Abstract

Reading is not using.

Takajo AI Lab exists to help people bring generative AI all the way into real work. The learning program is built as a single loop — lessons, submissions, reviews — designed together, not stitched together. It is not a video course. It is not a packet of slides.

Learning generative AI has to pass through the hand — write, run, read, revise — before it turns into capability. Reading along and nodding is a different activity. This program builds the iteration in as structure: every submission is actually executed, and the feedback is specific to what the execution produced.

The problems a rubric will not catch,
surfaced by running the thing.

Philosophy

A review begins with running it.

In most AI courses, a submitted prompt is read and nothing else. Someone looks at the text, praises the structure, issues a templated suggestion. What the student learns from that cycle is how to write a prompt that looks correct on the page — which is not the same as a prompt that works.

A review in this Lab begins by executing the submission on a current model. Only then do the real problems appear — the kind where, for instance, a prompt generating historical records drops the character for "Soga" in "the Soga clan" during the 645 Taika Reform entry, because the model is running out of breath in a bulk-generation task. These are things a rubric cannot see.

Reviews of this density are expensive to produce, which is why Premium — the human-reviewed plan — is restricted to companies and capped at five teams (currently full; an opening is possible in about three months). For individuals, the review process is being automated as the AI Review plan, which will launch once CriticChain is tuned.

The standards discovered in human review are fed back into the design of the automated review — teaching is, in the end, how this lab does research. That circuit is the shape of Takajo AI Lab.

Two plans

Two review plans.

Two plans. The curriculum and assignments are shared; what differs is who the plan is for and who writes the review. Premium is human review for teams; AI Review is automated review for individuals.

Premium Full For teams / human review Hatanaka reviews every submission in person. Prompts and code are actually executed against a current model; the feedback covers not only pass/fail but what to learn next. Capped at five teams. Companies only. Currently full; an opening is possible in roughly three months, so teams that are interested are encouraged to make contact in advance.
AI Review Soon For individuals / CriticChain Automated review powered by CriticChain (published under AGPL-3.0). A multi-stage pipeline in which one AI audits another's review as "too lenient" or "poorly grounded", and requires revisions until the review itself meets a standard. For individuals; available once CriticChain's tuning is complete.

Real reviews

From the review archive.

Even a passing submission comes back with this density of feedback. The excerpts below are drawn from Premium reviews, anonymised.

01 — Executing the prompt

Submissions are run. That is where the review starts.

Assignment — improving a bulk-generation promptVerdict: Pass

What was strong.
The prompt is organised under clear headers — role, context, request — so the model can parse the instructions without ambiguity. An example output is provided with the expected format and level of detail, which keeps output variance low.

Suggestion: the model "running out of breath" on bulk tasks.
Running your prompt against a current model, I found that in the entry on the 645 Taika Reform, the character for "Soga" was dropped — the text produced "Shi clan" where it should have said "Soga clan". This is a textbook hallucination triggered by the model being forced to produce too much in a single pass.

Tasks of that volume should not be run in one go. Break them into steps and inject Human-in-the-Loop (HITL) checks at intermediate points. That is the pattern to reach for here.

— The reviewer actually executed the submission and found a specific hallucination. A problem no rubric could catch, surfaced by running it.

02 — Beyond the pass mark

It passed. Here is the next wall.

Assignment — a self-verification loopVerdict: Pass

What was strong.
The implementation cleanly separates drafting, back-translation, verification, and revision — and crucially, it keeps the intermediate outputs visible rather than collapsing them. It is possible to see where in the chain the model's reasoning happened, which is the point (the glass-box approach).

Suggestion: split the session to avoid context contamination.
The implementation meets the basic requirements perfectly. However: the model has context memory, which means it is reviewing its own text while remembering the intent it wrote that text with. Just as a human proofreading their own writing tends to miss things, a model self-evaluating in the same session is biased toward reading its own output generously.

When stricter verification is required, split generation and verification into separate sessions, so the verifying model is blind to the writer's intent. That is the technique.

