DEV Community: aiFetchly

Build Your Own Local Knowledge Base So an AI Agent Can Write Client Emails That Sound Like You

aiFetchly — Sun, 01 Nov 2026 23:02:34 +0000

Disclosure: I am from the aiFetchly team.

Most generic AI emails fail for a simple reason: the model does not know your business. It does not know how you describe a service, which cases you accept, what you never promise, or how you follow up. A local knowledge base fixes that — not by adding more prompts, but by giving the agent a small, curated source of truth to draw on.

1. Start with your own documents, not the whole internet

Collect the material you already trust: service pages, FAQs, proposal templates, past replies you were happy with, pricing notes, and objection-handling answers. Keep it in one place and prune ruthlessly. If a document is outdated, remove it or mark it superseded. A knowledge base with 30 accurate pages beats one with 300 stale ones.

A practical structure that works well:

What we do / do not do: one page per service, with scope and limits.
Voice and tone: 5–10 approved examples, plus phrases to avoid.
Proof points: case summaries, timelines, and common questions, written in your words.
Follow-up rules: when to wait, when to nudge, and when to stop.

2. Keep the knowledge base fresh

Set a monthly 20-minute review. Check three things: prices or offers that changed, services you no longer offer, and replies that needed correction. Update the source document, not just the last email. If your agent keeps quoting an old offer, the problem is almost always an un-updated source, not the model.

Version your notes simply: add a date at the top of each page, and retire pages instead of leaving competing versions side by side.

3. Use it as context, not as a script

When drafting outreach, let the agent retrieve only the relevant pages for that prospect — their industry, problem, and stage — and then write in your voice. Ask it to stay close to your approved wording for claims, and to leave anything uncertain as a bracketed question for you, rather than inventing a detail.

A useful checklist before you send:

Does this email reference something specific about this prospect, not just their company name?
Is every claim traceable to a document in the knowledge base?
Would you be comfortable if the prospect replied and asked for detail?

4. Deduplicate before you send

Duplicate outreach is one of the fastest ways to look automated. Before a campaign goes out, deduplicate by domain and contact, merge contacts from the same company where appropriate, and filter out people you have already contacted recently or who already replied. Keep a simple activity log — who was contacted, when, and about what — so a second campaign does not restart a conversation from zero.

5. Keep a human in the loop

For important prospects, review the draft, edit one or two sentences to make it yours, and only then send. For larger lists, review a sample from each segment, fix recurring mistakes at the source document level, and resend only the corrected segment. Automation should save drafting time; it should not remove judgement.

A short workflow you can copy

Gather and prune your own documents into a local knowledge base.
Write one voice guide with approved examples.
Draft with retrieved context, not from memory.
Deduplicate, filter, and check recent activity.
Human-review the important drafts, fix the source, then schedule sending via SMTP on a timed workflow.

This is the approach we built aiFetchly around: multi-engine lead research, organising publicly available contact details, using your own documents and knowledge base as context for personalised emails, and SMTP scheduled sending with deduplication, filtering, and activity management, plus cron-based timed workflows and plugins for business knowledge. A short demo of a local knowledge base / private memory in action is here: https://www.youtube.com/watch?v=FTGsQ27HsfA — and the official site is https://www.aifetchly.com .

On pricing, for transparency: a free Community version is available, Plus starts at $10/month, and AI Content Assistant is Plus Only. Start with the free Community version to test whether a small, well-kept knowledge base improves your replies before you pay for anything.

What is the one document in your business that, if an AI agent always had it to hand, would most improve your emails?

When replies slip through a shared inbox: a triage routine that keeps leads from going cold

aiFetchly — Sat, 31 Oct 2026 23:02:55 +0000

Disclosure: I am from the aiFetchly team.

A shared inbox looks efficient until a reply slips through it. Two people answer the same prospect with different messages, nobody owns the thread, and the follow-up that should have gone out on Tuesday goes out a week later — if at all. The prospect does not see your internal handoff problem; they just feel ignored, and a warm lead goes cold.

