Operator-grade methodology for AI-driven publishing.
Leefan Reports is a publishing-operations practice for teams running programmatic SEO and AI-assisted
content at scale. We publish three open libraries - OpsLab, ToolLab,
DiagKit - and are opening a small number of first-quarter implementation slots.
The work is specific: card-shaped instructions for AI runtimes, an audit log that survives staff turnover, and
diagnostics that catch silent failures before they reach production.
The schema and redacted pattern behind the decision log: what shipped, what broke, what rolled back.
The opposite of “trust the runtime’s self-report”.
The three columns are deliberately uneven in tone. OpsLab is the practice; the regression DB is the
receipts; DiagKit is the gun on the table - small, usable artifacts a reader can apply tonight
without us in the room.
Methodology
[VISUAL SLOT: METHOD-01] horizontal process diagram, 5 stations
The Leefan Reports loop, in one sentence
We treat every AI-assisted operation as a card → run → verify → log → rollback-ready
unit. The loop is boring on purpose. Boring loops are the ones that survive a 200-page build.
What the loop forbids
We do not paste prompts into a chat window and call it a workflow. A workflow lives in a tracked file with an
ID, a hard gate, and a named output.
We do not trust an AI runtime’s “done” without reading the output against the original card.
Self-report is a hypothesis, not a result.
We do not ship 200 templated pages with one prompt. The same prompt produces the same template smell; the
regression DB exists so we know which deltas matter.
We do not buy paid SEO tools, paid stock images, or ranked link placements as a substitute for an audit log.
Tools accelerate a working loop; they do not create one.
What the loop produces
The card - a versioned instruction that a junior operator or a second AI runtime can re-execute.
The output - the actual pages, datasets, or configurations that ship.
The audit entry - what was decided, what was rejected, what to watch for in week two.
This is the same loop we run on our own libraries. The OpsLab pages on this site were each written as a card.
Side-by-side comparisons run on our own publishing operation. Each page states the use case, the weighting,
the posture, and a counter-case where the loser is the right choice.
T01 — Claude Code (terminal) vs Cursor (IDE) numeric rerun pending
S01 — Firecrawl vs Browserless numeric rerun pending
Every library page is written under the rule that someone not on a Leefan Reports
engagement should be able to read it, copy what they need, and walk away.
Who this is for / who it is not for
[VISUAL SLOT: FIT-01] small two-column “right fit / not right fit” graphic
Tends to fit when…
A team is already publishing 30+ pages per month and wants to ship 10× without 10× the headcount.
A team has tried Claude / Cursor / agent CLIs and has 40 stalled prompts in a doc nobody reads anymore.
A content site that was growing has plateaued, and the root cause is not visible.
The team needs an audit log because the work is regulated, sponsored, or audited by a parent org.
Wrong fit when…
The brief is “write our first ten blog posts”. Hire a writer.
The brief is “rank us for a head term in 90 days”. We do not guarantee rankings.
The team wants to outsource thinking. The card practice requires an in-house operator. We can train one;
we do not replace one.
Service inquiry
[VISUAL SLOT: INQUIRY-01] single calm visual, “evidence packet” still life
Engagements
We are opening a small number of paid first-quarter implementation slots. The shape of the work, the indicative
pricing band, and the scoping process are on the services page.
There is no contact form on this page on purpose. The inquiry surface is one of these two:
Email — contact@leefan.co.jp
(address shown as plain text; no form, no auto-responder; the intake mailbox exists; no mailto link is used)
A scheduling page — gated by the public-launch approval; no public scheduling link is used.
If you write to us, please include:
one paragraph on what your team is currently shipping
one paragraph on the specific bottleneck
whether you already use Claude Code, Cursor, or an agent CLI
a rough sense of timeline
We read inquiries personally. We reply within two business days, or we reply with a short note declining the
engagement. We do not place inquiries onto a drip sequence.