LinkedIn maxxing with Codex: 12 prompts that turn a project folder into your content system

LinkedIn maxxing with Codex means treating OpenAI's Codex agent as the working folder for your LinkedIn content: it keeps your business context, research and drafts as files it can reopen in any thread, and the 12 prompts below take you from a voice guide to a monthly report. Codex does the reading, counting and drafting, and you keep the judgment and the publish button.

Handing an AI a topic and asking for a post skips the work that makes a post worth reading: knowing your customers, finding an angle, gathering evidence and knowing what you can credibly claim. These prompts turn that work into a process you can repeat, from market research to a content calendar to tracking which posts led to real business conversations.

By the end you will have:

• A voice guide built from the way you actually write and speak.
• A research routine that surfaces ideas before the calendar runs empty.
• A drafting and review loop with explicit quality standards.
• A monthly report that tells you what to write next.

Codex suits this job because it works on files in a project folder. Your context, research and drafts stay on disk as documents, so nothing vanishes into an old chat. Bookmark this page for your next planning session.

What context should Codex have before you run any prompt?

A job title and "make it sound human" are not enough. Codex needs to know who you sell to, what you believe and which claims you can back up.

Start with this:

Treat the details below as the standing brief for all my LinkedIn content. Store them in a reference file I can reuse and tell me its location. Company: [name, website, what we sell, usual length of a sales cycle]. Readers: [job roles, company sizes, industries, regions]. What troubles them: [problems]. What makes them hesitate: [objections]. What makes them start looking for help: [triggers]. How we describe ourselves: we help [audience] reach [outcome] by [approach]; what sets us apart is [differences]; we are the wrong choice for [disqualifiers]. My opinions: three things I believe about my industry are [opinions], and I disagree with [common advice] because [reason]. Proof I'm allowed to use: [case studies, results, examples, personal experience]. Never make up customers, revenue numbers, stories or quotes. How I sound: [10-20 posts] plus [transcripts or writing samples]; stay away from [phrases, tones, formatting habits]. What I want from LinkedIn: my main goal is [qualified conversations / deal support / hiring / awareness], my offer is [offer], and the next step I want an interested reader to take is [action]. House rules: only use connected sources I have access to and permission for; if you can't reach a source, ask me for an export or a paste. Publishing, comments, invitations and messages on LinkedIn stay with me. Put drafts in [review location], keep evidence and inference apart, and flag anything that needs my input.

Update this brief whenever your offer, audience or opinions shift.

Where does Codex keep that context?

Give LinkedIn a folder of its own, such as linkedin/. Open it as a project in the desktop app, or run codex from inside it in the terminal. Ask Codex to save its answer to the prompt above as AGENTS.md in the root of that folder. Codex reads AGENTS.md automatically at the start of each session, so every later prompt here runs with your context already loaded. If you would like Codex to create the file skeleton, run /init first.

Keep AGENTS.md brief and have it link to longer files such as context/business.md, context/evidence.md and voice-guide.md. A working version can be as short as this:

# LinkedIn workspace
Read context/business.md and voice-guide.md before any draft.
Evidence lives in context/evidence.md.
Never invent customers, numbers or quotes.
Drafts go to drafts/{week}/, reviews to review/, reports to reports/.
I publish, comment, connect and message on LinkedIn myself.
Mark every claim as evidence or inference.

Here are the files this guide creates and where Codex expects them:

~/.codex/AGENTS.md                 personal rules for every project
~/.codex/config.toml               MCP servers, web search, memory
~/.codex/review.config.toml        read-only profile for prompt 8
~/.agents/skills/linkedin-draft/   reusable drafting skill
linkedin/AGENTS.md                 this project's context
linkedin/.codex/config.toml        project-only settings

Codex also has a memory feature that can carry preferences and corrections between threads; in the CLI, /memories shows or resets it. It is still rolling out and may be switched off on your account, so count it as a bonus. Anything Codex must follow every time belongs in AGENTS.md.

Part 1: research and writing (prompts 1-8)

Prompt 1: how do you build a voice guide with Codex?

A voice guide earns its keep when it explains the decisions behind your writing: how you set up an argument, where examples go and how you move from one idea to the next. Labels like "conversational and authoritative" don't help anyone write.

