Skip to main content
#ai-agents#technical-writing#publishing-workflow#automation

How JJo takes a blog post from a reader’s question to publication

The editorial workflow covers checking sources, equations and figures, parallel review, version control, approval and verification after publication. The article identifies the steps already tested and the remaining cloud and Blogger integration work.

Can a convincing AI-generated draft go straight online? A sentence may read smoothly while quietly combining results from two different experiments. A figure may have a source credit even though nobody has checked its terms of use. If a sentence changes after review, does the earlier review still cover the manuscript in front of us?

JJo’s blog automation addresses this gap. Beyond drafting quickly, we want to know what was read and explained, which version reviewers saw, and what readers actually received. This account is for people producing technical articles with AI. By the end, the differences between preparing a work order, performing a review, and confirming publication should be clear.

This description starts from the process in use on October 9, 2026. Its account of progress was checked against private operating records on October 23. Those records document reviews of actual manuscripts, repairs to the way tasks were connected, and code tested locally. They leave open the review status of other research drafts and the reliability of a system that would produce articles in the cloud every day and publish to multiple channels.

#What the operating record establishes

The private queue checked on October 23 contained 45 articles. The research input repository also contains material outside that queue and revisions of the same article. A count of its records would therefore overstate the number of articles reviewed.

In the latest work recorded, the active agent read primary sources and supplementary material and produced new private revisions. Several drafts improved. Checks still outstanding included current equation inventories, figure rights, search data, reviews from five reader perspectives, and the rendered pages. Some private preview builds succeeded. On October 23, the operator also verified the invocation and termination of three subagent tasks that requested Luna. All three reported unfinished checks. Their findings were incorporated into one private revision. The requested model name alone could not establish which model the tool actually ran. This update involved neither a public GitBlog deployment nor a Blogger publication.

To interpret a status record, ask three questions together. Which article and version does it cover? Which checks actually passed? What evidence is needed next? A file count or a label such as “ready” cannot answer all three.

#1. Start with a reader’s question

The starting point is a question a reader wants answered. For a news story on a power plant, a reader might ask, “Does the increase in capacity mean the plant has already generated that much extra electricity?” An explanation needs to distinguish the power the plant can deliver from the energy it has actually produced over time. Readers meeting MW and MWh for the first time first need to know what each unit measures.

Readers who want to go further can follow the same example through the comparison conditions, operating time, assumptions behind a calculation, and scope of the announcement. The explanation gains precision as it goes. It need not suddenly switch to technical vocabulary after a simple summary or repeat the same definition for different education levels. It should also distinguish the study’s results from an example the author constructed to explain them.

Length and figure count are poor measures of how well an article answers the question. Have the necessary concepts been defined first? Are the cases compared under the same conditions? Which conclusion would change if a condition changed? An equation or infographic earns its place by helping with the explanation. This process article uses prose and short examples to show the relationships.

#2. Agents do the work, and the harness connects their work

An agent reads material, exercises judgment, and produces a draft or a review. A source reviewer checks claims against the original. A methods reviewer examines equations, units, and analytical conditions. An editor tests whether a reader can follow the explanation. Dividing these roles gives reviewers a chance to bring different evidence to the same judgment.

The harness is the code and rules that connect those activities. It binds a review to its sources and manuscript, identifies missing results, and prevents an old decision from carrying forward when the draft or review criteria have changed. A work order with a checklist can exist before anyone reads the paper. A version check can establish that two files match while leaving their scientific accuracy untested.

In practice we distinguished preparing work for a role from carrying it out. Generating instructions and a handoff for the next role is preparation. Execution means opening the material, examining it, and recording which claim was checked and where its evidence was found. Generating a work order is insufficient evidence of a model invocation.

#3. Check sources, equations, and figures in both directions

Review from source to draft asks, “What important evidence did the article leave out?” Read the relevant methods and supplementary materials as well as the main text. Link key figures, panels, equations, and procedures to the review inventory. Leave anything unread marked as unverified. A PDF text extraction can succeed before a reviewer has examined a single claim.

Review from draft to source asks, “Where did this explanation come from?” Trace each number, calculation, figure, and account of a method back to its evidence. Distinguish equations stated in the paper from rewritten notation, prior methods, and teaching examples. When an announcement contains no equation, examine the procedure and physical quantities in words. A formula invented to fill a template would misrepresent the evidence.

