The right AI for the job. No hype.
We use AI on every Stacklumen engagement: code review, content drafts, research, internal docs. We also keep it away from plenty. No single tool wins everywhere, so here is the kind of tool we reach for, task by task, and why.
Five rules we don’t break. Everything below assumes them.
Not a legal disclaimer, just what has held up across our own AI-assisted work.
- 01
Verify before you ship.
Treat AI output as a confident-sounding first draft, never a finished answer. Names, numbers, citations, legal language — all hallucinate-prone. Read every word before it leaves your hands.
- 02
Never paste anything you wouldn’t email.
Anything you put in a public AI chat may be used for training and is stored on someone else’s servers. Client data, internal financials, customer info, anything under NDA — keep it out, or use an enterprise plan with a zero-retention agreement.
- 03
Disclose when it matters.
Marketing copy and brainstorm output — fine. Anything where authorship matters (proposals, expert articles, client deliverables) — say AI helped. It’s usually a non-issue if you mention it; it’s a real issue if you hide it and someone notices.
- 04
Keep the human on the hook.
A person should still own every decision the AI helped with. The model can draft a refund policy; a human owns the consequence of shipping it. "The AI told me to" is not a defense.
- 05
Bias is real and quiet.
Every model is trained on a slice of the internet. That slice carries its biases — about who counts as default, what English sounds professional, whose problems get treated as universal. Watch for it.
Pick a task. Get the ranked pick.
For each task, the kind of tool we reach for first and what to look for in it, then the alternatives and when they are the better fit. Brand allegiance isn’t a strategy.
Long-form writing + reasoning: the pick and 2 alternatives
Long-form writing + reasoning
Briefs, internal docs, policy drafts, anything where you need a model that holds onto a lot of context and reasons through it cleanly.
- 01
A frontier assistant with a long context window Pick first
It has to hold a long document without losing the thread and write in a voice that doesn’t sound like every other AI. Try two or three on your own material: the right one follows nuanced style instructions and keeps following them.
- 02
A fast general-purpose chat assistant
Quick, casual drafts where tone doesn’t matter much. A free tier covers a lot of everyday writing.
- 03
An assistant with live web search
When you need to pull in current search results mid-draft.
Writing or debugging code
Pair-programming, refactoring, generating boilerplate, debugging a tricky stack trace.
- 01
A coding agent that works in your repository Pick first
An agent that reads the codebase, runs commands and edits files directly does real refactors, not snippets. Judge it on your own code, not on a leaderboard. If it supports MCP, it can work on your Terminal sites too.
- 02
Inline completion in your editor
Suggestions while you type. Less context-aware than an agent, but lower friction.
- 03
A general chat assistant
Quick standalone code questions when you’re already there for something else.
Customer support + drafted replies
First-draft responses to inbound messages, FAQ writing, policy explanations.
- 01
A careful assistant, grounded in your real policies Pick first
Safe defaults around tone and edge cases matter more than speed here. Give it your actual policies so it is less likely to confidently invent a refund policy you don’t have.
- 02
A fast general chat assistant
Lower-stakes, high-volume support where a quick draft matters more than a careful one.
Quick everyday questions
Definitions, summaries, "what does this acronym mean", brainstorming a list of ideas.
- 01
Whichever free chat assistant is already open Pick first
Free, fast, ubiquitous. For low-stakes one-shot questions, the easiest tool wins.
- 02
A long-context assistant
When your "quick question" actually involves a long document you want it to read.
- 03
An answer engine that cites the web
When the answer needs a citation from the live web.
Image generation
Logos, illustrations, marketing visuals, mockups. Stacklumen does not ship AI-generated images for clients — we hire illustrators or commission photography — but here’s where the tools stand if you’re generating your own.
- 01
A dedicated image model Pick first
The dedicated image tools have the highest aesthetic ceiling of anything we’ve tested, with a learning curve on prompts. Run the same brief through two or three before you commit to one.
- 02
Image generation inside your chat assistant
You’re already in the assistant and want something passable in one prompt with no setup.
