Context has an owner.
The Analyst curates the inputs before drafting begins. The Writer receives a bounded brief instead of being asked to perform every role at once.
HolyCalm Content Engine
One brief.
A sequence of
deliberate handoffs.
A content workflow that gives each agent a defined job: gather the right context, draft within constraints, critique the output, then bring it back to a person.
Workflow architecture · Based on workflow instructions
Follow the handoff.
Each role has a defined input and handoff. Read the sequence below.
01 · Route the request
Orchestrator
Start by choosing the context.
Extract the content type, topic, channel and quantity. The topic selects persona material; brand guidance is always loaded.
Needs: the user’s request.
“If content_type is unclear, ASK before proceeding. Do not guess.”
Source excerpt · Step 1
- Specified handoff
- A defined request and selected context files for the research roles.
02 · Analyst
Required research
Curate before generating.
The Analyst reads brand guidance and routed personas, extracting relevant language, voice rules and constraints into a brief.
Needs: the request, brand foundation and selected persona files. Can run alongside Context.
“Do NOT write any content yourself. Just extract and organize the raw material.”
Source excerpt · Step 3
- Specified handoff
- A brief of up to 500 words for the Writer.
02b · Context — conditional
Optional research
Add facts when the request needs them.
For pricing, competitor or feature messaging, the Context agent reads market data and selected PRD sections. Other requests skip this role.
Needs: the topic, market material and product documentation. Does not depend on the Analyst’s output.
“Only needed when topic includes price comparison, competitor positioning, or feature-based messaging.”
Source excerpt · Step 3
- Specified handoff
- Up to 300 words of supporting context, when this role is used.
03 · Writer
Generation
Wait for the briefs.
The Writer combines the Analyst’s brief with Context output when relevant. It drafts twice the requested quantity so the Critic has candidates to compare.
Needs: Analyst output, plus Context output if that role was required.
“No fake numbers, no fake reviews, no FOMO language”
Source excerpt · Step 4, rule 10
- Specified handoff
- Candidate content with a hook, body, call to action and rationale.
04 · Critic
Evaluation
Return reasons with the selection.
The Critic applies five criteria and explicit rejection rules. It ranks candidates and attaches a suggested improvement to each selection.
Needs: the Writer’s candidates and the supplied brand rules.
“For each selected piece, note 1 specific improvement suggestion”
Source excerpt · Step 5
- Specified handoff
- Ranked selections, criterion scores, suggestions and a rejection summary. No actual scores are shown here.
05 · Human review
Return to the conversation
Make the result reviewable.
The orchestrator presents the selected and rejected pieces, then invites adjustments. The instructions end with review and revision; they do not specify automatic publishing.
Needs: the Critic’s selections, scores and reasons.
“Would you like to adjust any of these? Or generate more variations of a specific one?”Source excerpt · Step 6
- Specified handoff
- A result that the person can inspect and refine in the conversation.
This diagram explores the documented architecture. It does not run agents or display generated content from a recorded execution.
The interface behind the output.
The design work here lives in the instructions: what information enters, which role can act next, and how a person sees the result. HolyCalm’s brand rules and audience research become context for a coordinated drafting and review process.
This is a separate content-production workflow. It does not describe HolyCalm’s in-app experience, meditation audio generation or a chat feature inside the app.
Read the system
Parallel where independent.
Sequential where it matters.
The Analyst and optional Context agent can work together. The Writer waits for their briefs. The Critic waits for the draft.
The complete dependency map
Always loadedBrand foundation
Voice, vocabulary and guardrails
Selected by topicAudience material
Relevant personas and language
Orchestrator
Turn the request into a brief.
Extract content type, topic, channel and quantity. Route to audience material. Ask if the content type is unclear.
TypeTopicChannelQuantityLanguage
Analyst · required
Extract relevant context.
Read the brand and routed personas. Return a concise brief for the Writer.
Output: brand + audience briefContext · conditional
Add product and market facts.
Runs for price, competitor or feature messaging. Reads market data and selected PRD sections.