Read the posts and transcripts I've given you and turn them into a working LinkedIn voice guide. Describe how I start a post, how I explain an idea and how I wrap up an argument. Record my typical words, sentence and paragraph lengths, and how often I use lists. Show how I sound when I disagree, when I'm unsure, when I joke and when I'm confident. List phrases that come naturally to me, quoting them exactly with the source. Flag habits that only show up in my weaker pieces, so they don't become rules. Add a list of things to avoid: needless repetition, invented drama, certainty I can't support and too many one-line paragraphs. End with three before-and-after edits that move generic copy into my voice without changing what it says. Make the guide concrete enough that a different writer could follow it.

In Codex: put your strongest posts and a few transcripts in a samples/ folder and reference them with @ instead of pasting, so Codex reads whole files instead of a cut-off paste. The desktop app can open PDFs and documents in the sidebar preview, which lets you check what it is quoting. Have it write the result to voice-guide.md, then add one line to AGENTS.md: "Read voice-guide.md before drafting anything." When you notice you have corrected the same phrase twice, ask Codex to add a rule to that file instead of trusting memory to keep it.

What you gain: feedback you can reuse. You fix the instructions once and every future draft starts from them, so you stop correcting the same wording every week.

Prompt 2: how do you find the customer questions worth answering?

Sales calls and customer conversations contain questions no keyword tool will show you. The useful ones usually sit around a decision someone is finding hard. Codex handles this well because it can count across a folder of files instead of skimming one pasted chat.

Go through the sales calls, support threads, customer interviews and lost-deal notes from [period]. Sort the questions that keep coming up into five groups: problems people are trying to make sense of, worries that slow a decision, alternatives they weigh us against, expectations that end in disappointment, and the questions they ask just before deciding. For every group, give an anonymised quote, where it came from, how many separate conversations mention it, the person's role and how far along they were. Propose two LinkedIn angles per group, one that helps a reader name their problem and one that helps them judge a fix. Rank the ten strongest by how often they recur, how much they matter commercially and how good our evidence is. Keep what customers said separate from what you think it means.

In Codex: export call notes into a calls/ folder, one file per conversation, so Codex can actually check the "how many separate conversations" figure. If the notes sit in Gmail, Google Drive, Slack or Notion, a Codex plugin for that app can bring them in; in the desktop app, plugins live under a single Plugins entry in the composer's plus menu. Codex edits your files locally while the model runs in the cloud, so remove any names you aren't allowed to share before the run. Save the ranked list as backlog/customer-questions.md.

What you gain: a backlog of topics tied to conversations your company is already having, phrased in words customers use, with evidence you can check.

Prompt 3: how do you spot topics moving from X into your market?

An argument picking up speed on X can make a good LinkedIn post, as long as you explain why it matters to your customers and add something beyond a rewrite.

Search the live web, any sources I've connected and public pages you can reach for the past [seven days] of conversation about [topics] and from these accounts: [accounts]. Pick out up to 10 developments or debates that matter to [reader]. For each one give the original source, its author, the date and the link; a neutral summary of what happened or what is being argued; the best competing readings of it; what it changes for our audience day to day; and an angle we could back with our own experience or evidence. Treat like and comment counts as a snapshot, don't take popularity as proof, and check factual claims against the original source where you can. Recommend the three ideas most worth developing and tell me what I'd have to add before each could become a credible LinkedIn post.

In Codex: there is no built-in X feed, so this prompt relies on web search. Search is on by default in a cached mode that reads a pre-built index, which lags too far behind for a seven-day window. Start the session with codex --search, or set live search as the default for this folder in the project's .codex/config.toml:

web_search = "live"

If you already have links or an export of the posts you follow, put them in research/x-links.md and tell Codex to start there. In the desktop app, open each returned source in the built-in browser before you trust the summary.

What you gain: research gets you started, and your judgment gives readers a reason to follow you. A good adaptation ties the debate to a decision your readers really face.

Prompt 4: how do you analyse posts that beat their authors' usual results?

