Compose · grounded sequence

Compose a grounded message sequence

The engine joins company and people by Company ID, researches only missing context, and keeps every result behind human review.

  1. 01WorkbookUpload + validate
  2. 02SellerOffer + evidence
  3. 03Message planTouches + writer
  4. 04PeopleSelect recipients

Excel input

Choose the company + people workbook

.xlsx
No workbook selected.
Required columns and matching rules
companyCompany ID, CompanyDomains, evidence URLs, and review notes are used when present.
peopleCompany ID, Current TitleFirst Name or Full Name is required. Email and phone are not sent to the model.
Evidence + safetyHow RevMessage protects the handoff 4 guarantees · 0 automatic sends

Evidence contract

No dummy facts. No guessed URLs.

Workbook evidence is used first. If it is incomplete, the engine searches the public web and accepts only source-validated evidence.

  1. 01

    Company and people rows join by exact Company ID

  2. 02

    Placeholder domains are rejected outside tests

  3. 03

    Web-researched URLs must pass configured-provider source validation

  4. 04

    Every result remains in review-only mode

2Required workbook sheets
1Exact Company ID join
4Grounded elements per touch
0Automatic sends

Seller setup

Choose what you sell

Loading
Seller featuresUsed as offer context, never as prospect evidence.

Web search and the separate catalogue deduplicator merge only high-confidence semantic duplicates. Both use the configured live provider and can consume provider quota. Superseded records stay recoverable in the audit log.

Message composerMessage plan and prompts 3 LinkedIn + 4 email
Build prompts from an existing sequence

Paste labeled touches such as “LinkedIn 1” and “Email 1”. Count, channel, order, and identity are always detected deterministically.

CTA definitions used by classifier and writer
No sequence analyzed.

Use either channel or both. At least one and no more than ten total touches are required.

Show prompt previewRead-only, with placeholders before research
Open this section to load the preview.
GenerationGeneration settings Loading writer · critic retry on · batch continues
Suggestions never trigger retries.
Off by default. Failed contacts remain visible while the batch continues.
Generation dock Start with a workbook. Select a person to keep their identity, readiness, and pipeline together. Review-only · nothing is sent.

After Generate Join → Research → Normalize → Evidence → Context → Qualify → Write → Contract → Critic → Save

GenerationConnecting to runtime…
Review only

Checking provider availability…

Current settingsLoading writer · critic retry on · batch continues

Confirm action

Continue?

Review this action before continuing.

Seller profile

Create seller

One offer per line, pipe-separated. An asset touch may only offer a resource listed here, using one of its exact phrases. Leave empty if this seller has no value-first asset.

These fields form the frozen seller policy. They guide positioning but never prove a buyer fact.

Seller capability

Add feature

Audience values: marketing-growth, ecommerce-merchandising, technical-product, operations-supply, sales-revenue, finance-procurement, people-hr, security-legal, general-management, or role-neutral.

Manual edits are saved as operator-supplied and unverified. Domain-extracted features retain their cited source until edited.

Human approval gate

Approve this exact revision?

This marks the sequence approved for downstream handoff (contract value SEND). This app still does not send it.

Type to confirm.

Token and quota warning

Generate with Default from frozen facts?

Research, evidence normalization, qualification, seller-feature matching, and the sequence plan will be reused. Only Default writing, deterministic checks, and the independent critic will run.

RevMessage may generate and criticize up to three complete candidates. It creates a separate REVIEW run and never approves or sends outreach.

Type exactly.

Same research · multiple writing styles

Compare drafts from frozen facts?

Every selected style receives the same frozen facts, provenance, seller-feature matches, exemplar, and sequence plan. Research and qualification are not repeated.

Select 2–13 compatible styles

The recommendation is deterministic: only REVIEW + critic PASS drafts qualify; fewer advisories and rewrites rank higher. A tie remains a human decision. Selection never approves or sends outreach.

Type exactly.

Token and quota warning

Review generation cost and people

Generation and independent critic calls can consume significant tokens or provider quota. No job is created until this dialog is confirmed.

Type exactly.

Immutable skill snapshot

Writer skill

SKILL.md
Referenced instruction files
Compiled instructions
Manifest and checksums
Version and usage history

Draft then Git review

Edit as new draft

Before / after diff
Save the draft to generate a server-validated diff.

Destructive skill action

Delete or archive writer skill?

Historical run snapshots remain. UI deletion never removes Git files.

Type exactly.

Immutable raw layer

Add knowledge source

Saving computes a SHA-256 checksum and immutable fragments. The raw source cannot be edited through the UI.

Compiled wiki layer

New knowledge article

Select one or more immutable sources. New articles cite the full source; API clients may bind exact fragment IDs.

Escaped read-only content

Knowledge detail


      
Provenance and revisions

Human knowledge gate

Validate article?

Type exactly.

Typed and source-backed

New knowledge item

Allowed preview stages

Saving never activates runtime prompts. Editing a validated item creates a new review-pending revision.

Preview-only identity

New knowledge bundle

Immutable item manifest

New preview bundle version

Only validated item revisions are selectable. The manifest checksum freezes their IDs, revisions, and checksums.

Exact human gate

Confirm knowledge action

Type exactly.

User guide assistant

Ask RevMessage

Ask a question or say “show me” for an on-screen walkthrough. The guide never generates, approves, or sends.

Workbook and privacy details

When you press Ask, your question and recent Help context go to the configured model provider. Walkthrough mode also sends a bounded semantic snapshot of visible control labels, interface status, entity labels or IDs, and non-sensitive selection states. Free-form values in text fields, textareas, and editable regions—plus passwords, tokens, and confirmation values—are excluded. Answers-only mode sends no interface snapshot or action receipts. RevMessage does not persist a Help conversation; context lasts only for this page session. An attached .xlsx workbook is staged into Compose only when you press Ask.