An Open-Source Plugin That Keeps Your AI Sales Agent From Making Things Up

Hand a typical AI sales assistant a customer thread and you'll hit the same failure modes every time:

It treats an offhand "maybe we'll buy next month" as a hard fact and writes it straight into the CRM.
It "sends" the email or changes the deal stage before you've confirmed anything.
It blends one customer's private details with reusable sales playbook content.
When it drafts a customer email, it can't tell which lines are backed by evidence and which it invented.

SalesFlow exists to close exactly these gaps. It's now officially open source (MIT license), and you can install it directly in Codex, Claude Code, ChatGPT, or Synaro.

Repository: github.com/YIING99/salesflow-plugin

What Is SalesFlow, Exactly?

In one sentence: SalesFlow is a cross-platform sales workflow orchestration layer — not another CRM.

It doesn't store your customer data and isn't tied to any single company's internal system. It does one thing: translate a vague sales request into an evidence-grounded, confirmable workflow:

evidence -> identity check -> fact/judgment split -> seller artifact -> write plan -> user confirmation -> readback

This sequence is SalesFlow's core contract. Here's what each step means:

Start from evidence — not memory or guesswork.
Check identity — confirm which customer and which company you mean.
Split fact from judgment — what the customer said is a fact; "they'll probably buy" is a judgment. The two are never mixed.
Produce the seller artifact — an email draft, a meeting brief, a competitive battlecard.
Show a write plan — tell you exactly what it intends to write, and where.
Wait for confirmation — it never persists anything just because it was triggered.
Read back — after writing, it reads the result back for you to verify.

For a salesperson, that turns the AI from an "eager but unreliable intern" into an assistant that reports first, then acts.

8 Sales Workflows, Ready Out of the Box

SalesFlow ships with 8 bilingual workflows. Both English and Chinese triggers work, dispatched by a single salesflow router:

Workflow	English triggers	What it does for you
`account-research`	account research, company research, customer profile	Turns scattered information into a structured account profile
`communication-summary`	call summary, meeting recap, extract sales notes	Pulls hard facts and action items from a conversation, fact vs. judgment kept separate
`followup-draft`	draft follow-up, draft outreach, follow-up email	Drafts a follow-up within the evidence boundary — no invented promises
`daily-briefing`	daily briefing, today's follow-ups, sales briefing	Ranks which accounts to touch first today
`meeting-prep`	meeting prep, call prep, prepare for customer meeting	A one-page brief: context, agenda, risks, goals
`competitive-brief`	competitive brief, battlecard, objection handling	Competitor comparison plus objection-response cards
`pipeline-review`	pipeline review, account review, weekly sales review	Walks the pipeline, flags stalls and next steps
`quote-readiness`	quote readiness, quote prep, quotation check	Pre-quote self-check: is the info complete enough to quote?

You don't need to memorize commands. Just say "help me prep for tomorrow's call with Account A" in plain language, and SalesFlow routes it to meeting-prep.

Why "Bilingual" and "Cross-Platform" Matter

Bilingual, because real export and global sales constantly switch between languages. The customer's email is in English, your internal review is in Chinese — your AI tool shouldn't force you to pick one. Every SalesFlow workflow supports both languages natively, in triggers and in output.

Cross-platform, because nobody wants to be locked into a single tool. The same sales workflows should run in whatever AI you already have open:

Codex: use plugins/salesflow-plugin/.codex-plugin/plugin.json
Claude Code: claude plugin marketplace add YIING99/salesflow-plugin, then claude plugin install salesflow-plugin@salesflow
ChatGPT: copy adapters/chatgpt/PROJECT_INSTRUCTIONS.md into a custom project
Synaro / other agents: use adapters/synaro/SALESFLOW_WORKFLOW_PACK.md

The workflow contract is vendor-neutral — it targets abstractions like sales_knowledge_base, personal_knowledge_base, lightweight_tracker, and quotation_workflow, not any specific product. You can feed it pasted notes, files, CRM exports, or any knowledge base.

Install in Codex Quickly

# Add the repo as a local marketplace
codex plugin marketplace add github:YIING99/salesflow-plugin

# Install the plugin
codex plugin add salesflow-plugin@salesflow

For local development or to run the validator:

git clone https://github.com/YIING99/salesflow-plugin.git
cd salesflow-plugin
codex plugin marketplace add .
codex plugin add salesflow-plugin@salesflow
python3 plugins/salesflow-plugin/scripts/validate_salesflow.py

Want Persistent Memory? Connect KnowSales

SalesFlow is stateless by default — it works from whatever you paste in right now, but it won't remember anything for you. That's deliberate: workflow and memory are two different layers.

If you want your team's sales experience to accumulate, be reused, and persist across conversations, connect SalesFlow to a memory layer:

KnowSales as the sales memory layer — customer profiles, follow-up activities, product knowledge, objection cards, and competitor intelligence all persist here.
KnowMine as the personal / team learning layer — reusable judgment, preferences, and methodology.

Once connected, SalesFlow's "write plan" has somewhere to land: it produces a draft, you confirm, then it writes to KnowSales. SalesFlow owns "how to do it"; KnowSales owns "what to remember." We'll dig into these two layers in the next post.

If you are designing the broader sales system around it, read: An AI Sales OS Is Not Another CRM.

These integrations are optional. SalesFlow itself is fully vendor-neutral and works standalone without any memory layer.

Privacy and Safety by Default

Open source doesn't mean wide open. SalesFlow's safety rules live in the plugin's core/safety-gates.md and are enforced by default:

Never write just because it was triggered — every persistence step requires a write plan and your confirmation first.
Label key conclusions with evidence levels — from L1 (observed / raw record) to L5 (talk track / description), so you can see at a glance how trustworthy each line is.
Keep fact and judgment separate — the customer's words and the AI's inference are always labeled distinctly.
Never blend customer-specific info with reusable knowledge.

Summary

SalesFlow takes the most tedious, error-prone parts of selling — gathering evidence, producing artifacts, writing safely to durable systems — and turns them into an open-source, bilingual, cross-platform standard.

8 high-frequency sales workflows, ready out of the box
Natively bilingual; runs in Codex, Claude Code, ChatGPT, and Synaro
Confirm-first, write-second — the AI stops touching your CRM on its own
MIT-licensed open source, free to fork and customize

👉 View SalesFlow on GitHub and give your AI sales assistant a workflow you can actually trust.