The fix is not a bigger inbox. It is a small, repeatable triage routine that gives every reply an owner, a next step, and a deadline.

1. One daily triage pass, with an owner for every thread

Once a day, go through new replies and assign exactly one owner to each thread. No owner means no reply — that is the rule that matters most. The owner does not have to answer immediately, but they are responsible for the next step and for closing the loop. A shared list where everyone assumes someone else will handle it is how leads quietly die.

Keep the pass short: new replies, replies awaiting an answer, and threads where you promised something. Anything else can wait.

2. Deduplicate by company and domain, and suppress recent contacts

Before anyone writes, check whether the same company or domain is already in an active conversation. Duplicate outreach to different people at the same company, sent days apart, looks uncoordinated and wastes the goodwill you just earned.

The same applies to individuals: suppress people you contacted recently. A polite pause beats a second, near-identical message. Deduplication and filtering are not just tidiness — they protect your sender reputation and your relationship with the prospect.

3. Answer from your own documents, not from memory

Generic follow-ups are the second way leads go cold. "Just checking in" tells the prospect nothing. A reply that cites a real detail — the product they asked about, the pricing question they raised, the document your team already shared — shows you actually read their message.

That is why it helps to draft from your own documents and knowledge library: past proposals, product notes, FAQs, and earlier threads. Ground the reply in what you have already said and promised, so the prospect gets a consistent, specific answer instead of a fresh guess each time.

4. Send in controlled batches through your own SMTP, with campaign management

When the triage pass produces a batch of replies, do not blast them all at once. Send through your own SMTP in controlled batches, and keep campaigns organized so you can see who got what, when, and what comes next. Batching makes mistakes easier to catch, keeps volume manageable for a small team, and makes the whole process reviewable rather than a mystery.

5. One polite follow-up on a schedule, not a daily drip

If a prospect does not reply, one well-timed follow-up is usually enough. Put it on a schedule — for example, a cron-scheduled reminder a few business days later — and then stop. A daily drip trains people to ignore you; a single, useful nudge (a new detail, a short answer to their question, a clear next step) respects their time and keeps the door open.

A routine you can start this week

Morning: one triage pass, one owner per thread.
Before drafting: deduplicate by company/domain and suppress recent contacts.
While drafting: pull details from your own documents and knowledge library.
Sending: controlled batches via your own SMTP, tracked as a campaign.
Afterwards: one scheduled follow-up, then move on.

Where aiFetchly fits, in passing

aiFetchly (https://www.aifetchly.com) is an open-source, local-first desktop AI agent for business automation that can support this routine: multi-engine lead research, organizing publicly available contact details, writing personalized emails using your own documents/knowledge library as context, SMTP scheduled sending with deduplication/filtering/campaign management, cron scheduled workflows, and plugins and business knowledge. A Free Community version is available, Plus starts from $10/month, and the AI Content Assistant is Plus Only.

For a walkthrough of automated email replies grounded in local knowledge, see: https://www.youtube.com/watch?v=EQnmwuP6FBw

The tool matters less than the routine. Assign an owner, deduplicate, cite real details, send in batches, and follow up once. Do that consistently, and far fewer replies will slip through.

A local-first way to run weekly lead research and personalized email campaigns from one desktop

aiFetchly — Fri, 30 Oct 2026 23:03:27 +0000

Disclosure: I am from the aiFetchly team. This is a workflow note, not a sales pitch — the steps below work with a spreadsheet and your own email tool, too.

Most weekly outreach fails for a boring reason: the list is messy, the message is generic, and nobody remembers who was already contacted. A calmer routine fixes that before you send anything.

A weekly routine that fits on one desktop

1. Define the ICP first, in one sentence

Write down who you are trying to reach, what problem they already feel, and who is not a fit. If a prospect does not match that sentence, leave them out. A smaller, better list beats a long, cold one.

2. Research in batches, and keep only publicly available contact details

Use multiple search engines and sources, then record the company, role, and why this person is relevant. Keep a source note for each row so you can check it later. Do not buy scraped private data and do not guess personal addresses.

