My LinkedIn strategy as code: building a multi-agent content pipeline

Last Sunday, 22:00. On my desk are eleven Markdown files, next to them an open spreadsheet with 90 days of LinkedIn analytics. Cold mate in the glass, an honest question in the back of my head about why I am building a pipeline rather than writing the next post.

The spreadsheet gives a clear answer. My profile has 511 followers and delivered 44,483 impressions over the last 90 days at an engagement rate of 0.18 percent, a factor of 11 to 16 below the LinkedIn benchmark of two to three percent. On 52 of those 90 days the feed stayed entirely silent, without a single reaction to any of my posts.

One post from 20 April 2026 alone delivered 23,703 impressions and accounted for more than half of the quarter’s reach. Eight engagements followed, an engagement rate of 0.034 percent on a post the algorithm clearly chose to push. Distribution worked, attention did not.

The audience itself is right. 35 percent of my reach is top-management in the DACH region, most of it around Frankfurt, classic Mittelstand. What I am missing are five writing routines that I knew in theory and dropped after a long consulting day: three hook variants per post rather than one, a quality gate before every live post, five substantive comments in the 30 minutes before posting, alternating formats rather than two text posts in a row, and a 50/30/20 pillar rotation rather than ad-hoc topic choice.

So instead of writing more posts, I am building the pipeline that enforces those routines.

What the pipeline actually is

Inside pfisterer-consulting/agents/ there are eleven Markdown files. Each file is a phase-agent with a clearly defined task, a voice anchor, and a YAML output header for the next phase. The master orchestrator sequences and parallelises the whole thing in waves.

The phases run in this order:

• Pillar Selector decides which content anchor carries the post (A for ERP reality in the Mittelstand, B for AI in the Mittelstand, C for build-in-public solo consulting). Default ratio 50/30/20.
• Trend Scanner scans sources from the past week and proposes five to ten topics. Skipped when the topic is already in mind.
• Research Brief Writer builds a defensible position, collects eight to fifteen dated sources, and drafts three hook variants in the bold-claim, story, and data styles.
• Visual Storyteller generates two ComfyUI workflows for FLUX.2 Dev Q8_0 with the Mistral 3 Small encoder, one for the 1536x1024 hero image and one for the 1024x1024 LinkedIn post.
• Article Writer DE+EN drafts two independent articles with a counter-position section and FAQ in ### headings. Not translations of each other.
• SEO+GEO Frontmatter finalises the title, SEO title, description, tags, and reading time, plus JSON-LD schemas for FAQPage and Article.
• Social Post Writer writes three hook variants per platform for LinkedIn DE, LinkedIn EN, and XING DE, each with its own first-comment text.
• Anti-AI-Slop Editor scans everything from phases 3 to 5 against pattern lists and blocks anything above slop score 2.
• Engagement Strategist builds the pre-engagement plan, eight comment-reply templates, and the DM candidate list for T+24h.
• Lead Nurturer runs asynchronously for seven to fourteen days after publishing and manages the DM sequence to ICP matches.

The pattern is not my invention. Anthropic published the architecture for their multi-agent research system on 13 June 2025, an orchestrator-worker model with specialised subagents that manage their own contexts (Anthropic Engineering). What I added on top is that every agent lives as a Markdown file with YAML front matter, just like Skills and subagents in the current Claude Agent SDK. I chose Markdown over code so that the content of every phase can be read and recalibrated in plain text.

Phase 6 is the actual bottleneck

The Anti-AI-Slop Editor blocks more drafts than any other phase. Most of the time the trigger is not an obvious buzzword but a structural pattern: echo sentences, rule-of-three lists, smuggled-in antitheses, colon-reveal spam, telegraph-style numeral stacks.

On the dry run for this very article my first hook draft came back at score 13. Dramatic-pivot antithesis, machine-listed phase names, a banned word in quotation marks that I had hoped to defuse with the quotes. Phase 6 flagged all of it as slop and sent the draft back to Phase 5. Three iterations later the score was at one.

The uncomfortable part was not the score itself. It was the realisation that roughly eighty percent of my recent posts used the same patterns. Phase 6 now forces me to rewrite on Sunday night and spares me the comments on Monday morning in which someone politely notes that the post reads as AI-generated.

