Concept Generator (DNA-Lens)

Three-stage pipeline that replaces the host-agent-freestyle script-generation gap. Each stage runs as a separate script; chain them via the daily-loop workflow.

Pipeline

1. generate_concepts.sh -> writes runs/<date>/concepts.md
2. pick_concept.sh -> writes runs/<date>/concept-pick.md
3. draft_script.sh -> writes runs/<date>/script.md + concept-meta.json

Inputs

• Brand workspace at <brandsroot>/<brandslug>/ with at minimum a

brand-voice.md. Richer corpora (voice-profile.md, deliverables/recent/) produce better output.

• Trend research input from daily-loop Step 4 (markdown file with operator's

hot takes on selected trends).

Modes

• interactive (default): Stage 3 elicits operator specifics before drafting

any concept that claims a personal fact. Recommended for own-brand work.

• autopilot: Stage 3 never elicits; concepts claiming personal facts are

down-ranked at Stage 1 so the auto-picker avoids them. Recommended for agency / scaled work.

Mode resolves from: --mode flag > brand config.json concept_mode field

default interactive.

Stage 1: generate_concepts.sh

Loads brand corpus via lib.conceptcorpus.loadbrandcorpus, reads the trend input, and prompts the host agent (via the conceptgeneration template in prompts.json) to produce N ranked concepts. Each concept has:

• format (one of v0.7.0's seven formats)
• close_action (save / share / comment / soft)
• arc (1-line summary)
• whythisworks (3-5 sentences)
• claimspersonalfact (true if any concept-level fact requires operator

confirmation in Stage 3)

• 3-5 hook_variants, each with text, pattern (one of six in

references/hook-patterns.md), and 360 shock-score

• concept_score (aggregate 0-10)

Output: runs/<date>/concepts.md with markdown sections per concept, ranked top-to-bottom by concept_score.

In autopilot mode, concepts where claimspersonalfact: true get a -2 score penalty so the auto-picker avoids them. Operator still sees them.

Stage 2: pick_concept.sh

In interactive mode, the operator (or upstream agent) provides --concept N to pick the Nth ranked concept. In autopilot mode (no --concept), the script picks the highest-ranked autopilot-safe concept (claimspersonalfact false). If all concepts claim personal facts in autopilot, the script exits non-zero with a clear error.

Output: runs/<date>/concept-pick.md (a copy of the picked concept's section + the chosen hook variant; first hook variant by default).

Stage 3: draft_script.sh

Reads concept-pick.md. In interactive mode, if the picked concept has claimspersonalfact: true, prompts the host agent to elicit specifics from the operator using the scene_elicitation template. The agent asks the minimum questions documented in that template ("Have you actually X?", "What's your real number / date / situation?"). Operator answers; agent collects.

Then prompts the host agent (via scriptdraft template) to write the full script. Format slot template loaded from lib.formatregistry.get_format; slide bodies match alias headings.

Pre-output lint chain (any failure aborts):

1. lib.voicelint.linttext (v0.6.1): em-dash, AI-tell patterns, banned

words. --no-lint bypasses.

1. lib.formatlint.lintscript_structure (v0.7.0): slot count, word count,

close vocab. --no-format-check bypasses. Per-style style.yaml wordcountoverride (v0.8.0) applies when present (set BRANDSTYLEDIR env var to point at the brand's style directory).

1. lib.savefilter.checksave_worthiness (v0.8.1): heuristic gate (numbers,

dates, named frameworks, paired quotes, tactical asides) + optional LLM fallback. --no-save-filter bypasses. Adapts to effective word-cap: tight slots (max <=12) need >=1 marker; standard slots (>=13) need markers from >=2 categories.

Output: runs/<date>/script.md (frontmatter + slot bodies) and runs/<date>/concept-meta.json (data foundation for v0.8.1+ ingestion loop).

Non-dry-run invocation contract: the script expects script.md to already be present in --run-dir (the host agent writes it after reading the scriptdraft template + concept-pick.md). draftscript.sh then runs the 3-step lint chain and emits concept-meta.json. Bypass any step with --no-lint, --no-format-check, --no-save-filter.

References (one level deep)

• references/hook-patterns.md: six hook patterns with exemplars
• references/concept-patterns/: kit-shipped concept exemplars (six files)
• references/carousel-examples/: operator-supplied annotated carousel

reference (empty in default install; ingested when Matt provides)

• references/formats/: v0.7.0 format YAML files (consumed via

lib.format_registry)

Authoring a new hook pattern

The six patterns are sufficient per Bug 12. To add a seventh: append an entry to lib/hookpatterns.HOOKPATTERNS with description, exemplars, and whentouse. Mirror in references/hook-patterns.md. Tests enforce sync.

Authoring a new concept pattern

Drop a markdown file in references/concept-patterns/ matching the existing shape (Shape / Hook style / Format hint / Body density markers / Close style / Why it saves / Source exemplars). The skill loads all *.md files in the directory automatically.