3. Deduplicate before you draft

Merge duplicates by company and domain, remove people you contacted recently, and filter out roles that never reply to this kind of mail. This one step protects your sender reputation more than any clever subject line.

4. Use your own documents as context

Your pricing page, case notes, and product notes are better context than a generic template. Pull one specific detail per prospect — a product line, a recent hiring post, or a public announcement — and write the first two lines around it.

5. Draft personally, then review like a human

Draft one message per segment, then read it aloud. Remove anything you would not say to that person in a first meeting. Keep the ask small: one question, one next step, and an easy way to say no.

6. Schedule with SMTP, then follow up on a schedule

Send in controlled batches through your own SMTP setup, with campaign management so you can pause, filter, and avoid sending the same person twice. Use a cron-style scheduled workflow for a single, polite follow-up — not a daily drip.

Where aiFetchly fits, in passing

aiFetchly is an open-source, local-first desktop AI agent for business automation. In this workflow it can help with multi-engine lead research, organizing publicly available contact details, writing personalized emails using your own documents and knowledge library as context, SMTP scheduled sending with deduplication, filtering and campaign management, and cron scheduled workflows, plus plugins and business knowledge. A Free Community version is available, Plus starts from $10/month, and the AI Content Assistant is Plus Only. Learn more at https://www.aifetchly.com — and this B2B leads walkthrough shows the idea in practice: https://www.youtube.com/watch?v=PvioetGLCaI

A checklist to copy

ICP sentence written and agreed
Public sources noted for each lead
Duplicates and recent contacts removed
One specific detail in each first paragraph
Batch sent, campaign paused if replies need attention
One scheduled follow-up, then stop if there is no reply

What part of your weekly routine breaks first: the list, the drafting, or the follow-up?

How Small Teams Can Triage a Shared Inbox Without Losing Leads: First Pass, Deduplication, Your Own Materials as Context, Then Scheduled Follow-Ups

aiFetchly — Thu, 29 Oct 2026 23:28:09 +0000

We are the aiFetchly team. This is not a product pitch, but a shared-inbox triage method a small team can use directly: set up the process first, then talk about tools.

First pass: turn the inbox into a to-do list, not a chat log

Set a fixed "first pass" time each day. Classify first; do not start with long replies to every message:

Needs a reply today: customer questions, clear partnership intent, and time-sensitive requests.
Needs materials checked: first note which contract, quote, or document is missing, and do not answer from memory.
Can be archived or handed off: notifications, duplicate CCs, and messages with no clear action item.
Needs follow-up later: move these to a separate follow-up list instead of leaving them in the inbox to be buried by new messages.

The goal of the first pass is not to empty the inbox, but to give every message a clear next step and an owner.

Deduplication: keep only one main thread for each matter

The reason a shared inbox loses leads is often not that nobody reads it, but that the same customer and the same issue get answered separately by several people, or discussed inconsistently across different threads.

A practical approach is:

Merge by customer and topic, rather than handling each message separately by sender;
Record the latest conclusion in the main thread, and bring later replies back to that main thread;
Organize and deduplicate repeated inquiries and repeated contact details first, then decide who will reply on behalf of the team.

aiFetchly follows the same idea: organize public contact details, and use deduplication, filtering, and activity management in scheduled SMTP sending, so the same contact does not receive duplicate messages.

Use your own materials as context: check the materials before writing a reply

Many replies go wrong not because of a bad attitude, but because policies, prices, and delivery scope for different customers get mixed up. Before writing a reply, confirm three things:

What has this customer discussed with us before?
Which document applies?
What cannot be promised and needs internal confirmation?

When using your own documents as context to write personalized messages, the key is verifiability: reflect the original meaning of the materials, and distinguish facts, inferences, and items to confirm. If the materials do not contain it, state clearly that it needs confirmation instead of inventing an answer that looks complete.

