional"
        }
      ],
      "rejected_alternatives": [
        {
          "title": "candidate title",
          "reason": "duplicate, weak evidence, low fit, or weak argument"
        }
      ]
    }
  ]
}

Keep exactly one status=selected card. Include at least three evidence signals spanning social and web sources. Record at least two rejected alternatives so selection is comparative rather than arbitrary.

Validate:

node <skill_dir>/scripts/validate_artifact.js \
  --type topic-selection \
  --input 03_topic_selection.json

Pass condition: the selected topic has a clear hook direction, thesis, audience value, evidence path, and no near-duplicate in history.

Checkpoint 6: Draft Longform Plan Locally

Create 03a_longform_plan_<topic_id>.json yourself:

{
  "input": {
    "profile": {},
    "topic_card": {},
    "longform_plan": {
      "plan_id": "longplan_<topic_id>",
      "target_duration_sec": 240,
      "topic": "spoken subject",
      "public_entry": "first spoken hook",
      "topic_reveal": "second spoken sentence naming the subject",
      "viewer_promise": "what the viewer will gain",
      "core_thesis": "one defensible thesis",
      "stance": "creator judgment",
      "audience_filter_turn": "how the opening narrows to the target audience",
      "outline": [
        {
          "section": "section label",
          "claim": "one main claim",
          "points": ["supporting point"],
          "bridge_to_next": "why the next section follows",
          "evidence_role": "none"
        }
      ],
      "fact_boundary": ["claim requiring verification or softer wording"]
    }
  },
  "options": {
    "language": "zh-CN"
  }
}

Allowed evidence_role: none, example, analogy, supporting_evidence, direct_proof. Omit bridge_to_next only on the final outline item.

Required logic:

public_entry: first spoken hook
topic_reveal: next sentence explicitly naming the subject
viewer_promise: why the viewer should continue
each outline item has one claim
each non-final item has bridge_to_next
each item declares evidence_role
analogies are not labeled as direct proof

Validate:

node <skill_dir>/scripts/validate_artifact.js \
  --type plan \
  --input 03a_longform_plan_<topic_id>.json

Pass condition: a viewer hearing only the first three fields knows the topic and expected payoff, and every adjacent section has a stated logical relationship.

Checkpoint 7: Review Plan

Run:

luma-cli agent plan-review \
  --input 03a_longform_plan_<topic_id>.json \
  --model basic_model \
  --output 03b_plan_review_<topic_id>.json

Inspect decision, score, issues, and revision instructions. A backend pass does not override local topic-reveal, continuity, or evidence-role failures.

Failure branch:

Revise only the plan.
Save a versioned plan.
Rerun plan review.
Do not start material collection before acceptance.

Checkpoint 8: Plan And Collect Materials

Use the bundled material-planning and collection scripts:

node <luma-find-material>/scripts/plan_from_topic_review.js \
  --review 03_topic_selection.json \
  --topic-id <topic_id> \
  --longform-plan 03a_longform_plan_<topic_id>.json \
  --max-web-queries 2 \
  --max-image-queries 1 \
  --output 04_material_plan_<topic_id>.json

node <luma-find-material>/scripts/collect_from_material_plan.js \
  --plan 04_material_plan_<topic_id>.json \
  --execute-collection \
  --results-dir materials/<topic_id> \
  --deliverables-dir materials/<topic_id>/final_assets \
  --output 05_material_assets_<topic_id>.json

Validate:

node <skill_dir>/scripts/validate_artifact.js \
  --type materials \
  --input 05_material_assets_<topic_id>.json

Required:

every chapter has a ready asset or generated-component spec
factual claims have verified evidence or softened wording
rejected and failed assets remain recorded
materials/<topic_id>/final_assets/deliverables_manifest.json exists
search snippets never count as proof
official pages and primary documentation outrank media and social signals
social signals show attention, not factual truth
generated components explain the argument but never prove it

Do not continue with an empty material result.