What changes for every post

Four mechanics now run automatically:

1. Three hook variants per platform. Phase 5 writes a bold claim, a story, and a data opener. Phase 6 scores them on slop. The orchestrator picks the slop-cleanest variant with the strongest pillar fit. Previously every post had exactly one hook, usually the weak method opener that averages 410 impressions in my own 90-day data (n=12).
2. No external link in the LinkedIn body. Phase 5 consistently places the link in the first own comment, three to five minutes after posting. External links in the body cost roughly 60 percent of reach, the first-comment pattern performs markedly better.
3. Pre-engagement list in Phase 7. Five substantive comments on ICP-relevant posts, 25 to 30 minutes before my own posting. The algorithm reads this as a priming signal and gives the new post more initial distribution.
4. 60-minute Golden Hour protocol. Phase 7 ships eight reply templates for the most frequent response patterns in the Mittelstand consulting context, from the agreeing CEO voice to the vendor employee with a counter-position. That cuts response time during the 60-minute window after posting in which LinkedIn decides how far to push the post.

These mechanics appear in every current write-up about the LinkedIn 2026 algorithm. The difference is only that they no longer depend on my daily form, because a pipeline phase requires them.

What the pipeline cannot do and is not supposed to be

A few honest limits.

For a profile with 511 followers an eleven-file pipeline is objectively over-engineered. Anyone starting at 30 followers should write manually for four weeks and test hooks by gut feel first. My case is different because the profile has five and a half years of substance and an ICP demographic I genuinely care about. Architecture here is more economical than another quarter on the same spot.

The pipeline does not write for me either. Phase 3 takes the outline from the research brief, holds the voice anchors from the pillar playbook, and produces a draft. What follows is reading, cutting, and rewriting of whole paragraphs whenever a specific memory cannot be reconstructed from the brief. The pipeline does not know what I discussed with a CFO at the workshop table on Thursday, or which half-sentence in a phone call with an IT lead in Frankfurt ends up carrying the article. Those anchors stay human.

ComfyUI runs deliberately outside the automated flow. Phase 2 hands me the JSON workflows, I launch them locally on my RTX 4090. Roughly three minutes per image for rendering, followed by manual review. Logo hallucinations occur regularly, as do readable letter shapes that resemble vendor lettering. Manual review after each generation protects against images I would not approve.

Frequently asked questions about the multi-agent content pipeline

How long does a full pipeline run actually take?

A full run takes 35 to 40 minutes:

Compared to a purely manual set of an article plus social posts that is more time, not less. The actual gain sits in hook quality and quality-gate protection, not in speed.

Why does a solo consultant need three hook variants per post?

My 33 posts from the last 90 days show a clear performance spread. The method opener averages 410 impressions (n=12). The bold claim averages 7,034 impressions (n=2) at 0.2 percent ER. The story opener averages 1,580 impressions (n=4) at 0.4 percent ER. Without A/B testing at the brief stage I cannot know which variant fits the substance of a given post. The pipeline therefore produces three every time, and Phase 6 picks the slop-cleanest variant with the strongest pillar fit.

Will the pipeline make my writing uniform?

Several mechanics work against that. The 50/30/20 pillar rotation enforces different voice anchors for the ERP pillar, the AI pillar, and the build-in-public pillar. The Anti-Slop Editor flags word-frequency stuffing from the third occurrence of a term. And the voice bonus in the slop score rewards openings with a specific person, time, and place, which rules out noun pile-up almost on its own. Without a pipeline, style drifts toward the slop middle over time because no systemic mirror exists.

Which phase is most disposable when time is short?

Phase 0, the Trend Scanner. With the topic already in hand, the orchestrator skips straight to Phase 1. Phase 6 is the wrong place to save time. It is the only protection against AI slop reaching the live post, and Phase 7 delivers the Golden Hour protocol that makes the first 60 minutes after posting economically usable in the first place. Both phases provide the highest performance gain per invested minute.

Next step

Are you thinking about building a similar pipeline for your content operations, or are you a solo consultant wondering whether this model fits your situation?

The pipeline grew over several weeks, with the corresponding dead ends, and all of that is captured in the playbooks. If you want to evaluate the model for your marketing or for internal knowledge work, let us talk for 45 minutes. No pitch, an honest read on whether the model fits your case, or which simplifications you can adopt right away without copying the whole architecture.

→ Book a no-cost intro call

→ Or read first: AI automation in the Mittelstand · Website relaunch with AI instead of an agency

Sources and references: Anthropic — How we built our multi-agent research system, 13 June 2025 · Claude Agent SDK — Skills and subagents · GSD — Get Shit Done in context engineering · OpenClaw — orchestrating autonomous agents · Agents as Markdown personalities

More on pfisterer.xyz: Shadow AI in your own workflow · Website relaunch with AI instead of an agency · OpenClaw as a personal AI assistant