During a review of an industrial announcement, we found that the draft had added capacities belonging to different contracts. A teaching calculation already in the manuscript was also missing from the review inventory. Recomputing the numbers would have left both attribution and scope unresolved. We identified which value belonged to which contract, distinguished a company’s stated reason from an independently established fact, and narrowed a claim of absence to the documents actually checked. Media, reader, and rendering reviews remained open.

Figures require the same reading. A photograph can show appearance and surroundings, a diagram can show relationships, and a measured plot can show results under stated conditions. A photograph of another power plant might illustrate the background. It cannot establish that construction under the contract being discussed is complete. A caption should tell the reader where to look, what is being compared, what can be observed, and what the image leaves unresolved. Attribution and permission to use the figure require separate checks.

In another research review, we withdrew an initial judgment about which material had been analyzed. That judgment had rested on a figure legend alone. We compared the other reviewers’ findings with the paper and supplement and recorded the discrepancy. Confidence in an interpretation and the number of completed reviews do not resolve an inconsistency in the source.

#4. Parallel reviewers need the same version

Once a draft exists, independent checks can proceed in parallel. One reviewer can check sources, another the equations and methods, and another the links between claims and evidence. The parent agent gathers the findings. Comments on a single paragraph become one set of edits, starting with factual errors, followed by missing logical steps and readability. In our trial, the parent grouped findings about the same issue and corrected a comparison table’s denominator and the interpretation of the test period. All three tasks ended. Full source coverage, reader review, and the public page still needed checking.

The role-session code prepares instructions and an empty result form for each role and marks the work as “not yet run.” It can construct arguments for parallel calls, but an actual agent tool must still be invoked. A Luna-profile work order alone establishes neither that Luna ran nor that a subscription delivered a completed parallel review. Check the requested name, the records of invocation and termination, and the evidence the worker returned.

Additional roles do not automatically make reviews independent or accurate. Asking a model to read from the perspectives of a nonspecialist, student, or researcher can expose an unexplained jump. It is an editorial check. Measuring learning would require actual readers and an appropriate study. Claims of severalfold time or token savings would likewise require measurements.

If several workers edit one file at once, comments and versions can get mixed. We assign distinct review scopes and files, then apply manuscript changes together. Replacing a figure after review requires another look at its rights, caption, connection to the surrounding explanation, and actual display. Even a title change can alter the scope of a claim.

We use content hashes to check whether the manuscript, evidence, and settings are identical to those reviewed. A hash identifies file contents. It does not score scientific accuracy. If the input or criteria change, the old pass remains in the history and the new version needs current evidence. In one resumed session, the team repaired stale references and reviewed the current material again without rewriting the manuscript.

#5. Review and publication approval answer different questions

A completed source review establishes that claims were checked within its scope. A completed reader plan establishes the question, required concepts, and sequence of the explanation. Asking an early reviewer to inspect the rendering of an illustration that does not yet exist can block useful work. Completing the plan alone still leaves the article unready for publication.

Before publication we compare the current manuscript, media, destination channel, and required review results with the scope authorized by the operator. Approval of an earlier article or deployment does not automatically cover a different revision or set of images. Within a valid approval there is no need to ask the same question again. Scientific accuracy, rights, search evidence, reader review, and screen checks still require their own evidence. A changed draft or asset may need review or renewed approval. Blogger and Instagram have their own channel permissions and connections.

The workflow for a new article also requires checking actual Google Search Console or Naver search data first. On October 22, we opened the three-month web search report in a connected browser. The dates actually displayed were August 18 through October 20. The query table contained zero rows and the pages table contained one row for the homepage. We had accessed the tool, yet obtained no article-specific queries. Those results establish neither the indexing status of every page nor the absence of search demand. A general web search or a test with mock data cannot supply the missing article-level evidence. At the time of that record, this process article itself was kept as a draft. Its presence in the repository did not establish publication.

These distinctions also tell us what can proceed. When source access or permission is missing, record what has been checked and what will allow review to resume. Meanwhile an editor can plan an explanation or look for usable photographs. Finding a candidate leaves the rights check open. Producing or approving this process article likewise does not make an existing science draft ready to publish.