Checkpoint 9: Build Writing Brief

Run:

node <skill_dir>/scripts/build_script_writer_payload.js \
  --profile 01_profile.json \
  --profile-extra 01_profile_extra.md \
  --topic-selection 03_topic_selection.json \
  --topic-id <topic_id> \
  --longform-plan 03a_longform_plan_<topic_id>.json \
  --plan-review 03b_plan_review_<topic_id>.json \
  --material-plan 04_material_plan_<topic_id>.json \
  --material-assets 05_material_assets_<topic_id>.json \
  --deliverables materials/<topic_id>/final_assets/deliverables_manifest.json \
  --duration-sec <seconds> \
  --output 06_script_writer_payload_<topic_id>.json

This file is a local writing brief. Do not submit it to backend script.write.

Pass condition: the brief contains the selected topic, approved plan, claim constraints, material coverage, and section briefs.

Checkpoint 10: Write Script Locally

Write 07_script_<topic_id>.json yourself:

{
  "topic_id": "<topic_id>",
  "title": "internal and cover title",
  "hook": "first spoken hook",
  "topic_reveal": "spoken subject sentence",
  "viewer_promise": "spoken payoff sentence",
  "sections": [
    {
      "section": "section label",
      "spoken_text": "complete spoken copy including transition",
      "claim_ids": ["claim_001"],
      "material_asset_ids": ["asset_001"],
      "evidence_role": "supporting_evidence",
      "bridge_to_next": "spoken transition logic",
      "visual_intent": "what the audience should see"
    }
  ],
  "full_script": "complete concatenated spoken copy",
  "evidence_notes": [],
  "risk_notes": [],
  "estimated_duration_sec": 240
}

full_script must contain the same spoken content as the ordered opening and sections. The title is not spoken context unless its words also appear in full_script.

Mandatory opening order:

public_entry -> topic_reveal -> viewer_promise -> audience_filter_turn

Mandatory section rules:

spoken text, not outline prose
one argument per section
explicit bridge between adjacent sections
source role stated accurately: example, analogy, supporting evidence, or direct proof
only ready asset IDs in material_asset_ids
no title-card-dependent context

Validate:

node <skill_dir>/scripts/validate_artifact.js \
  --type script \
  --input 07_script_<topic_id>.json

Pass condition: the full script can be understood with the screen turned off.

Checkpoint 11: Review And Revise Script

Create 08_script_review_payload_<topic_id>.json:

{
  "input": {
    "profile": {},
    "topic_card": {},
    "longform_plan": {},
    "plan_review": {},
    "material_assets": {},
    "script": {}
  },
  "options": {
    "language": "zh-CN"
  }
}

Use the accepted profile snapshot, selected topic card, approved plan, latest plan review, material coverage summary, and complete local script. Do not submit only full_script.

Then run:

luma-cli agent script-review \
  --input 08_script_review_payload_<topic_id>.json \
  --model basic_model \
  --output 09_script_review_<topic_id>.json

If the decision is revise or major_revise:

Map every issue to a script location.
Revise the local script, not the approved thesis unless the review identifies a plan defect.
Save versioned script, payload, and review files.
Rerun local validation and backend review.

Stop after three failed cycles and report the repeated blocker. Do not silently lower the standard.

Pass condition:

backend decision accepted
local topic-reveal, continuity, evidence, and material-ID gates pass
full_script is complete spoken copy

Checkpoint 12: Upload Durable History

Upload the accepted artifacts:

luma-cli content artifact upload \
  --profile <profile_id> \
  --type longform_plan \
  --input 03a_longform_plan_<topic_id>.json \
  --name longform_plan.current.json \
  --topic-id <topic_id> \
  --topic-title "<title>"

luma-cli content artifact upload \
  --profile <profile_id> \
  --type script \
  --input 07_script_<topic_id>.json \
  --name script.current.json \
  --topic-id <topic_id> \
  --topic-title "<title>"

luma-cli content artifact upload \
  --profile <profile_id> \
  --type script_review \
  --input 09_script_review_<topic_id>.json \
  --name script_review.current.json \
  --topic-id <topic_id> \
  --topic-title "<title>"

Verify with:

luma-cli --json content history --profile <profile_id>

Final handoff to luma-workflow-original-ip-talk:

accepted 07_script_<topic_id>.json
accepted 09_script_review_<topic_id>.json
05_material_assets_<topic_id>.json
deliverables manifest
explicit selected topic ID and title

Do not begin TTS or video production without this handoff.

luma-content-script

Luma Content Script SOP

Non-Negotiable Route