- 03
An open model run locally
You want local, private, free generation and you’re comfortable with technical tooling.
Research with citations
Anything where the answer needs to point at a source — competitive research, market sizing, finding case studies.
- 01
A search-native answer engine Pick first
Search-native. Every claim cites the source so you can verify.
- 02
Your chat assistant with browsing
Already in your subscription — works when you don’t want to spin up another tab.
- 03
A long-context assistant with web search
When the research is also a long-form synthesis and you want one tool to do both.
Voice + real-time conversation
Hands-free brainstorming, language practice, having something explained while you cook.
- 01
An assistant with a real-time voice mode Pick first
Look for low conversational latency, natural voices and support for interruption. The closest thing to a real conversation.
- 02
A voice mode that can use your camera
When you want the model to see what your camera sees while you talk.
Private + local
Anything that involves sensitive client data, internal financials, or anything you wouldn’t paste into a public web form.
- 01
An open-weight model running on your own machine Pick first
Free, open-source tooling that runs on your laptop. Nothing leaves your machine.
- 02
A frontier model API on a zero-retention enterprise plan
You want frontier quality AND data privacy guarantees — pay for the enterprise tier.
Blog + long-form content
SEO-targeted articles, thought-leadership posts, case studies. The places where 1,500+ words have to read like a human wrote them.
- 01
A long-context assistant, working from your brief Pick first
The strongest models hold structure and transitions without collapsing into AI-voice at length. Pair one with a brief + outline you write, not a one-line prompt.
- 02
A template-driven AI writing tool
You want blog flows with SEO scoring built in instead of doing the SEO check separately.
- 03
The AI inside your docs workspace
The article lives there and you want generation inline with your other docs.
Short-form social (LinkedIn, X, threads)
Posts, threads, hooks, repurposing long-form into bite-size formats. Tone matters more than depth here.
- 01
Your assistant with a voice prompt Pick first
Feed it five of your best past posts as the voice anchor, then ask it to write in that voice. Avoids the LinkedIn-default tone every model defaults to.
- 02
A LinkedIn-native writing tool
Built around hooks, post scheduling, and pulling proven templates.
- 03
A thread drafting and scheduling tool
X threads with a clean drafting UX and built-in scheduling.
YouTube — scripts
Video scripts, intro hooks, B-roll callouts, end-screen CTAs. The hard part is the first 15 seconds; the rest is structure.
- 01
A long-context assistant with a hook framework Pick first
Holds a consistent narrator voice across a 6-10 minute script. Drop in your channel’s top three intros as examples and it patterns them well.
- 02
A fast chat assistant
You want quick variation on a script you already have — generate 5 different intros, pick one.
YouTube — titles, descriptions, tags
Where YouTube SEO actually lives. Title is the click; description is the rank.
- 01
A YouTube keyword tool Pick first
Plugged into YouTube’s actual search data, so its suggestions are grounded in real keyword volume, not guesses.
- 02
Another tool in the same category
Pick whichever subscription you already have.
- 03
Your assistant
You’ve done the keyword research and just need 10 title variations from a known seed.
YouTube — voiceover + narration
You wrote the script and don’t want to be on camera. Or you want consistent narration across a long series.
- 01
A voice-cloning text-to-speech tool Pick first
Voice quality is the whole job. The best tools clone your own voice from a few minutes of clean audio, or offer a library of natural-sounding presets.
- 02
A simpler text-to-speech tool
Lower-volume use or a friendlier UI. Less natural ceiling but easier learning curve.
- 03
Voiceover inside your video editor
You’re already editing there and want voiceover inside the same tool.
YouTube — editing + clip extraction
Cutting a long-form video into short clips, removing filler words, generating captions, syncing B-roll.
- 01
A transcript-based video editor Pick first
Edits video by editing the transcript. Cut a word, the video cuts. The single highest-leverage tool for a small content team.
- 02
An automatic clip extractor
Pure clip extraction — feed it a podcast or webinar, get short-form clips back.
- 03
An animated-caption tool
Animated captions + subtitle overlays for shorts and reels.
