Track “I'm interested” comments in one record, then draft the DM reply for approval
Draft the comment intake, the DM reply, and a five-question fit-band form. Hold every send for your approval. Free prompts and contracts for your own tool setup.
3
build recipes
1
week of content
0
sales calls
PLATE 02PUBLIC TEASER
What's inside.
01A reply script for “I'm interested” comments, with a tracked-record field map and a DM draft for approval
02One core idea, one platform, a full week of content
03A five-question form that places responses into draft fit bands for your review, with no automatic send
BRIDGE 01THE PATH
2-Step No-Call Engine
This free kit handles the first narrow problem.
01 · FREEThis free kit
Draft a comment-to-reply path, plan one platform's content week, and set initial fit bands for human review.
02 · DFYDone-For-You Install
Installs the mapped engine in your tools with the agreed human gates.
PLATE 03THE GIVE
The three builds.
Each build includes prompt blocks, run steps for your own tools, and an acceptance test. Live integrations stay held until their canaries pass.
01
CAPTURE / REPLY
Turn comments into tracked leads
How to work with your assistant
What your agent does Drafts the intake, reply, record map, and reconcile contract from the inputs you provide.
What you do Supply the account access, approve every send, judge fit, and run one real canary before wider use.
Ongoing help: bring the saved contract, the failed acceptance step, and the redacted log to your working session. Do not paste credentials into a prompt.
RUNTIME / SELF-SETUP
Runner-neutral contract. Test it in your chosen agent runner before live use.
You are helping me stand up an Instagram comment and inbound-DM capture flow in my own assistant.
Ask me for these inputs one at a time: the Instagram account, the approved comment and inbound-DM triggers, the qualifying-form URL, two real replies I wrote, my CRM choice, and the human approver.
Build three separate stages: INTAKE, REPLY, CAPTURE. Accept only the two event shapes in the intake contract and fail closed on malformed input. Put a human-approval gate before every send. Never request or print a secret. When a console key is needed, mark it WHAT I DO and wait.
Return: the two-branch event contract, the shared human-hold and draft-reply contract, the CRM field map, the idempotency rule, the polling/reconcile backstop, and a three-event acceptance test covering one comment, one inbound DM, and one malformed inbound DM. Stop after each stage for approval. Live webhook wiring stays deferred.
INTAKE
Watch the approved Instagram inbound surface for either configured event.
COMMENT: type=instagram_comment, event_id, account_id, post_id, comment_id, sender_handle, received_at, message_text.
INBOUND DM: type=instagram_inbound_dm, event_id, account_id, thread_id, message_id, sender_handle, received_at, message_text.
Reject any other or malformed shape before a write. Normalize either branch into one awaiting_review intake record with a stable submission_id. Reserve that identifier before the CRM write. A matching repeated event must return or update the same record. A conflicting repeat must fail closed.
Route both branches through the same held relay: human_hold=true, approval_required=true, auto_send=false. Do not send a message. Present the intake record and reply draft to the human approver. Live webhook mapping and credentials stay deferred to the operator-controlled OD-20 environment.
REPLY
Draft one first-touch reply in the voice examples I provided. Answer only the question asked, then offer the qualifying-form link. Mark the draft AWAITING HUMAN APPROVAL. Never send it automatically.
SASSISTANT NAMEAI
Opening message template
Hi [first name]. I am [assistant name], an AI assistant for [business]. I can answer the question you asked and point you to the right next step. Ask for a person anytime and I will hand this over.
[one-line answer]
If you want us to look at this properly, these five questions tell us where you are: [form link].
Placement contract: keep the AI label visible before and throughout the exchange, self-identify in the first message, and keep the human escape line. Do not add a blocking disclosure modal. The operator must recheck final guidance before launch.
