PILLAR Generator — pipeline orchestrator (v3.3)

Aktuell für September 2026

Lizenziert

Verfügbar in DE

Schnelle Auszahlungen

Nur 18+

The pipeline generates one SEO-optimized iGaming page based on the variables in variables.md. The final artifacts are index.md (a Markdown article with embedded image placeholders) and a set of promptN.txt files (prompts for illustration generation), lying in the task root next to variables.md.

A note on Russian section names in variables.md

The section headers inside variables.md (e.g. ## РЕЖИМ, ## KEYWORD (HEAD), ## ПУЛ КЛЮЧЕЙ, ## ROLE_IN_SITE, ## LINKING, ## OUTPUT_FILENAME) are intentionally kept in Russian. They are technical identifiers used by the upstream site-builder pipeline (which generates these files automatically as part of multi-page sites). Do not translate or rename them. Treat them as opaque labels — the agent parses values by these exact labels.

The values written under these labels can be in any language depending on the task: ## ЯЗЫК СТАТЬИ may say DE, EN, IT, JP, etc., and the article will be written in that language.

Commands

• „run the full pipeline“ — execute stages 00→06 in sequence with automatic validation between them.
• „run stage NN“ — execute a specific stage (NN from 00 to 06). Used for debugging or re-runs of an individual stage.

After receiving a command, the agent does not ask clarifying questions and does not request confirmations. All decisions where the source instructions say "choose yourself," "select," "determine" — are made by the agent autonomously inside the relevant stage.

Stage registry

Each stage is described in its own file with inputs, actions, outputs, and a gate criterion of readiness.

Workspace

• _workspace/ — folder for intermediate artifacts of stages 00–05. Created by the agent at the start of the first stage, located at the task root next to CLAUDE.md. Never deleted automatically under any circumstances. After the pipeline finishes (regardless of whether the outcome is success or a validation failure) the folder stays on disk. Site build scripts on the developer side ignore it by the rule „files and folders with leading underscore — not for production“. The employee or developer removes it manually if they decide it is no longer needed.
• index.md — the final article with a fixed name. Placed at the task root next to variables.md. The name index.md does not depend on OUTPUT_FILENAME or slug — it is always the same, because the task folder itself carries the page name through its path in the site structure.
• prompt1.txt … promptN.txt — prompts for generating the article’s images, placed at the task root. Naming is preserved from _workspace/.
• <name>.meta.json — sidecar with page metadata (slug, title, headline, description, keyword, clusters, H2s, the formulated intent, outgoing links). Placed at the task root. <name> is taken from the OUTPUT_FILENAME: field in variables.md (primary mode, in a site context) or from the slug field in the index.md frontmatter as a fallback (for one-off runs). Used by future stages 7+ (site structure, internal linking, homepage). Not opened by the employee.

Master principles (priority over all specific rules)

1. Solving the user task from _workspace/00_intent.md. Not merely formal closure of pool clusters. A concrete semantic outcome that the reader should walk away with after entering KEYWORD into search. All subsequent stages (plan, writing, images, validation) serve this task. Formal closure of all clusters is a means, not an end. If a pool cluster does not help solve the user task, it is closed compactly (a single paragraph or a row in a summary table) so as not to dilute focus.
2. Substantive treatment of named options. If the user task implies an applied answer (where exactly to play, which sites accept X, which providers are available), then the concrete entities (operators, brands, legal entities, sites, providers, payment systems) that answer this query receive a dedicated H3 breakdown of at least 2 substantive paragraphs each. This rule outweighs the formal „TOP-10“ template: the number of breakdowns is dictated by the number of relevant options in the GEO (from 1 to N), not by a fixed figure. Mentioning an option only in a summary-table row without an H3 breakdown in the body is a failure, not an option.
3. Information gain. At least one original angle, calculation, or connection of facts that is not present in the competitors‘ TOP-10.
4. No padding. If the topic is honestly covered in three paragraphs under an H2 — three are written. Stretching paragraphs with repetition, generic statements, or paraphrasing the same thought is forbidden.
5. Specificity. Names, dates, figures (only verifiable ones), regulator names, products, legal entities. No abstract „many experts believe,“ „people often say,“ „usually.“

If a specific rule below conflicts with these principles, the principles take priority. In particular, principles 1 and 2.

Meta-contamination guard (strict)

The final article body and image alt tags never contain words and phrases that expose the machine nature or the back-office machinery: "cluster," "semantic core," "focus keyword," "information gain," "SEO," "LSI," "PAA," "SERP," "intent," "prompt," "AI," "GPT," "Claude," "neural network," "model," "according to the brief," "as required," "as part of the task," "target audience," "content plan," "article structure," "in this article we will examine," "as mentioned above," and any of their equivalents in ЯЗЫК СТАТЬИ. The final text reads as material from a human expert written for a reader, not for an algorithm.

Agent operating mode

• No file or folder deletions. The agent never performs delete operations (rm, rmdir, del, Remove-Item, unlink, os.remove, shutil.rmtree, fs.rm, fs.unlink, „cleanup,“ „tidy-up,“ and any of their equivalents) — not automatically, not on its own initiative, not at the end of the pipeline, not on validation failure, not as „cleaning up after itself.“ This applies to everything: contents of _workspace/, service temporary files, empty folders, duplicates, anything at all. All deletions are done by the user manually at their convenience. The rule is unconditional: if a stage or checklist says „delete X“ — that item is ignored, the files stay on disk, the user cleans up manually when needed.
• Quiet operation. The chat receives only short one-line reports on each stage’s completion. No long retellings of the content. The final message is a single line with metrics plus a present_files call.
• Quiet gap handling. If during research a figure cannot be verified, it is silently replaced with a qualitative wording. No reporting in chat.
• Automatic validation. After completing each stage, the agent checks against that stage’s gate criterion. If the gate is not passed, the stage is re-run.
• Cross-checking with the intent. In stages 03 (plan), 04 (writing), and 06 (final validation) the agent must re-read _workspace/00_intent.md before beginning work. The intent artifact is required context for all substantive stages, not reference material.

Stage error-handling logic

1. After completing a stage, the agent applies its gate criterion.
2. If the criterion is passed, move to the next stage.
3. If not passed, the stage is re-run with an explicit pointer at the failed item. Already-completed previous stages are not re-done.
4. Limit — 3 attempts per stage. If the gate is not passed three times in a row, the pipeline halts with a message of the form: „Stage NN failed validation on item X after 3 attempts. Intermediate artifacts in _workspace/ are preserved for debugging.“

Universality

The pipeline works for any iGaming topics in any GEO: legal casinos, TOP operators, provider reviews, payment methods (PayPal, Neteller, Skrill, Klarna, paysafecard, etc.), game types (slots, live, roulette, blackjack, baccarat, crash), bonuses and free spins, licenses and regulators, mobile apps, responsible gambling, new casinos, specialised picks. The specific task, form, and set of clusters are determined by the variables in variables.md, the intent formulation in _workspace/00_intent.md, and the keyword pool.

Search intent across topics is fundamentally different. For one keyword the user wants a TOP-list of operators. For another — a map of where a specific payment method works. For a third — a regulatory rules breakdown. A universal "review pillar" template does not work across topics. That is why stage 00 (intent) is the mandatory first step on every task.

What to do right now

If the variables in variables.md are filled in, wait for the command "run the full pipeline." On receipt of that command, execute the stage registry top to bottom, checking against gate criteria.

Verfasst vom Team von „Deutsche Slots Online”.