Holding every post up against a celebrity's biggest hit skews your research. Compare each post with its own author's typical numbers first, then look for what explains the gap. Codex can do the maths in code, which is more reliable than asking a model to estimate medians in prose.

Use this LinkedIn swipe file and whatever performance data comes with it: [file]. Where there is enough comparable data, set each post's reactions and comments against the median for the same author, format and period. Show your working, and flag small samples, missing numbers and posts that are too new to judge. For the clearest outliers, describe the topic and intended reader; the opening, argument, evidence and close; the format and rough length; whether the comments suggest real customer interest; and anything else that could explain the result, such as a launch, a news story or unusually wide distribution. Turn this into five reusable structures, then suggest an original idea for each based on what I know. Don't copy distinctive phrasing, and don't present an observed pattern as a rule of the algorithm.

In Codex: store the swipe file as a CSV in the project and ask Codex to work out the medians with a short script it runs itself, then show you the script beside the table. "Show your working" then becomes something you can rerun next month. To check a single post against its author's usual numbers by hand, use our LinkedIn engagement rate calculator.

What you gain: a better set of ideas to test, without assuming that copying a hit post's layout will copy its results.

Prompt 5: how do you turn recorded know-how into original posts?

A transcript can hide a strong post inside a story, a throwaway explanation or a disagreement that deserved more airtime. Long recordings are where a file-based agent beats a chat window: the full transcript stays on disk and can be cited line by line.

Read this transcript from an interview, podcast, webinar or meeting: [source]. Find eight moments that could carry an original LinkedIn post, favouring decisions, mistakes, surprising observations, practical explanations and opinions grounded in experience. For each, give the exact quote with its timestamp or location; the main idea and who it would help; any context needed so the speaker isn't misrepresented; facts we already have that support it; and one follow-up question that would sharpen it. Choose the three best moments and write a draft from each in my voice, following my voice guide. Keep the speaker's level of certainty, strip out anything confidential, and put placeholders wherever a result or example still needs confirming.

In Codex: give each transcript its own thread. The desktop app runs parallel threads in one project, so you can mine three recordings at once without quotes getting mixed up, and its summary pane lists the sources and outputs of each thread, which shows where every quote came from. Ask for the results in transcripts/{episode}-moments.md with timestamps, so the drafts in prompt 7 can cite the exact line.

What you gain: posts built from experience you already have, without someone having to invent a new opinion every time the calendar needs a slot filled.

Prompt 6: how do you build a four-week LinkedIn content calendar?

A good calendar weighs what your customers need to understand against what you have the evidence to say. Filling every slot with another version of your offer makes the feed repetitive. Codex can check the plan against the backlog files it has already written, so each slot traces back to research.

Plan four weeks of LinkedIn posts at [three a week], using only ideas we approved during research. Mix in practical explanations that help readers solve a problem, opinions backed by experience or evidence, customer stories we have permission to share, answers to objections and decision questions, and the occasional offer post with a relevant next step. For every post, name the reader, the one goal, the main argument, the source material, a suggested format and the action you want the reader to take. Look for repeated topics, overlapping arguments, claims without support and too many sales posts. Write full briefs for week one and shorter outlines for weeks two to four. Keep one open slot a week for something timely, since the plan shouldn't be fixed in stone.

In Codex: ask for the plan as calendar.csv plus one brief file per week-one post. The desktop app previews spreadsheets, so you can check the grid without leaving Codex. Codex can also generate images, which helps you rough out a visual for a slot marked "carousel" before building the real slides in our LinkedIn carousel maker. For the timing column, check the best time to post on LinkedIn for your audience instead of letting Codex guess.

What you gain: less planning each week, with room left for better ideas, new evidence and changes in the market.

Prompt 7: how do you get drafts that have something to say?

Settle the argument from the research before Codex starts smoothing sentences. Skip that and you get tidy prose that takes several paragraphs to make an obvious claim.