CAPTURE / FIELD MAP
Write one lead record to GHL or HubSpot with exactly these fields:
first_name: captured or null
email_or_handle: captured value
interest_or_qualifier: captured value
source: instagram_comment or instagram_inbound_dm, copied from the validated event branch
submission_id: the stable intake identifier
status: awaiting_review
human_hold: true
auto_send: false
Use submission_id as the idempotency-key on every write. A matching repeated key must update or return the existing record, never create a twin. A repeated key with different event content must fail closed. After the event, poll the system of record and reconcile by submission_id. If no record appears, flag NONDELIVERY and stop for human review. No branch may send automatically.
Launch-gating leads to verify on live documentation
Webhook retry behavior. Default-safe: idempotency-key plus polling/reconcile stays mandatory. Verify the current retry rule for the InboundMessage seam.
Signature transition date. Default-safe: target Ed25519 X-GHL-Signature now. Verify the live removal date for the legacy header.
Reels and Stories comment coverage. Default-safe: name any gap and use the ManyChat Pro fallback for unsupported surfaces. Verify current GHL coverage.
ManyChat Pro price. Default-safe: show no public price. Verify the billing screen before any price is stated.
Thread co-ownership. Default-safe: omit handover protocol from v1. Verify only if true multi-app thread control becomes required.
Canary acceptance
What your agent does: creates one comment event, one inbound-DM event, and one malformed inbound-DM event. It checks the CRM for each stable submission_id, confirms both valid events reach the same human-held relay, and confirms the malformed event creates no record or relay.
What you do: approve each valid test reply, confirm the disclosure is visible before or at first interaction, and compare each CRM record to its source message before widening the flow. Keep the live webhook held until the OD-20 environment is supplied and this canary passes.
02
ONE-PLATFORM CONTENT
Run one platform's content engine
How to work with your assistant
What your agent does Finds candidate ideas, reads recent results, plans the week, drafts each piece, and recuts one claim across one platform.
What you do Pick the platform and claim, provide real examples and results, and read every draft before it ships.
Ongoing help: keep the chosen claim, source notes, platform constraints, and your corrections together so the next run starts from evidence.
RESEARCH
For [one platform], find 10 current candidate ideas around [audience problem]. For each, return the source signal, the exact audience question, and one claim I can defend. Do not invent performance data. Mark anything without a source UNVERIFIED.
ANALYSIS
Read the recent posts and results I provide. Report what each post did in plain terms using only the supplied figures. Separate observation from inference. Do not claim which post earned a buyer. Return three patterns worth testing next week.
PLANNING
Take the one claim I select and plan a real week for [one platform]. Use only formats that platform supports. Give each day one job, one format, one opening line, and one proof point. Keep the same claim across the week.
SCRIPTING
Draft the planned week from the real voice examples I provide. Keep the claim and proof intact. Label every unsupported assertion CHECK. Return drafts only. I will read and approve each one before it ships.
REPURPOSE
Re-cut the approved source piece into the planned week for [one platform]. Preserve the claim, vary the angle and format, and show which source sentence supports each cut. Stop before posting or scheduling.
This package contains one editable 1080 x 1920 Remotion composition and its render command.
What your agent does
Reads content.json and checks that every value is real.
Places your clip.mp4 in public/.
Runs npm install, then npm run render.
Verifies that out/starter-video.mp4 exists and reports its duration and dimensions.
What you do
Provide the clip and approve the three content lines.
Confirm your Remotion license covers your use.
Watch the rendered file before publishing it.
Paste-ready contract
Read this package before changing it. Ask me for the source clip, title, proof line, and closing line. Write only those approved values to content.json. Put the clip at public/clip.mp4. Install the declared packages, render StarterVideo, and verify the output file, duration, and 1080 x 1920 dimensions. Stop and ask if the clip or any approved line is missing.
PACKAGED EDITOR / WHOLE
The video template and runner
The package carries the editable composition, content file, and render command. Add your own clip, edit the supplied content values, then render the file you own.
Remotion is free for individuals and companies of three people or fewer. A recipient with four or more employees needs their own remotion.pro company license. The license travels with the tool.
Build one approved clone video from your own accounts.
How to work with your assistant
What your agent does Guides setup, locks the script, creates the chunk plan, drives your HeyGen and ElevenLabs connections, assembles the clip, and returns a review packet.