— A perfect implementation is graded as a pass, and the next real-world wall (context contamination) is shown ahead of time.

03 — A growth path

Beginner to advanced, as a route.

Assignment — role-setting and verificationVerdict: Pass

What was strong.
Running the task with a role assigned and then without, and comparing the two — that comparative posture toward LLM behaviour is exactly right, and exactly what prompt engineering actually is.

Suggestion — three steps.

1. Structure. Separate role, goal, and output format under Markdown headers.
2. Observability. Emit the intermediate work so the model's criteria are visible (glass-box).
3. Role experiments. Try several different roles and notice that more than the tone shifts — the substance of the advice changes too.

— One assignment, two things at once: the fix suited to the current level, and the growth path that continues past it.

AI Review pipeline

How CriticChain works.

AI Review's engine — CriticChain — is a multi-stage review pipeline. A draft review is written by one AI; another audits it and forces revisions until the review clears a standard.

Submission Lint Structural analysis Hallucination detection
Draft review Critique (leniency audit) Refine (rewrite)
Consistency check Scoring Final review
  • One AI's review is audited by another as "too soft" or "insufficiently grounded"
  • Hallucination detection goes beyond the fact itself, tracing the propagation path
  • Every stage is logged; the grounds of the verdict are readable

CriticChain works against the criteria defined in the prompt-standard prompt-as-code. Define a standard, then review against that standard automatically — that coherence is what the quality rests on.

Curriculum

23 lessons, 5 phases.

From the fundamentals of generative AI to building systems that hold up in production. The axis is not "keeping up with the buzzwords" but turning AI into something a team can actually run on.

Phase 01
Foundations How generative AI works, how LLMs fail, choosing between the major tools
Lesson 01–03
Phase 02
Environment Python setup, prompt engineering, building custom skills
Lesson 04–06
Phase 03
LLM implementation LLM APIs in practice, LangChain applications, tokens and cost
Lesson 07–11
Phase 04
Applied RAG, AI agents, fine-tuning, multimodal, UI construction
Lesson 12–18
Phase 05
In production Evaluation and quality, ethics and governance, career design, capstone project
Lesson 19–23

Who it is for

The two audiences.

The two plans address clearly different audiences. It is worth deciding, here, which one you are.

Premium For teams
  • Companies considering in-house training or organisational rollout
  • Leadership and business units who want an organisation that can handle AI, not just individuals who can
  • Teams that need high-density review where the submission is actually executed
  • Organisations that want the prompt-as-code and CriticChain thinking internalised

Currently full (max 5 teams). An opening is possible in about three months. Interested companies are encouraged to make contact in advance. Can be combined with the technical advisory / AI Strategic Partner engagement → Work

AI Review For individuals
  • Engineers who want to take generative AI all the way to running in their day job
  • People who are done using AI tools and want to understand them as systems
  • Anyone moving from prompt engineering toward multi-agent development
  • Anyone who wants to submit at their own pace and get fast, specific feedback

In preparation — launches once CriticChain's tuning is complete. Register for launch notification at info@1stpiece.io.

FAQ

Questions that come up.

What is the difference between the two plans? Premium is for teams — Hatanaka reviews submissions personally, after actually executing them against a current model. AI Review is for individuals — CriticChain returns automated review immediately. Curriculum and assignments are shared across both.
Is Premium open? Currently full (companies only, capped at 5 teams). An opening is possible in about three months; interested teams are welcome to reach out in advance and we will respond in order.
When does AI Review launch? In preparation for individuals. Launches once CriticChain's tuning is complete. Write to info@1stpiece.io and we will notify you at launch.
Is it a watch-the-videos-at-your-own-pace course? No. The program is designed around an iteration: study, submit, get reviewed, revise.
Can individuals take it now? At this moment, only AI Review (still in preparation) is planned for individuals. Premium is restricted to companies. We recommend registering for the AI Review launch notification.
Is CriticChain open source? Yes — AGPL-3.0. GitHub: taka4rest/CriticChain

Start with an inquiry.

For Premium availability, AI Review launch notifications, or a company pilot — we respond case by case.