Using my business context and voice guide, write three LinkedIn drafts from these approved briefs: [briefs]. In each draft, make one clear argument for one reader. Open with a specific observation, tension, result or question that the brief supports. Walk through the reasoning with a concrete example. Use bullets where they make parallel items easier to scan. Vary sentence and paragraph length, with the odd one-line paragraph for weight. Close with a useful consequence or a natural next step. Aim for [180-300 words] unless the idea needs a different length. Don't restate the same idea at the end, invent personal stories, overpromise or lean on formula contrasts. Give two other possible openings and say briefly how each shifts the focus. List missing evidence apart from the copy that's ready to publish.

In Codex: you will run this prompt every week, so save it as a skill. Create the folder ~/.agents/skills/linkedin-draft/ containing a SKILL.md like the one below, with the prompt pasted under the header:

---
name: linkedin-draft
description: Use when I ask for LinkedIn drafts from approved briefs in briefs/.
---
Write three LinkedIn drafts from these approved briefs ...

Codex reads the description to decide when the skill applies, so write it as a trigger and leave summaries out of it. After that, $linkedin-draft calls it up in any thread and /skills lists what you have. In the terminal you can also put it in one of the ten /presets slots. Have Codex save each draft as drafts/week-1/{slug}.md. If a draft starts filling up with padding while Codex is still writing, send a correction straight away: a message sent mid-response changes course at once instead of waiting for the end. For more first lines to set against its two alternatives, run the brief through our LinkedIn hook generator.

What you gain: clear limits produce drafts that are easier to judge, so your review can focus on the argument instead of cutting padding.

Prompt 8: how do you set up a separate editorial review?

Give the reviewer one clear job: find what is weak in the draft and explain how to fix it. A comforting "this is strong" is no editorial process. In Codex, the cleanest way to separate the roles is a different profile with different permissions.

Check these LinkedIn drafts against their brief, the source material, my business context and my voice guide. Give each a score out of 100: relevance to the reader 25, specificity and evidence 25, strength and clarity of the argument 20, match with my voice 20, readability and layout 10. Show what each score is based on. Verify every factual claim, mark generic passages, and flag repetition, abrupt one-line fragments and paragraphs that would read better as lists. Return a revision ready to publish, a short note on the changes that matter, and any open questions. Invented evidence, leaked confidential details or a misquoted source block publication whatever the score. Keep all of it in the review document until a person signs off.

In Codex: a separate review means a fresh thread, never the one that wrote the drafts. In the CLI, create the profile file ~/.codex/review.config.toml:

sandbox_mode = "read-only"
approval_policy = "on-request"

Start the review with codex --profile review. The reviewer can read every draft and can't quietly overwrite any of them. It returns the scores and suggested revision in the thread; save them to review/week-1.md from your normal session, where the human sign-off happens. Before you approve, paste the revision into our LinkedIn post analyzer for a second opinion from outside.

What you gain: a rubric makes your standards visible. The score organises the feedback, and the source checks plus your own judgment decide whether the post is ready.

What should you do in the first hour after you publish?

Prompt 8 ends with a revision you approve and post. From that moment every MAXXING method works the same way: in the first hour after the post goes live, collect as many reactions as you can.

Start with people you can reach. Colleagues, friends, family and clients are the best people to ask, because you can contact them directly. Send them the link as soon as the post is up.

If you don't have that circle yet, use a community. Choose one where real people act by hand and someone handles moderation. The groups in the next section are ones we have already reviewed in other guides.

LinkedInLikes is our community for this step. It is free, and you only share a public post link.

• Members react, comment, repost, save and follow by hand from their own LinkedIn accounts.
• Comments are written by members in their own words, and post owners can't dictate the text.
• If someone confirms an action they didn't take, you can report it within 72 hours and the points are taken back.
• New accounts get 100 starter points for their own posts; you get more by supporting other members' posts.

Add your post on LinkedInLikes once it is live.

Which LinkedIn groups can you join for first-hour reactions?

These are free communities we checked on Oct 21, 2026, with member counts as each platform showed them that day. Every group has its own admins, so read the pinned message before you post there.

To get reactions on your own posts without hunting through groups, create a free LinkedInLikes account.

Part 2: pipeline and performance (prompts 9-12)

Prompt 9: how do you write posts around deals that got stuck?

Some of your most useful posts may answer a worry several prospects raised in private. The aim is to explain the issue well enough that the next prospect arrives better informed. This is also where Codex's access settings matter most.