What you do Own and pay for both accounts, provide a voice and likeness you have the right to use, keep credentials in your own secret controls, approve the script and spend before generation, then watch and approve the finished clip before publication.
Ongoing help: bring your setup record, approved script, run manifest, vendor receipts or credit usage, and the failed review item back to your working session. Never paste an API key, password, voice id, or avatar id into a public prompt or support message.
Evidence: We have produced real clone videos through this audio-first flow. Your first run stops at a human review to protect your accounts and stay inside each platform's rules.
This is an agent-run contract with marked human gates. A manual-only step list does not meet this build.
SELF-SETUP / READINESS
You are helping me set up one clone-video workflow in my own assistant using my own paid HeyGen and ElevenLabs accounts.
OUTCOME
Return a redacted readiness record for this flow: approved script -> ElevenLabs voice audio -> HeyGen avatar video -> assembled review clip. Do not generate media during setup.
RULES
- Ask one question at a time. Never request, display, or store a password, API key, access token, voice id, or avatar id in chat.
- Use only a voice, likeness, script, music, font, logo, and other asset that I confirm I own or am authorized to use.
- Treat vendor plans, connector availability, commercial-use terms, generation credits, and disclosure rules as live facts to verify in current official documentation and my billing screens.
- Stop at every WHAT I DO item and wait for my confirmation.
SETUP
1. WHAT I DO: create or select the HeyGen and ElevenLabs accounts owned and billed by me. Confirm the intended channel, audience location, and commercial or non-commercial use.
2. WHAT MY AGENT DOES: return the current official links and a checklist for the required paid-plan features. Mark any unverified plan or price NOT VERIFIED. Do not choose or buy a plan for me.
3. WHAT I DO: record the avatar source as one continuous two-minute take. Use 1080p minimum, with 4K ideal. Put the camera at eye level and frame from chest to just above the head. Use even lighting, a simple background, a natural expression, and keep hands below chest. Make no edits and keep only me on camera.
4. WHAT I DO: provide about thirty minutes of clean voice audio from a quiet space with a moderate-to-professional microphone. Existing call or class recordings are acceptable when they are clean.
5. WHAT I DO: in ElevenLabs, complete the vendor's consent process, create my authorized voice clone from the approved source, and store its reference in my own secret or environment controls.
6. WHAT I DO: in HeyGen, complete the vendor's consent process, create my authorized avatar from the approved source, and store its reference in my own secret or environment controls.
7. WHAT MY AGENT DOES: guide me through connecting the current official HeyGen and ElevenLabs MCP servers or approved connectors to my assistant. Credentials stay in my own secret controls. Confirm access by listing the selected voice and avatar without generating media.
8. WHAT MY AGENT DOES: return READY or NOT READY for accounts, rights and consent, source quality, connections, secret handling, output format, disclosure requirement, and a client-set spend ceiling.
STOP CONDITION
Do not move to generation. Present the readiness record and wait for my approval.
01
Accounts and rights
Agent: checks current plan requirements and records unknowns.
You: open and pay for both accounts; confirm consent and usage rights.
02
Voice and avatar
Agent: checks the source recordings against the worked capture recipe without receiving credentials.
You: record, consent, create both clone assets, and retain their private references.
03
Assistant connections
Agent: guides the official MCP or connector setup and runs a non-generation listing check.
You: enter credentials only in your own secret controls.
04
Readiness record
Agent: returns READY or NOT READY with evidence and open checks.
You: set the spend ceiling and decide whether generation may begin.
Record source material that holds up.
Avatar source Two continuous minutes, 1080p minimum with 4K ideal. Camera at eye level, framed chest to just above the head. Use even lighting, a simple background, a natural expression, hands below chest, no edits, and only you on camera.
Voice source About thirty minutes of clean audio in a quiet space with a moderate-to-professional microphone. Existing call or class recordings are acceptable when they are clean.
SCRIPT / PREFLIGHT CONTRACT
Prepare my script for one clone-video run. Do not generate media.