#6. Check the build, the deployment, and the public page

On GitBlog, the first checks build the site from the manuscript and run its automated tests. Continuous integration (CI) runs checks when changes are submitted and can catch syntax errors or broken internal structure. A successful build leaves other questions unanswered. Does a citation support its claim? Does a long equation overflow on a phone? Does the figure illustrate the accompanying explanation?

Next, check that the approved change reached the deployed site. The repository change, the deployment job, and the reader’s public URL are three observable steps. A local preview verifies only the preview. Open the public address again and check the title, body, images, equations, and internal and external links.

Some problems appear only on screen. A detailed research table could obscure the default view, expanding the sources could make the reading position difficult to recover, or the homepage could display an entire long article. Inspect both desktop and mobile layouts. Keep the main explanation and its limitations in the default reading path, with long derivations and supporting evidence available when needed.

We also read operating records back after saving. Creating a write request, verifying the stored value, and checking that the result is readable are separate observations. During an actual handoff, we compared destination values with existing entries and checked for duplicates after retries. That check covered the handoff in question. It did not establish daily production success or reverify every published article.

#7. Maintain the process article from a single source

A single source of truth (SSOT) gives an explanation one authoritative version so copies do not diverge through independent edits. This operating guide has a source document. The public article is derived from the material suitable for disclosure. Account information, private file locations, and internal task identifiers are unnecessary for understanding the process and stay private.

The description of the process and the record of an actual run serve different purposes. When a procedure changes, update the description and its rationale. Record execution only within the scope actually verified. A status query is not a completed implementation. If another edit is discovered, reread the current version and preserve it before applying changes.

The goal is to keep explaining changes and lessons in this article. When a run changes something a reader needs to know, update the relevant section. Revising the authoritative source does not update the public page on its own. Updating the public version requires review, approval, deployment, and verification at the public URL.

#8. Moving the workflow to the cloud

A reproducible transfer needs more than the harness code. Role configurations, skills, operating documents, and tests travel with it. Keep credentials and private operating data outside public deployment bundles. Handle original source files and photographs according to their terms of use. In the new environment, check file integrity, dependency installation, task preparation, and actual execution in that order.

Earlier handoff records document installation, file hashes, and test runs. This update did not repeat the original cloud-environment verification or the full test suite, so their historical counts are not presented as current successes. A passing test with test inputs does not establish that an actual article was reviewed, a model ran, or an article was published.

Earlier subagent runs, handoffs to Sheets, merges into main, CI, deployments, and public-page checks each need to be traced to the original execution evidence and content version. This update did not revalidate the entire history. The name of the requested model, evidence of an invocation, and an identity authenticated by the tool are separate facts. Where the identity could not be established, it remains unknown. Blogger publication, the next scheduled run, and publication of research drafts still awaiting review each need separate confirmation.

The same boundary applies when preparing post-publication analysis. Open the follow-up task once observations from the public page match the current work record. Testing a handoff with mock data establishes neither that an article was analyzed nor that a new post may be published.

Blogger has an additional dependency. The local publishing adapter can create or update a post through the API, but the cloud handoff bundle did not include the publishing tool and account authentication together. The new environment needs its own checks of the tool, dependencies, and credentials. Keep credentials out of public manuscripts and distribution bundles. Verify sign-in, actual creation or update, and the public result separately.

The channels also have different editorial purposes. A Blogger article should explain one or two core relationships well enough for the reader to understand them there. GitBlog can take the same reader through the models, methods, comparison conditions, and sources in greater depth. A cross-link should still leave each article able to answer its own question. The next objective for automation is to bring what the article promises closer to what readers actually see.

A completed run starts with a reader’s question, builds an explanation, checks its evidence and usage terms, obtains review and approval for that version, and verifies what readers receive. When the next operator and the reader can distinguish completed work from open questions, automation contributes to reliability as well as drafting speed.

Scope of the October 23 update

We verified the structure of the parallel-review skill and confirmed that 115 related session tests passed. The three actual subagent results were accepted with their original open items, and the findings were incorporated into one revision. The test count does not establish that all articles were reviewed or that automatic publication succeeded. Detailed identifiers for manuscripts and runs remain in the private operating ledger.

Connect