How we made 275 task guides easier to use

If you run a business, clear steps help your team get work done. We reviewed 275 guides and added setup help to the downloads we build. Choose one job, connect the tools it needs, and check the result.

Work in progress. Updated September 10, 2026. This is the record of one real guide-improvement job. It does not report 275 completed client jobs.

Why this job started

Dennis Yu opened the shared-memory guide and found an empty colored box at the top. The first explanation was too hard to follow. Links often sent readers to a provider’s help site before explaining our own process.

He asked for a clear start: who the page helps, why the work matters, what to do, and how to check the result. He then extended the job to all 276 entries in the Task Library and its ZIP downloads.

Our Goals, Content, and Targeting brief, often shortened to GCT, made the task specific. The goal was to make one useful first job easy to start and check. The content was the actual guides, page sources, download files, and test results. The target reader was a business owner who does not need to know how AI tools work inside.

What we started with

The selected article list had 85 pages. We separated task recipes and playbooks from topic pages, opinions, and other supporting material. We also checked one related skills-and-agents article, bringing the current opening review to 86 pages.

The library had 276 task entries. Its source contributors had marked 125 complete, 128 needing work, and 23 as gaps. Those labels did not prove that someone had tested each guide, installed a skill, connected an account, or completed the work for a client.

We followed Create or Update a Definitive Article. A definitive article is the maintained main recipe for a task. This page is its meta article: the record of what happened when the recipe was used.

What changed

The article rules and definitive-article guide now call for a short, plain opening and a meaningful picture or diagram on the first screen. The recipe must explain its starting point, needed inputs, steps, measured result, checks, and next handoff. New terms get a short explanation and an owned teaching link.

The library’s published release contains 274 reviewed local guides and one reviewed, pinned external guide for Positive Mentions. Each review belongs to the exact source version. Changing that source makes the old review expire.

All seven ZIPs now contain START-HERE.md. It tells the reader to choose one job, share the guide and needed files, use a first-run prompt, and check the saved result. It explains account access and the difference between doing one job and setting up repeated work. Downloading a file alone does not install a worker or start a schedule.

Our skills and plugin explanation uses the same terms as the system map. A skill gives the steps. A pack groups useful skills. A plugin packages supported features for an app. An agent is the AI worker that uses the instructions. A job supplies the trigger and work context. Repeated jobs need an actual schedule and access to their inputs.

What we checked

The one remaining unreviewed live import is the external Google Ads analyzer. A reviewed contribution is prepared, but its upstream version has not changed. The old contributor labels stay visible beside the separate review evidence. Missing canonical articles and real-run examples also remain visible.

Where this fits in the Content Factory

The Content Factory turns real source material into useful content, publishes it, and helps the right people find it.

This job uses Process to improve the guides and Post to save and check them. It has no measured Promote result. Within the SEO Tree, the task’s main recipe sits under a real topic hub. A run record like this supports that recipe with evidence. Both connect to the Task Library; they do not become competing main guides.

What went wrong and what we learned

The old publishing attempt failed its comparison check and rolled back. Later normal public reads showed the intended pages. We kept the failed attempt in the record and did not claim that the old publisher had been repaired. The former article-publishing block no longer stops this work.

Review caught task rewrites that had lost useful evidence rules. We restored named sources, limits on relationship claims, and clear handling of missing proof. Fictional teaching examples are labeled as fictional. They do not become completed runs.

We also found two misleading links. The old Thank You Machine route now opens the Face Socks recipe, so it cannot certify the video workflow. The Dollar-a-Day coaching offer and the actual how-to serve different purposes. The affected task guides now explain those limits and link to the how-to where they teach the method.

Our visual checker had a separate bug. It could measure a page before its stylesheet arrived, then save a screenshot after the layout had moved. The timing repair passed 25 browser tests and is merged. It keeps the existing visual requirements and records an unstable page as a failure. We are using it to check the remaining diagram changes.

The accepted guide changes have separate authoring and independent-review receipts. One bounded routing trial timed out after about 60 seconds and produced no accepted guide. Total model cost, tokens, active human review time, client revenue, and performance across all task types were not measured.

September 10: keep improving in small batches

We checked the live library again: 275 of 276 tasks have reviewed instructions. That count does not mean 275 jobs were completed for clients. The contributor status labels and proof of real runs remain separate.

The new maintenance guide explains recursive self-improvement (RSI): check a real result, save the lesson, improve the recipe, and test the revised steps on the next run. It defines where a worker saves the checkpoint so an interrupted job can resume without repeating finished checks. A daily maintenance task is now configured to review changes and known gaps in batches of up to three guides. Configuration alone does not prove that its first scheduled run has fired.

Review also clarified downloads. Library-built ZIPs include Start Here instructions. An outside provider’s full suite may use a different setup guide and must have its own files and access checked. Downloading either archive does not activate a worker or schedule.