INPUTS: draft script; audience; destination; target aspect ratio and duration; approved voice examples; pronunciation notes; avatar framing; caption style; required disclosure text; spend ceiling; readiness record.
WORK
1. Check the script against the supplied facts. Label every unsupported claim CHECK.
2. Return a final reading script and a chunk manifest. Each chunk must end at a sentence boundary and stay under 60 seconds.
3. For every chunk, show the exact text, estimated duration, pronunciation notes, framing cue, and expected vendor generation step.
4. State which plan, price, credit, connector, license, or disclosure facts still need current official verification.
5. Give a pre-generation estimate using only figures visible in my own vendor billing screens. If exact spend cannot be known, state the known unit and the uncertainty.
HUMAN GATE 1
Mark the result AWAITING SCRIPT AND SPEND APPROVAL. Stop until I reply with the exact approved script version, approved chunk manifest, and APPROVE GENERATION. Any later script change cancels that approval.
PRODUCE / ASSEMBLE CONTRACT
Produce one review clip only after HUMAN GATE 1 is satisfied. Use my connected HeyGen and ElevenLabs accounts. Never publish, post, schedule, or send the clip.
INPUTS: approved script version; approved chunk manifest; approved spend ceiling; private voice and avatar references from my secret controls; output specification; required disclosure asset or text.
FLOW
1. Create ElevenLabs voice audio for one approved chunk at a time. Use model eleven_multilingual_v2 with stability 0.4, similarity boost 0.8, style 0.0, and speaker boost on. If that model is unavailable, stop and mark the substitution NOT VERIFIED. Save each returned audio file and its vendor job reference. Retry only a failed chunk and record any extra credit use.
2. REQUIRED TWO-PASS LOUDNESS STEP: run a first FFmpeg loudnorm measurement pass at I=-14, TP=-1.5, and LRA=11. Run a second linear loudnorm application pass using the measured I, TP, LRA, threshold, and offset. Re-measure and log integrated loudness and true peak. The worked recipe corrected -40.6 to -14.9 LUFS without clipping. Raw voice output can be whisper-quiet, so stop if this step is skipped or the measured result misses the approved destination target.
3. Create the matching HeyGen avatar clip from each approved audio file, one chunk at a time. Save each returned video file and its vendor job reference.
4. Check that every audio and video artifact exists, opens, matches the manifest order, and has a plausible duration. Stop on a missing or mismatched artifact.
5. Assemble the avatar clips in manifest order with the media tool available in my environment. Add only the approved captions, disclosure, and brand assets. Do not add an unapproved person, voice, claim, music track, or logo.
6. Run technical checks for dimensions, duration, audio presence, loudness, clip order, black frames, and visible seams. Return the review clip, manifest, check results, and vendor-reported credit or cost usage.
STOP CONDITION
Mark the clip REVIEW REQUIRED. Do not call it publish-ready and do not release it.
REVIEW / RELEASE CONTRACT
Prepare the clone-video review packet. Do not publish or schedule anything.
CHECK THE FULL CLIP AND EACH CHUNK
- Exact approved script and chunk order
- Authorized voice and likeness only
- Lip sync, pronunciation, pace, expression, framing, and continuity
- No dropped audio, black frames, stitch seams, or audio-video drift
- Captions, brand assets, and required synthetic-media disclosure match the approved inputs
- Final format matches the destination specification
- Vendor-reported credit or cost usage is recorded, including retries
Return PASS or FAIL for every item with a timestamp and correction owner. Any failure routes back only to the failed step or chunk, followed by a fresh review packet.
HUMAN GATE 2
Mark the final result AWAITING HUMAN OUTPUT APPROVAL. I must watch the entire clip, confirm the disclosure and destination rules, and reply APPROVED FOR PUBLISH. Your role ends at the approval packet. Never publish, upload, schedule, or send the video.
01Approved script
Locked version and sub-60-second chunk manifest.
02ElevenLabs voice
One authorized audio file per approved chunk.
03HeyGen avatar
One authorized avatar clip per approved audio file.