Thumbnails + ad creative
YouTube thumbnails, ad creative variations, social graphics. The work where one good visual outperforms ten okay ones.
- 01
An image model, then your design tool Pick first
An image model for the base imagery, then composite text + brand layers in your design tool. AI generates the hero asset; a human ships it.
- 02
An online design tool’s templates
Lower-stakes thumbnails where speed matters more than uniqueness — built-in text + templates.
- 03
An ad-variant generator
You’re running paid ads and need 20 creative variants for testing, fast.
SEO — content briefs + optimization
Keyword research, competitor analysis, content briefs that hit the structure search engines reward.
- 01
A SERP-based content brief tool Pick first
Reverse-engineers ranking pages into briefs you can hand to a writer (or your assistant). Scoring is grounded in actual SERP data.
- 02
An enterprise content-optimization tool
The higher-end option for content teams at scale. Strong topic modeling.
- 03
A lighter brief tool
A cheaper option for solo creators or small teams.
Email campaigns + newsletters
Newsletter drafts, drip sequences, cold outreach, transactional copy.
- 01
Your assistant for drafts, your email platform for delivery Pick first
Generative quality matters more than the sending platform. Draft in your assistant, paste into your ESP. Don’t pay extra for "AI inside an email tool" when standalone is better.
- 02
A cold-email coach
Specifically for cold sales email — scores subject lines and bodies against open/reply rates.
Customer research + interviews
Synthesizing customer interviews, finding patterns across support tickets, summarizing survey free-text.
- 01
A long-context assistant (with the raw transcripts) Pick first
Take the largest context window you can get. Paste 50 pages of interviews, ask for themes, get them.
- 02
A research repository
You’re doing this regularly and want tagging and search across every study.
The wider stack, by category.
The finder covers tasks where AI is the work. These are the tools where AI sits inside an existing job and quietly multiplies it. Reach for them as you hit the problem they solve, not all at once.
Workflow + automation
The connective tissue. Wire AI into the rest of your stack so it triggers automatically, not manually.
A no-code automation platform
The default. Thousands of integrations, AI steps built in, dead simple no-code triggers.
A visual scenario builder
More power, more complexity. Visual scenarios for branching logic a simple trigger chain can’t do cleanly.
A self-hosted workflow engine
Open source. Run it on your own server, pay nothing in API fees.
A code-first workflow runner
For when you want to write actual code in the workflow steps.
Meeting + voice intelligence
AI-generated meeting notes, action items, follow-up emails. Highest-ROI category for most small teams.
A local-first note taker
What we look for first. Never auto-joins a call, and writes notes you actually use.
A meeting recorder with a free tier
A genuinely usable free tier. Strong on summaries and CRM integrations.
A team meeting-analytics tool
Better for larger teams that want sentiment + speaker analytics on top of notes.
A dedicated transcription service
When transcript quality matters more than the notes built on it.
CRM + sales
AI-augmented pipelines, lead scoring, draft replies, deal summaries.
A CRM with AI built in
Auto-enriches contacts, drafts replies, builds reports.
The AI suite in the CRM you already use
Already on a CRM with one? Use it. Don’t move CRMs just for the AI.
A programmatic enrichment tool
Lead enrichment + outbound. Stacks many data sources behind one workflow.
Customer support
AI-first ticket triage, knowledge-base answers, draft replies for human review.
An AI support agent
Resolves a real percentage of tickets end-to-end without a human, once it has your help centre to work from.
The AI in your help desk
Enterprise-tier support. Plays nice with the workflows you already run.
A help desk built for B2B
For dev-tool / B2B support. AI built in, made for small teams.
Notes, docs, knowledge
AI-aware writing surfaces. Where drafts live before they ship anywhere.
The AI in your docs workspace
"AI Q&A across your entire workspace." Strong for distributed teams.
A doc + database hybrid
Stronger structured-data AI workflows than a plain docs tool.
AI-native notes
Auto-tags, auto-links, auto-surfaces related notes as you write.
Project management + scheduling
Where work gets tracked + when it happens. AI is mostly a feature here, not a category.
