poindexter

TTS pronunciation configuration. Controls how written text is normalized to natural spoken English before being passed to the Speaches/Kokoro TTS engine. Two DB-configurable app_settings keys let operators add or override pronunciation rules without code changes.

Image-direction prompts for the content pipeline. Generate featured-image prompts, build stock image search queries, run the image-director reasoning that picks which sections get visuals (image_gen vs pexels, style, and prompt/query for each) plus one featured hero image, and caption rendered images with vision-derived alt text.

Generate a native, community-appropriate Reddit value-post from a published blog post

Blog drafting & revision prompts. Turn a topic + research context into a publication-ready markdown article, then refine it against reviewer feedback. Includes the writer system prompt, the draft request, the anti-fabrication initial-draft prompt, an SEO-metadata helper, and the iterative-refinement prompt. Use during the content-generation stage of the pipeline.

Video director — given a post body + podcast script + target duration, produces a JSON shot list (ordered shots with per-shot source plugin, prompt/query, and duration) for the post's video. Enforces the no-AI-humans + stylized-not-photoreal policies. Used by the generate_video_shot_list pipeline stage. Operator brand is templated via {site_name}.

Classify a published post into brand content-types (multi-label). Populates post_content_types via ClassifyContentTypesJob.

General content-utility prompts — the writer-persona system prompt, content summarization, and free-form content-to-JSON conversion. Use for the writer system role, condensing content, or coercing content into structured JSON.

Prompts used by the composable atoms in services/atoms/ — the founder-voice dev_diary narrator and the LangGraph pipeline architect. These wrap the lower-level building blocks the TemplateRunner chains into pipelines.

Look up a past decision or memory specifically — filtered to memory-type embeddings (decision logs, project notes, feedback). Use when the user says "what did I decide about", "what was decided", "why did we choose", "remind me about [project]", or "what's our policy on".

Content quality assurance — the adversarial QA gate pack. Topic-delivery, internal-consistency, publication-readiness review, aggregate rewrite, writer self-review for contradictions, self-consistency sampling, an LLM quality rubric, and vision-QA for inline images and rendered preview screenshots. Use after a draft exists, during the multi-model QA stage, to score, critique, and revise content before publication.

Podcast script generation. Rewrite a finished blog article into a single-narrator, TTS-ready podcast script — translating every written/visual convention (URLs, "see below", headings, abbreviations) into natural spoken English. Use after a post is written, when generating the audio/podcast surface for it.

Data-retention and memory-hygiene summarization. Compress one calendar day of source-table rows, or a cluster of older same-source memories, into a single dense paragraph. Use when bounding storage growth while preserving the gist for retrieval.

Topic research & analysis. Turn raw web-search results or internal records into structured topic decisions: analyze search results, rank topic candidates against weighted niche goals, and distill a post angle from internal records. Use during topic discovery and ranking, before a post is written.

System prompts for the Poindexter voice agent — the local-LLM "Emma" persona spoken on the Pipecat / LiveKit surfaces, and the TTS-friendly override used when a voice room bridges to Claude Code. Resolved by services/voice_prompts.py::resolve_voice_prompt.

Two-pass blog writer for niche-driven content. Drafts a post from topic + angle + background context, then revises it to substitute [EXTERNAL_NEEDED: ...] markers with verified external facts. Use as the canonical_blog writer during content generation and revision.

Approve a content task for publishing. Use when the user says "approve task", "approve post", "looks good, approve it", or "approve [id]".

Show AI/LLM spending, budget usage, and operational cost metrics. Use when the user says "how much have we spent", "cost report", "show costs", "budget status", or "spending breakdown".

View or update pipeline settings stored in the database. Use when the user says "show settings", "set the budget to", "change the API key", "update config", or similar.

Short-form video narration. Turn a published blog article into a ~150-word spoken script for TikTok / YouTube Shorts — hook first, two or three takeaways, a closing call to action. Use during the media-script stage of the pipeline, after a post is written.

SEO metadata generation. Produce titles, meta descriptions, excerpts, keywords, tags, and category matches for a post. Use during the generate_seo_metadata pipeline stage, after the draft is written.

Operator-persona alert triage. Diagnose an alert plus curated database state into one short paragraph the brain dispatcher posts to the operator. Use when the firefighter service responds to a system alert.

Social-media copy generation for the content pipeline. Research current trends for a topic, draft a platform-native post, and produce char-limited promo copy for a published blog post (single-tweet + LinkedIn formats). The promo templates are vendor-agnostic — the caller supplies the character limit, so the same prompt works across X / Mastodon once an adapter exists. Use after a post is published, during distribution.

Create a new blog post or content task about a given topic. Use when the user says "write a post about", "create content about", "draft an article on", or similar.

Show content quality scores and QA results for tasks that have cleared the pipeline. Use when the user says "quality report", "show quality scores", "how's the content quality", or "QA results".

Content quality evaluation. Score a draft for quality and return a structured verdict. Use after a draft is written, to gate it before publication.

Long-form blog drafting and narrative composition. The one-shot blog-generation task prompt plus the bundle-grounded narrative reporter persona. Use when generating a blog draft from a topic, or composing a grounded narrative summary of shipped work.

Marketing-automation prompts — generate an email campaign for a topic. Use when an operator-facing automation surface needs a structured campaign draft.

Business analysis prompts — financial impact, market analysis, and performance analysis for a given topic. Use when an operator-facing business module needs a structured first-pass analysis.

List content tasks, show the pipeline queue, or check task status. Use when the user says "show tasks", "what's in the queue", "pipeline status", "list posts", or "what's pending".

Reject a content task and send it back for revision. Use when the user says "reject task", "reject post", "send it back", "needs more work", or "reject [id]".

Check Vercel deployment status, recent deployments, and domains for the public site via the Vercel REST API. Use when asked about the website, site status, or Vercel deployments.

Turn off the LiveKit voice bridge for this Claude Code session. Use when the user says "voice off", "/voice off", "stop voice", "leave the voice room", or "end voice".

Turn on the LiveKit voice bridge for this Claude Code session. Use when the user says "voice on", "/voice on", "let me talk to you", "voice mode", "join the voice room", or "start voice".

Speak text into the active LiveKit voice bridge — TTS the assistant's reply into the room. Use when voice mode is on and the assistant wants to vocalize a reply, or when the user says "say this out loud".

Recall something said earlier in this voice conversation — semantic search over the voice_messages table. Use when the user says "what did I say about X earlier?", "remind me what we talked about", "did I mention X already?", "what was that thing I said about". Different from recall-decision (curated memory) — this is the in-conversation transcript.

Find blog posts semantically similar to a topic — pulls from the posts embedding source. Use when the user says "have we written about", "any posts on", "find similar posts to", "what's our coverage of", or wants to avoid duplicating an existing post.

Search semantic memory across all embedded content (memory, posts, issues, audit). Use when the user says "what do I know about", "search memory for", "recall", "find references to", or asks open-ended questions about prior context.

Manually publish a completed content task. Use when the user says "publish post", "push it live", "publish task", or "publish [id]".