04Assembled clip
Ordered media, captions, disclosure, and technical checks.
05Human release
Full watch-through and explicit approval before publication.
Client-owned accounts, rights, and cost
Paid accounts: this flow requires your own paid HeyGen and ElevenLabs accounts. You choose the plans, accept the terms, and pay both vendors directly.
Variable spend: generation consumes vendor credits or usage. Script length, selected models, resolution, retries, and plan rules change the total. No fixed public price is promised here.
Separate billing checks: do not assume a web subscription includes MCP or API access, or that it covers every generation charge. Verify each billing screen before a paid run.
Rights and license: use only material you own or are authorized to use. Before commercial publication, verify the selected plans permit the intended voice, likeness, assets, API or MCP use, destination, and audience location.
Leads to verify for each client's first run
Current connector and plan access. Verify the official HeyGen and ElevenLabs MCP or approved connector requirements in the client's assistant and paid accounts.
ElevenLabs PVC source duration. Verify the current Professional Voice Clone source-audio requirement in official ElevenLabs documentation before the first run. The about-thirty-minute recipe reflects the validated run, not a standing vendor minimum.
Current billing and commercial-use terms. Verify plan price, included credits, extra usage, API billing, output rights, and any renewal or cancellation terms in the client's billing screens and official terms.
Current disclosure and destination rules. Verify the required synthetic-media label or disclosure for the publishing platform, audience location, and jurisdiction before release.
Client-contract dry-run. Have a first-time client follow the exact prompts without hidden operator context. Record unclear steps and missing assumptions, then refine the prompt wording. This tests client paste-readiness.
EVIDENCE · PROVEN FLOW
The flow is proven. Your first run makes it yours.
The underlying flow is proven. It has produced real clone videos through the same audio-first sequence described here.
03
QUALIFICATION
Score responses into draft fit bands for human review
How to work with your assistant
What your agent does Builds the form, scoring contract, instant reply, and the selected routing depth.
What you do Read the answers, decide who gets in, correct the weights against real outcomes, and approve any routed action.
Ongoing help: bring the five questions, the current weights, and three redacted outcomes to the next review.
Five questions, a score that sorts, and an instant reply
BOUNDED BUILD CONTRACT
You are building a bounded qualification path for my offer.
Ask me for: my offer in one line, who it is not for, where leads arrive, my form tool, two real replies I wrote, and the human approver.
Build these five questions:
1. What are you trying to fix right now?
2. What have you already tried?
3. Roughly what can you put toward fixing it? Use ranges.
4. When do you want to start?
5. Where are you now with this?
Create one editable scoring table. Return three bands: clear fit, worth a look, not yet. Draft one instant reply that answers the question and links the form. Never send without approval. Show me the form, score table, and reply for approval before wiring them.
Form, score, conversational follow-up, guardrail brief, and routing
WHOLE-PATH CONTRACT
You are building the whole qualification path for my offer.
Ask me for: my offer in one line, who it is not for, where leads arrive, my form tool, two real replies I wrote, and the human approver.
Build these five questions:
1. What are you trying to fix right now?
2. What have you already tried?
3. Roughly what can you put toward fixing it? Use ranges.
4. When do you want to start?
5. Where are you now with this?
Create one editable scoring table. Return three bands: clear fit, worth a look, not yet. Draft one instant reply that answers the question and links the form. Never send without approval. Show me the form, score table, and reply for approval before wiring them.
BUILD:
1. Conversational follow-up: ask only the next question warranted by the answer. Never invent facts or reassure for fit.
2. Guardrail brief: list who the offer is not for, claims the qualifier cannot make, and every condition that requires a human.
3. Routing table: CLEAR FIT -> flag for human approval; WORTH A LOOK -> send the approved resource after approval; NOT YET -> record the reason and hold.
4. Record: write the answers, score, transcript summary, route, and stable submission_id to the CRM.
OUTPUT: the working form contract, score table, follow-up contract, guardrail brief, routing table, and a three-case acceptance test. Stop for human approval before any route fires.