Output: supporting contextWait for the required briefs
Writer
Draft within the constraints.
Combine the briefs with the requested format. Produce twice the requested quantity for the Critic to filter.
Output: candidate contentCritic
Evaluate. Reject. Explain.
Apply the rubric and rejection rules, rank candidates, then attach an improvement suggestion to each selection.
Output: selections + rejection reasonsHuman review
Keep the final judgment visible.
Present the selections, scores and rejected pieces. Invite adjustments or more variations.
Output: a reviewable result in the conversationContext selection
A different topic opens
a different set of sources.
The brand foundation stays constant. The audience material changes with the request. These routes are excerpts from the documented routing table.
Sleep
Sarah + Grace
Two persona sources inform the brief.
Grief
Grace
A focused route for loss and comfort.
Price or comparison
Market-wide context
Buyer research and demographics, plus the optional Context agent.
These are research personas named in the instructions, not people interacting with the system or verified audience results.
Inspect the instructions
Small contracts.
Clear responsibilities.
Open a role to see its inputs, output and a curated excerpt from the source. These are instructions, not transcripts of a successful run.
AnalystExtract, don’t draft
Inputs: the request, brand foundation and selected persona files.
Output: a brief of up to 500 words with relevant voice rules, audience language and guardrails.
“Do NOT write any content yourself. Just extract and organize the raw material.”
Curated excerpt · SKILL.md · Step 3, Analyst
ContextOnly when needed
Inputs: market material and selected PRD sections.
Output: up to 300 words of competitive and product context. It can run alongside the Analyst.
“Only needed when topic includes price comparison, competitor positioning, or feature-based messaging.”
Curated excerpt · SKILL.md · Step 3, Context agent
WriterWait for the brief
Inputs: Analyst output, optional Context output and the requested format.
Output: twice the requested number of pieces, with a hook, body, call to action and rationale.
“No fake numbers, no fake reviews, no FOMO language”
Curated excerpt · SKILL.md · Step 4, rule 10
CriticMake rejection legible
Inputs: candidate content and brand rules.
Output: ranked selections, criterion scores, suggestions and a rejection summary.
“Theological violation → REJECT”
Curated excerpt · SKILL.md · Step 5, rejection rules
OrchestratorReturn control to a person
Inputs: the Critic’s selection and reasons.
Output: a presentation of selected and rejected content, followed by an invitation to refine.
“Would you like to adjust any of these? Or generate more variations of a specific one?”
Source excerpt · SKILL.md · Step 6
Evaluation design
A rubric, not a result.
The Critic is instructed to assess five dimensions. No measured scores or performance gains are represented here.
Voice
Does the language fit the brand?
Audience
Does it match the selected persona?
Opening
Does it begin with the person’s problem?
Price
Is the price framed as instructed?
Guardrails
Does it respect theological constraints?
A hard rejection is different from a low score.
The instructions reject prohibited vocabulary and theological violations regardless of the total. This makes some constraints explicit, but a model’s judgment still needs human review.
What this makes visible.
Dependencies shape the workflow.
Parallelism is used for independent research. Drafting and critique follow their inputs, making the handoffs easier to inspect.
Reasons travel with the output.
The person reviewing gets rejection reasons and improvement suggestions alongside the selected content.
The next questions to test.
A useful evaluation would compare actual drafts, critique decisions and human edits across repeated requests. Does the rubric agree with human judgment? Are source facts current? Which rejected pieces should have survived?
The source describes the intended workflow. It does not provide a benchmark, run history, measured quality improvement or an enforced publishing approval mechanism. Those are separate things to demonstrate.
Source: HolyCalm Content Engine, SKILL.md · Architecture, Steps 1–6 and Performance Tips. This page is a curated explanation of that local instruction file; it does not execute agents.
Inside the workflow
Instructions you can inspect.
- Present the selected and rejected pieces.
- Return the selections, scores and reasons for review.
- Would you like to adjust any of these? Or generate more variations of a specific one?