Look through the CRM records and call notes for qualified deals that got stuck during [period]. Find the objections that keep returning, and keep three kinds apart: objections people actually said out loud, explanations that are only guesses so far, and operational holdups that no post will fix. For each recurring objection, note the person's role, the deal stage, the evidence we have and the question a helpful LinkedIn post should answer. Suggest five content briefs. Each should present the worry fairly, show where our approach fits, admit where it doesn't and name the proof we would need. Draft the two strongest posts without naming any prospect or exposing deal details. In a separate list, suggest which existing, approved resources a salesperson could send in a relevant conversation.

In Codex: if your CRM has a Codex plugin or an MCP server, connect it read-only; otherwise a CSV export in crm/ works just as well. With an MCP server, config.toml lets you allow-list the tools Codex may call and require approval for anything not marked read-only. The URL and tool names below are placeholders for your own CRM's:

[mcp_servers.crm]
url = "https://your-crm.example.com/mcp"
enabled_tools = ["search_deals", "read_notes"]
default_tools_approval_mode = "writes"

Deal data should never end up in memory. Run this prompt with memory off, and if you pull records through MCP, setting disable_on_external_context = true under [memories] in config.toml keeps that thread out of it. Ask Codex to keep prospect names in the source files only and to use roles in every brief it writes to briefs/stalled-deals.md.

What you gain: posts that answer the questions your sales team hears over and over, and fewer persuasive posts about problems your customers don't care about.

Prompt 10: how do you prepare replies to LinkedIn conversations?

When someone responds to a post, the right next move depends on what they wrote and whether they would welcome a business conversation. A like on its own says nothing about business interest. Codex sorts and drafts here; the relationship stays with you.

Here are LinkedIn conversations I've collected: [comments, inbound messages or interactions I noted by hand]. Sort them into four groups: questions that deserve a helpful public answer, conversations where sending a resource would help, clear interest in our offer, and general discussion with no commercial next step. For each relevant one, summarise what the person said they need, any business context we have and what we still don't know. Draft a reply that deals with their actual message before bringing up anything else. Where it fits, suggest one follow-up question or a resource we already have. Don't read business interest into a like, don't pretend to know someone better than we do, and don't push a meeting request into a conversation that can't carry one. Leave every reply as a draft for me to check and send.

In Codex: copy the comments and messages into inbox/this-week.md. Codex doesn't read your LinkedIn inbox, and for this prompt it shouldn't. Ask for one reply per conversation in replies/this-week.md, then edit and send them yourself. For public replies under your post, our LinkedIn comment generator gives you a few more openings to compare with Codex's draft.

What you gain: more thoughtful replies without turning every exchange into a sales sequence. Sometimes the right result is just a useful answer.

Prompt 11: how do you report on LinkedIn results with honest attribution?

Impressions tell you about distribution. They can't tell you whether the right people got interested or whether a post helped close a deal. Codex can join and filter the data in code, so next month's report comes out the same way.

Merge my LinkedIn analytics, tracked website visits, inbound enquiry records and relevant CRM data for [period]. Report reach, reactions and comments by post, topic and format; qualified inbound conversations; resource sign-ups and demo requests where we track them; opportunities with a recorded LinkedIn contact; and the time spent producing and reviewing posts. Keep directly tracked conversions, discovery people reported themselves and possible influence in separate columns. Don't attribute a whole deal to LinkedIn because one contact reacted to a post. Call out posts that reached relevant prospects, posts that reached a wide but less relevant audience, and posts with too little data to judge. End with three findings, three limitations and one experiment to run. Where you can, compare posts at the same age so new posts aren't judged against older ones.

In Codex: download the LinkedIn analytics export and your website and CRM exports into data/{month}/. Codex can join them with a script, which turns the "same age" rule into a filter instead of a promise. Once the report looks right, save the prompt as prompts/monthly-report.md and run it non-interactively; -o writes Codex's final answer straight to a file:

codex exec -o reports/2026-11.md "$(cat prompts/monthly-report.md)"

codex exec doesn't pause for approvals the way an interactive session does, so run it by hand once and read the output before you put it on a schedule. To estimate reach for posts that lack impression numbers, use our LinkedIn impressions calculator.