Scheduled follow-ups: turn follow-ups from "when I remember" into "reminded at the scheduled time"

Follow-up failure is often not forgetting once, but having no fixed rhythm for a second or third follow-up. You can do it this way:

For each follow-up, write down the next contact time and the trigger condition;
Use scheduled workflows for repeated reminders instead of relying on personal memory;
Keep follow-up messages short: recap the last conclusion, say what this step needs the other party to confirm, and give a clear next step.

aiFetchly supports cron scheduled workflows and scheduled SMTP sending, which can fix this kind of repeated follow-up rhythm in place; it is still recommended that key replies be reviewed by a person before sending.

A weekly checklist you can copy directly

Does every important message have an owner and a next step?
Have duplicate threads been merged into the main thread?
Were replies checked against your own materials, rather than written from impression?
Has next week's follow-up been scheduled, rather than left in the inbox?

About aiFetchly

aiFetchly's relevant capabilities include: multi-engine lead research, organizing public contact details, writing personalized messages using your own documents as context, scheduled SMTP sending (including deduplication, filtering, and activity management), cron scheduled workflows, plugins, and business knowledge. The free Community version is available, and Plus starts at $10 per month; AI Content Assistant is available only on Plus.

Related video (about automatic email replies): https://www.youtube.com/watch?v=EQnmwuP6FBw

If your team also uses a shared inbox, please share: which step is your team most likely to lose — triage, deduplication, checking materials, or follow-up?

The quiet failure of scheduled work: how to notice when automation stops working

aiFetchly — Thu, 29 Oct 2026 03:17:28 +0000

Scheduled work has a special kind of failure. It does not crash loudly. It simply stops producing results, and nobody notices, because the absence of a result looks exactly like a quiet day. A report that should arrive each morning does not arrive, and the team assumes there was nothing to report.

Three checks prevent this quiet failure. First, every scheduled run should leave a trace, even when it finds nothing. "Ran, found nothing new" is a useful result. Silence is not. Second, someone should own the check. If a run is important enough to schedule, it is important enough for a person to review its trace on a regular rhythm, not only when something goes wrong.

Third, design the workflow so that a failure in one step does not silently become a success in the next. If a source is missing, the workflow should stop and say so, rather than producing a summary based on old data. A missing input is information, and hiding it is how a small problem becomes a wrong decision.

This is also why starting small matters. One scheduled routine, with a visible trace and a named owner, is easier to trust than ten routines that nobody watches. Trust in automation is not built by how much it does. It is built by how clearly it shows what it did, and what it did not do.

How do you notice when a scheduled task has quietly stopped working?

Documents are not a filing problem. They are an answering problem

aiFetchly — Thu, 29 Oct 2026 03:11:28 +0000

Most teams do not struggle because they lack documents. They struggle because, when a question arrives, nobody can find the right document quickly, or trust that the version they found is current. The folder is full, the search returns ten similar files, and the safest answer becomes, "Let me check and get back to you."

The fix is not a stricter folder structure. It is a smaller set of rules for how documents earn their place. First, every document that answers customer questions should have an owner and a date. If nobody owns it, nobody updates it. Second, old versions should be archived, not left beside new ones with a slightly different name. A file called "final-v2-new" is a warning sign, not a solution.

Third, keep the document that the team actually uses, not the document that was approved three years ago. If the team has a better note, a better checklist, or a better template, that working document should become the source, with the old one retired. Otherwise, people will keep using the unofficial copy and pretending otherwise.

When an agent searches documents, these rules matter even more. An agent will quote whatever it finds, including the outdated version, unless the set it searches is clean. A small, current, owned set of documents beats a large archive every time, because the agent's answer is only as trustworthy as the shelf it reads from.

Which document does your team search for most often, and is it the current one?

Research without the tab explosion: a small, traceable routine for a busy team

aiFetchly — Thu, 29 Oct 2026 03:10:58 +0000

Research usually fails in the same way. You open twenty tabs, you find a few useful facts, and by the end of the day you cannot remember which fact came from where. The result is not more knowledge. It is more doubt, because nobody wants to act on a claim they cannot check.

