At first, finishing one topic took me a whole day. A few months later, the same pipeline needs only a few short inputs. The agents collect, filter, verify and restructure the material, produce the bilingual web pages, and publish them. I mainly handle verification, confirming or rejecting the work at key checkpoints.
genai-playbook mostly pulls together material scattered across different sources, filters and verifies it, then restructures it into bilingual pages that people can read directly and other agents can call. The real time sink was never typing. It was keeping everything in my head at once: the current state, unchecked numbers, red lines that could not be crossed, and where the work had to go next.
First decide what is worth reading, then confirm it does not overlap with a topic already completed.
Read the original in full, checking item by item that numbers, quotes and context have not drifted.
Not a straight copy: rebuild the structure, then generate the Chinese and English pages separately.
Check links and charts, then update the index, the sitemap and the publish record.
If any one step in this chain was forgotten, everything after it kept running with the error baked in. Back then I was the scheduler, the executor and the checker all at once. The pipeline could finish, but every new topic still meant walking the whole thing through by hand again.
This line was not designed in one pass. It ran first, then exposed its problems; each time a particular failure kept recurring, I added another layer to the harness. Months later the person had stepped out of most of the execution, but the judgments from the start did not disappear. They moved somewhere else and kept doing their work.
First I let the pipeline hand work across sessions. Then I separated writing from checking, moved the stable red lines into code, and only at the end reduced the daily scheduling to a few words. The level of automation did not rise in a single move.
A long task spans sessions, and different agents may pick it up. When state exists only in context, whoever takes over next has no reliable way to know what has already been done or what constraints still apply.
Reading its own draft while still carrying what it meant to say, it quietly fills in the missing information, and its review keeps sliding along the same judgment that produced the draft.
The "close enough to pass" edge cases are the dangerous ones. On the model's in-the-moment judgment alone, the same rule can come out differently today and tomorrow.
The deep radar covers a fourteen-day window; one full sweep costs about 3M tokens and takes 15–20 minutes. Run it every day and most of what comes back is "already covered."
Keep only the final recommendation and you cannot see at which gate a candidate was cut, nor whether the filter quietly killed something worth keeping.
Topic selection is not handed to one agent to think through end to end. About 190 lines of topic-selection.js hard-code the control flow into four phases, and the script decides what runs in parallel, what waits and what may pass, while the agent only fills the current cell.
schema · SCORE_SCHEMAevery subagent returns fixed JSON fields, not a paragraph of prose.
await parallel(...)parallelism stays inside a phase; the barrier keeps partial results from flowing on early.
filter(isReal)even if an agent says keep, PR-speak or a forced so-what is still dropped in code.
A recheck does not ask the original agent to look again. Every round starts a subagent in fresh context, carrying no conviction from the previous round, to judge from scratch until it finds zero issues or hits the round limit.
restructure, write or build pages; holds no sign-off.
FACT · up to 6 roundschecked line by line against the source; exits only at zero issues.STYLE · prechecka script first clears the mechanical, certain problems.STYLE · up to 4 roundsthen a fresh auditor handles meaning and tone.Here a hook means one of genai-playbook's own pipeline checkpoints, not a global tool hook in ~/.claude. Each check fires at a fixed moment from a project script: the ones meant to block return a non-zero exit code, the ones meant for a human get gathered into a queue.
number-drift.py
reconciles visible numbers across the report and both pages; extras on a page are flagged as drift.
exit 1 · blockpublish-precheck.py
checks template leftovers, registration, homepage card, links, bilingual cross-links, figure sources.
exit 1 · blockpublished.txt
the published list drives stats, sitemap, search and MCP, so no outlet keeps its own copy.
single sourcepending-status.py
scans each topic's STATUS and surfaces unchecked manual items like GSC and git push.
human queueI still start the pipeline section by section, but I no longer execute each step by hand. Each line only states intent; the harness expands the plan from PIPELINE and the current topic's STATUS, and the agents plus deterministic checks take care of filling in this cell.
today's topics
do the first one
generate web
deploy
It looks simple today, but not because the work was ever simple, and not because I found some universal prompt. The short input is only the operating surface left behind after months of engineering: the search logic, the quality bars, the failure handling and the cross-session handoff have all been moved into the harness.
My work went from execution to verification: give direction, trigger the next section, confirm or reject at the key checkpoints. The repeated execution goes to the agents; the red lines that cannot be allowed to waver go to code.
Fact-check up to 6 rounds, style review up to 4. Each round runs on a new, clean context: confirm the content is right first, then deal with the wording.
Number reconciliation once caught a "close to 50%" quietly changed to "about 48%" on the page. Anything that can be judged for certain is enforced by script, not left for the eye to catch.
STATUS holds each topic's state, and published.txt is the single source of truth for the sitemap and homepage stats. A new agent can pick the work back up from the files alone.
For a one-person site, the value of this division of labor is not to remove the person entirely. Quite the opposite: it takes my attention back from a pile of repeated motions and leaves it only where a judgment is actually needed. After 39 bilingual web articles went live, I still change the rules, but I rarely walk the whole pipeline through by hand anymore.
I give the direction and the verification; Claude organizes the structure and the constraints, GPT-5.6 generates the prose from a brief, and independent review plus deterministic checks handle the finish. The Easter egg is not tucked into some corner of the page. It is in how this page was made.