What you gain: a reason to repeat or change what you are doing, and a report that shows uncertainty openly instead of turning patchy tracking into a confident revenue claim.

Prompt 12: how do you turn your edits into a better system?

Every draft you edit records something about your standards. Write the recurring lessons down so you spend less time fixing the same problems.

Compare the first drafts, my edits, the versions I published and any performance data for [period]. Find the voice corrections I keep making, claims that needed more evidence, ideas that got clearer once they were framed differently, layout changes that made posts easier to read, and topics that started useful conversations with prospects. Suggest changes to the voice guide, the research briefs and the editorial checklist. For each new rule, show an example, say what it is for and note when breaking it would be fine. Keep style preferences apart from guesses about performance. Never turn one post's success or failure into a permanent rule. Give me a short change log to approve and one controlled experiment for next month.

In Codex: turn the linkedin/ folder into a git repository and commit each draft before you touch it:

git add drafts/ && git commit -m "Codex drafts, week 3"
# edit the drafts yourself, then
git add drafts/ && git commit -m "My edits, week 3"

Codex can then read the history, and /diff shows exactly what you changed, so "my edits" are on record instead of half remembered. Ask it to propose the new rules as edits to voice-guide.md and AGENTS.md that you approve line by line, and to log the experiment in experiments.md.

What you gain: a process that gets easier over time. Your corrections become guidance, and performance ideas stay testable instead of hardening into superstition.

How should you roll out all 12 prompts?

Add the prompts in stages, so you can check each output before another recurring job joins the list.

• Week 1, the base. Load your business context, run prompt 1 and compare the voice guide with writing you would really publish. In Codex terms, AGENTS.md, context/ and voice-guide.md exist and are committed.
• Week 2, the research backlog. Run prompts 2-4 on a manageable set of customer conversations, market sources and LinkedIn examples. Switch on live search for prompt 3 and keep the median script from prompt 4.
• Week 3, the first batch. Use prompts 5-7 to turn approved material into a calendar and a few drafts. Create the linkedin-draft skill this week.
• Week 4, the review. Run prompt 8, publish what you approve, work the first hour on each post and record the changes you make. Line up the people and the community who will react in that first hour before the first post goes out.
• Weeks 5-6, conversations. Add prompts 9 and 10 once you have enough sales context and inbound activity to make them worthwhile. Connect the CRM read-only only at this stage, when you know which fields you need.
• Every month after that. Use prompts 11 and 12 to review results, improve the instructions and pick the next experiment.

For each recurring job, write down its inputs, where the output goes, how often it runs and who owns it. Test one run before trusting the schedule, and have Codex report any missing access or unfinished work.

Once everything runs, each repeat job sits here:

Codex has three ways to run the repeat jobs on a schedule:

• Automations in the desktop app. Schedule the Monday research run (prompt 3) or the monthly report (prompt 11) from the app. An automation can reuse an existing thread and keep the context built up there. Local scheduled runs need the computer on and the app open.
• Codex Cloud. Delegated runs carry on while your laptop sleeps, but they only see files and access that exist in the cloud environment.
• The CLI. codex exec runs one prompt to the end, so your own cron job or CI can start the report and save it with -o.

While a scheduled or cloud run is working, you can follow it in Codex on iOS, which shows background commands on connected hosts and lets you stop them. Codex prepares the plan; publishing stays in your hands.

You don't need 12 separate agents on day one. Start with the job that eats the most time and produces something you can judge reliably.

What can't Codex do for your LinkedIn content?

Codex can organise your research, find useful evidence, prepare drafts and learn from your edits. How good the posts are still depends on the experience and judgment you put into the system.

That means holding a real opinion, being willing to explain it, and noticing when a polished draft says something you would never defend face to face.

Load your context into AGENTS.md today and run prompt 1; better inputs make every later step more useful. When the first post is live and you want its first hour to count, add it to LinkedInLikes.

Related reading

• How to build an AI content system for LinkedIn
• LinkedIn maxxing with Claude Code
• LinkedIn maxxing with ChatGPT
• The 2027 LinkedIn marketing playbook