We recovered 52 saved article checks from the interrupted run; all contained the full published article text. An offline review of seven media exceptions found the saved image and video identities still present. The saved-source differences included filled-in image descriptions, extra theme elements, and a lazy-loading change. This does not prove that every image loaded, every caption worked, or the automatic descriptions were useful. Those checks retain their limits.

September 10: clearer diagrams and reliable source files

We updated nine more pages so readers can see the main steps and how the work connects. The changes cover Dollar-a-Day, website checks, weekly audits, article thumbnails, agency websites, clear replies, definitive guides, shared work records, and skills and agents. Each saved page matched its reviewed source. Normal public reads retained the full article text and source media. Phone and desktop checks are recorded separately from those source checks.

We also fixed a cause of broken rebuilds. Local build commands had pointed to a source folder that other tasks could change by switching branches. They now use a separate folder pinned to the reviewed source. The actual commands reproduced all 77 Dollar-a-Day output files, the main course alone, and the canonical skill-rule page exactly. The maintained Dollar-a-Day source includes the new title and diagram, so a rebuild keeps the repair.

The review caught defects before publication. One diagram kept the labels but changed the arrows. Another candidate left an unmatched HTML tag that broke the page layout. A later phone check exposed joined words and clipped text. The lesson is simple: check the meaning, check the source, and look at the real page. A passing count or a saved file cannot prove all three.

The broader public-page review also found a large duplicate title card hiding the first steps in How to Document a Task. We saved a small change in its page builder to hide that duplicate card. The useful diagram, instructions, links, and source media remain. Separate video-player checks are still open. The Task Library now records 275 reviewed guides out of 276. Its updated display explains why this differs from the older 125 contributor-reported complete labels. These edits remain part of the same maintenance run. They do not count as new client jobs or prove that all 276 catalog tasks have been completed.

September 10: check what each image really shows

We reviewed 76 images in the seven-page media check. A filled-in description was often just a file name. Even a description that sounded right could be wrong. We updated 72 image descriptions, replaced three empty or misleading diagrams, and removed one repeated screenshot. The six pages with images now have descriptions checked against the pictures themselves. The seventh page had no raster images to change.

The meta-article guide now shows where a run record fits in the Content Factory: produce the source, process the record, post after review, and promote when useful. The positive-mentions guide now adds its three scores correctly. The SEO Tree guide shows how the home page, main topics, smaller topics, and detailed examples connect.

We also checked the pictures in Metrics, Analysis, Action (MAA), the weekly MAA report guide, and How I Talk to One Agent. One photo caption claimed two people were golfing in Phoenix. The picture showed one person outdoors wearing a microphone. We removed the unsupported part. Reviewers also removed names from other image descriptions when the source did not identify the people.

The wider review checked 57 public pages in four screen settings each. It recorded 188 passes, 38 failures, and two errors. Those are checks of page layout and readiness, not completed business tasks. Some failures came from video settings that still need work. One page used a valid small-screen layout that the checker did not recognize. Another had a punctuation error in the check itself. Slow font loading and a late page shift also remain recorded. We kept the original results and added separate evidence instead of rewriting a failure as a pass.

The final SEO Tree save matched the reviewed source, but the first public check still received an old cached page. After that cache expired, the normal URL retained the new full text, diagram, and all 31 reviewed image descriptions. Public font loading still exceeded both the normal 10-second check and a separate 20-second diagnostic. Its final rendered check remains open. The new diagram was visually reviewed in the real-page preview; that does not replace verification of the public render.

These lessons now feed the Task Library maintenance guide. Check the meaning of arrows, view the real image, use the real page in a preview, and open any folded section before judging it. Then verify the saved page at its public address. This is recursive self-improvement: a real run teaches us how to make the next run better.

The final release test caught one more mistake in the work record: the source said the attempt had ended, but the saved dashboard copies still said it was running. We rebuilt those copies with the same tool that builds the library. All 91 tests then passed against the final files. We also fixed the run-record instructions: rebuild the public copies before testing, and commit them with the source record. The failed deployment stayed in the record.

The next handoff

This maintenance attempt has ended with useful repairs published and open follow-ups saved. The daily maintenance task will take the next small batch. One instruction source still needs review. Some video settings need approval to deploy, and font-loading issues need a focused fix. We also need records from real people or agents using the recipes. The 275 reviewed guides do not prove 275 successful business jobs.

To use the work now, open the Task Library Dashboard, choose one job you understand, and follow its Start Here path. Save what happened using the recipe and run-record guide. That evidence shows which steps work and which need a better recipe.

Execution record: recipe-audit-20260905-080140. Started September 5, 2026 at 08:01:40 UTC; first recorded September 6 at 01:18:27.649605 UTC. Status: partial; this attempt ended September 10, 2026 at 17:01:27 UTC. Published repairs and remaining follow-ups are recorded above. The starting definitive-article source hash is 73c6c011d5fe83b152a2192b175f320b7a6470ca81b4f71672332ae899f10b9b. Edits, retries, and this write-up remain part of that one execution.