A fast issue tracker
Issues, projects, cycles — everything fast. AI features are a bonus, not the reason.
An AI-scheduled calendar
Tell it your tasks, it auto-blocks time and reshuffles when life happens.
A calendar habit planner
The lighter option. Strong calendar habits without the full task layer.
Developer tooling
IDE + CLI tools where AI lives next to the code.
An AI-native code editor
Tab-complete that understands your whole repo, plus agentic edits.
A command-line coding agent
Runs commands + edits files directly. The kind of tool building Stacklumen, mostly.
Inline completion in your IDE
Still solid if you want IDE-only AI without the agentic surface.
A prompt-to-UI generator
Great for prototyping a page in 30 seconds, less great for production code.
Whichever agent you choose, it connects the same way.
Terminal and Baselumen each run an MCP server. Any agent that supports MCP works with them on your own plan, with no AI key to add, in three steps.
- 1. Copy the server URL.
- Terminal: www.stacklumen.com/terminal/api/mcp. Baselumen: www.baselumen.com/app/api/mcp.
- 2. Add it to your agent as a remote MCP server or connector.
- Wherever your agent keeps them. Any client that supports remote MCP servers works.
- 3. Sign in with your Stacklumen account.
- Your agent asks when it first connects. Allow access, and what it changes shows up in the app as it happens.
- An agent that cannot sign in uses a token.
- Make one in Terminal’s Agents & API or Baselumen’s Agents screen and send it in a bearer Authorization header.
Choose the model by the job, not the brand: long context for reasoning, a voice you can steer, code it can run and check, and defaults that ask before they guess. The connection stays the same.
We carry the subscriptions, so you don’t have to.
Stacklumen runs on the stack this guide recommends. Every engagement, whether a site built in Terminal, a custom application or a retainer, uses these tools.
- The subscriptions, already paid for.
- Frontier assistants and their APIs, image and voice models, a research engine, a video editor, an SEO brief tool, a coding editor, team automation, meeting notes, a CRM and an issue tracker. The seat fees sit behind every Stacklumen engagement: you get the output, we carry the cost.
- The integrations, already wired.
- Auth, billing, support triage, content pipelines, lead enrichment — we’ve already plumbed these workflows on our own platforms. When we drop them into a client engagement, we’re moving working patterns, not figuring it out for the first time on your time.
- The judgment, already calibrated.
- Which tool wins for which job changes month over month. We test new releases the week they ship and update what we recommend. You don’t have to track which model leads at refactors this quarter or which voice-clone tool just got better — we already did.
- Built for small teams scaling up.
- Small and mid-sized businesses can’t run the evaluation work a large company staffs for. We do it as part of the engagement and bring the tooling with us, so you start from what already works.
The tooling and the judgment come with the engagement, and you keep full ownership of the code.
Where we don’t use AI. A short list, by design.
Some work is faster and better when a person owns it from the start. This is what we keep off the AI’s plate at Stacklumen.
Naming
Brand names, product names, headline copy that has to land in one shot. Models default toward safe and forgettable. We brainstorm with one in the room, then humans pick.
Final brand voice
A trained-on-the-internet model produces internet voice. To sound like you, it needs a corpus of you. We use AI to draft fast, then rewrite in the actual voice.
Client image generation
Stacklumen ships hand-photography or commissioned illustration on client work, not AI imagery. The tools are good and getting better — they still have a tell, and they raise legal questions we don’t want to inherit.
High-stakes legal, financial, or medical language
Anything that creates legal exposure if it’s wrong. The right answer there is a human professional, not a confident-sounding sentence.
Replying as a person
If something has someone’s name on it, that person wrote it. AI drafts in a doc are fine; AI sending email as you is not.
What we’re building next. Nothing here ships yet.
An AI co-builder we are scoping, trained on the patterns behind our client work and the Baselumen library. None of it has a date. Each row says how concrete it is: planned means we know how we would build it, concept means the approach is still open.
Picking tools is the easy part. Wiring them in is the work.
Fitting AI into a small team’s workflow without breaking the parts that already work is what we do. Tell us where the time goes.