How to Use AI for a Resume Without Installing Anything

The free resume agents are worth using and worth understanding. Here is what installing one actually involves, where all of them stop, and how to cover the gap with nothing but what you already have open.

By Zaira

01The Problem

Free and good are both true. So is four commands.

There is a real, well-built open-source layer here, and it has a setup cost that nobody mentions, a hard ceiling on what it can reach, and one failure mode you cannot see.

The free tools are good, and they are also four commands

The open-source resume agents are real projects with real engineering in them — a GitHub install, an executable bit, two Python packages, knowing where your agent keeps its skills directory, and a restart. Some need a container runtime. One needs Node and a package manager and downloads a browser to render output. None of that is unreasonable for a developer. It is simply a precondition, and the difference between "paste a prompt" and "set up a local toolchain" is the whole of what this page is about.

They run once, on a document you paste, with nothing behind them

A local skill has no job board, no company research, and no memory. It reads what you hand it, produces an output, and exits. The searches, the second posting, the check on whether the role is still open next week, and the follow-up in nine days — every one of those is still a manual step you were already doing. That is not a flaw in the tools; it is what a document-in, document-out tool is. It just does not replace the part of the job that takes the longest.

The most common failure is silent

An agent loads a skill file only when your request matches the description it declares. If your phrasing misses — and phrasing is doing real work when you are describing a role you do not have words for yet — the agent simply answers without it, in the same tone, with no error and no indication that anything was skipped. People conclude the tool does not work. More often it was never invoked, and the way to tell is to ask the agent to state what it can do before you rely on it.

02The How-To

The install-free method, step by step

Copy it into Zaira. It starts by telling you what it can and cannot do — and what to paste instead for each thing it cannot — then reads two postings properly and writes only the half you can evidence.

Resume research with no local install
Automate it with Zaira

Step 0: Confirm the boundary, then work around it

Act as my resume research assistant. I have nothing installed, I am not installing anything, and I am not connecting anything to anything. Work only from what I give you and what is already in this conversation. Tell me plainly what you can and cannot do for this task: - read a job posting I give you as a link? [yes / no] - read a job posting I paste as text? [yes / no] - find new postings on a job board by itself? [yes / no] - run again tomorrow without me? [yes / no] - remember my last answer so I do not repeat it? [yes / no] Do not guess. Then for every "no", immediately follow it with the paste-in substitute: exactly what I put in my next message, and what I carry over from this one. Name no tool, package, or setup step — a pasteable substitute is the only answer I can use.

Step 1: Take the role, not the job

- The role I am targeting: [e.g. revenue operations analyst] - Where I want to work: [remote / city / country] - What I am actually good at, in plain words: [list, no adjectives] - What I have never done: [list — this matters as much as the first one] - My last three roles, roughly: [one line each]

Step 2: Widen the search before you narrow it

Build a plan of 12-18 queries across these angles: 1. the exact title I typed 2. adjacent titles doing the same work 3. the title as that employer would write it internally 4. skill-based queries with no title at all 5. sector-specific queries Show me the table first: query | angle | why. For each query also write the exact string I would paste into a search box, and one sentence on what result would make it worth my time. Then work from what I have already given you. For any query you cannot run yourself, do not guess at what it would return — leave it as a string for me to paste back in, and tell me how many results I should bring back so the next pass has enough to work with.

Step 3: Read two postings properly

Take the two best matches I have given you. For each, tell me: - the eight terms the posting actually weighs, and the sentence each comes from - which of those I can already evidence from Step 1 - which I cannot evidence, and therefore must not have written Show me the gap before you write anything.

Step 4: Write only the evidenced half

Draft the section using only what Step 3 marked as evidenced. Keep my Step 0 wording wherever you can. Do not add a number I did not give you. Do not add a tool I have not used. Collect every unevidenced term from Step 3 into a separate list headed "not claimed" rather than leaving it silently absent.

Step 5: Make the second run cheap

Nothing runs again on its own, so design tomorrow's run now, while you still have everything in front of you. Write me a carry-over block I can save and paste back: - my Step 1 answers, in the form I gave them, ready to re-use verbatim - the Step 2 search strings, formatted so I can paste them again unchanged - yesterday's "not claimed" list - today's date, and the day count since I started Then tell me exactly what I re-send, what I can skip because it has not changed, and what single question I add so the second run tells me what is new instead of repeating what I already know.

Output rules

Never present a posting I did not give you or that is already in this conversation. Never name a tool, package, extension, or setup step — if a capability is missing, the answer is a paste-in substitute, not an installation. If you are out of material to work from, stop and tell me what to bring you next.
03Why People're Using

Know the ceiling before you commit to it

Three outcomes, one loop. You find out what an agent can reach with nothing installed in twenty seconds instead of an hour, you get the terms a posting actually weighs instead of a guess, and you ship the evidenced half rather than a document that falls apart in the interview.

The boundary, and the substitute for each part of it

Step 0 answers all five, then writes a paste-in substitute for every "no" — the exact next message and what to carry forward — and is forbidden from naming a tool or a setup step. That single rule is why the ceiling on this page is a method instead of a dead end.

Terms with their source sentence attached

Step 2 hands you each query as a string to paste into a search box, plus how many results to bring back so the next pass has enough to work with. It is the query ladder doing the work a tool would otherwise do, and you end up holding a reusable search list instead of a one-off result.

A carry-over block that makes tomorrow's run cheap

Step 5 writes the save-and-paste block while everything is still in front of it: your answers, the search strings, yesterday's "not claimed" list, and the date. That block is the manual equivalent of a scheduled run — it costs you one paste and tells the second pass what changed instead of repeating what you already know.

04FAQ

Frequently asked questions

The practical questions people ask before deciding whether the free tools are worth their setup.

Yes, and the ceiling is lower than the marketing suggests and higher than the sceptics claim. What you need is a place to paste a prompt, and every part of the job that is text-in-text-out works there: the register work, the interview, the tells audit, the rewrite, the gap analysis. What it cannot do by itself is reach anything — find the fifteenth posting, open a link, or start again tomorrow. None of those require software. They require you to paste the result of one step into the next, and Step 0 of this method has the agent write that bridge for every capability it does not have, as an exact string you can paste. That is the whole difference between an agent that stops at your first message and one that keeps going across a search.

More than their readme implies. In practice: clone the repository, make the installer executable, install its runtime dependencies, know which directory your agent reads skills from, and restart the agent so it re-scans. Several need a Python toolchain; some projects require a container runtime; the ones that output a PDF generally need a document library, and one goes as far as pulling a headless browser at install time. All reasonable for a developer, and all of it a prerequisite the free tier does not remove. This page's position is not that they are bad — several are excellent — it is that you should know the cost before you start.

Because the trigger is a description match, not a command you ran. An agent loads a skill when what you asked for resembles the summary that skill declares, and the similarity check is a real gate rather than a formality. If your phrasing misses — and phrasing is hard when you are describing a role you do not yet have vocabulary for — the agent answers the question without it, in the same confident tone, and never says a skill existed. The reliable fix is to stop inferring from the output and just ask: state what you can do here, and name what you would need connected. That question exposes the silent skip every time.

You would be faster, not more correct. Without one, the method still does the register work, the interview, the six-question extraction, the tells audit and the rewrite — most of the value, all of it from pasted text. What you give up is reach: finding the fifteenth posting, opening a link rather than a copy, and running again without re-pasting. Every one of those has a paste-in substitute, and Step 0 writes them out as exact strings rather than describing them. The honest cost is repetition. That work was never going to be done by a tool for you — the open-source skills are local document tools, and a paid connection would only have moved the repetition into a setup step rather than removed it.