A smaller routine works better. Start with one question, written down in a single sentence. Then collect only sources that answer that question. For each source, record three things: where it came from, when you found it, and exactly what it says. Do not copy a summary of a summary. If a source is weak, say so in the note, instead of letting it quietly become a fact later.

Next, separate facts from conclusions. A price, a date, and a quoted rule are facts. What you should do about them is a conclusion, and it deserves its own line. When facts and conclusions are mixed, a team argues about the source when it should be arguing about the decision. When they are separate, both conversations become shorter.

Finally, close the loop. Write a one-paragraph answer to the original question, with the sources listed underneath. If the answer cannot be written, the research is not finished, and that is useful information too. It tells you which question to ask next, instead of sending you back to twenty open tabs.

What is the one research habit that has saved your team the most time?

Before you automate B2B outreach, automate the boring part: finding leads you can actually trust

aiFetchly — Thu, 29 Oct 2026 02:45:21 +0000

Most B2B outreach fails before a message is ever written. The list is stale, the company no longer matches your offer, or the contact has moved on. Automating the sending step only makes a bad list fail faster.

A better order is: verify first, personalize second, send last. Here is a workflow a small team can run without a large sales stack:

Define a fit filter in plain language. For example: industry, company size range, location, and one clear signal that the business needs what you sell. If a lead does not meet the filter, it does not enter the outreach list, no matter how easy it was to find.
Collect from public sources and keep the source. Record where each detail came from and when it was found. A lead with a dated source is easier to re-check than a lead copied into a spreadsheet with no history.
Score for outreach readiness, not just size. A smaller company with a current hiring post, a new office, or a recent product launch is often a better first contact than a large company with no recent signal.
Write one specific opening line per segment, not per lead. Reference the segment's real problem, then let a human adjust the line for the individual company before sending. That keeps volume up without turning every message into a generic template.
Stop and review after the first batch. Check reply quality, bounce reasons, and wrong-fit replies. Fix the filter before you scale the list.

The goal is not more leads. It is fewer wasted conversations and a list you can defend when a prospect asks, "Why did you contact us?"

Disclosure: I am posting as part of the aiFetchly team. aiFetchly is an open-source, local-first desktop AI agent for business automation, including research, document handling, customer communication, scheduled workflows, plugins, and business knowledge. A short demo of automatic B2B lead extraction is here: https://www.youtube.com/watch?v=PvioetGLCaI

What is the one fit signal you trust most before you contact a new B2B lead?

A practical way to automate customer email replies without sending your business knowledge to the cloud

aiFetchly — Thu, 29 Oct 2026 02:42:35 +0000

Small teams often want the same thing from email automation: faster replies, fewer missed messages, and answers that actually match how the business works. The usual shortcut is to paste everything into a cloud tool and hope the model remembers it. That creates a second problem: your prices, policies, and customer history now live somewhere you do not control.

A local-first workflow avoids that trade-off. Here is a simple version that works well for a small business:

Keep one source of truth on the desktop. Store the current price list, return policy, service notes, and a few approved reply templates in one folder. When something changes, update that folder first.
Let the agent draft, not send blindly. The agent reads the incoming message, finds the matching document, and drafts a reply based on that document. A person reviews anything that promises a price, a refund, or a deadline before it goes out.
Log what was used. For each reply, note which document the agent relied on. If a reply is wrong, you can fix the document once instead of arguing with the model again.
Start with one inbox and one message type. Common questions about opening hours, delivery, or appointment changes are a good first automation. Leave unusual or sensitive cases for a human.

This approach is less impressive in a demo than a fully autonomous inbox, but it is easier to trust in production. The business keeps its knowledge local, the agent stays grounded in real documents, and customers still get a fast, consistent answer.

Disclosure: I am posting as part of the aiFetchly team. aiFetchly is an open-source, local-first desktop AI agent for business automation, including research, document handling, customer communication, scheduled workflows, plugins, and business knowledge. A short demo of automatic email replies using local knowledge is here: https://www.youtube.com/watch?v=EQnmwuP6FBw

What is your rule for deciding which messages an agent may answer on its own, and which ones must wait for a human?