Agent/seo Bot

seo-bot — Cross-Product Article Generator

Spawns N×M×K fresh claude -p workers, one per (TA × promo × product) combo. Each worker writes one article in isolation; the orchestrator just loops + uploads via lc_api.sh.

Architecture (same shape as agentic-hive)

Project root (e.g. seo-system/)
├── .env                   # LC_API, LC_USER, LC_PASS, ARTICLES_TABLE_ID, …
├── .env.example           # config template
├── lc_api.sh              # sourced from developing/lattice-cast skill
└── seo-bot/
    ├── run.sh             # entry point (sources .env, calls orch)
    ├── orchestrator.sh    # pure shell — reads CSVs, loops product
    ├── worker.sh          # bash infra + `claude -p` for ONE article
    ├── audiences.csv      # TA personas
    ├── promotions.csv     # promotions / campaigns
    └── products.csv       # featured products

Config contract

.env must define:

| Var | Purpose | |---|---| | LC_API | Base URL, e.g. http://localhost:13491/api/v1 | | LC_USER | Login user_name for /login/password | | LC_PASS | Login password for /login/password | | ARTICLES_TABLE_ID | Table ID for the articles table | | TITLE_COLUMN_ID | Column ID for the title field in articles |

How it works

1. bash seo-bot/run.sh → sources .env, runs orchestrator
2. Orchestrator (rule-based, no LLM):

• Parse audiences.csv, promotions.csv, products.csv (header row + title,description)
• Build cross product (3 × 4 × 2 = 24 typically; generic for any size)
• Query existing rows via lc_row_list on ARTICLES_TABLE_ID
• Skip combos already covered (idempotent — safe to rerun)
• For each missing combo, exec worker.sh with the 3 (title, desc) pairs
• Sleep 1s between workers (rate-limit per CLAUDE.md)

1. Worker (bash + LLM):

• Compose a prompt embedding the three contexts
• timeout 120 claude -p --dangerously-skip-permissions ""
• Worker writes ONE markdown file to ~/.tmp/article-.md
• Bash creates row via lc_row_create then uploads doc via lc_doc_write
• No LLM in the upload path — bash handles it deterministically

1. Orchestrator logs in, derives LC_AUTH_HEADER at runtime, and emits

ALL_DONE when count reached; otherwise exits non-zero

Each worker is a fresh process — no shared context between articles.

CSV format

UTF-8, comma-delimited, first row is header. Two columns: title,description.

title,description
akai,26-year-old male office worker. WFH 3d/wk. Loves bouldering.
meiwen,22-year-old female grad student. Industrial design. Tight budget.

Title becomes the article-name component (used for the unique articles.Title like akai-spring-2026-ergopro-chair). Description is fed to the LLM as persona/context.

Article-title contract

Worker uploads with Title = "${a_title}-${b_title}-${c_title}".

This title is the dedupe key and the verifier's signal — it must be unique across all combos. Don't change the format unless you also update orchestrator's contains skip-check.

Worker rules

• One combo only. Don't mention any other persona/promo/product.
• ≥400 chars of substantive prose. The driver enforces this.
• First line is # .
• No filler / no echoing the prompt.
• Don't curl from inside the worker — bash handles upload via

lc_row_create + lc_doc_write. Same rule as agentic-hive.

Rate limit

sleep 1 between workers in the orchestrator. From seo-system/CLAUDE.md: > API rate limit: wait 1 second between each API call

Idempotency

Re-running run.sh after partial completion picks up where it left off — already-uploaded titles are skipped via lc_row_list.

Composition

Sources lc_api.sh from developing/lattice-cast skill (the thin LatticeCast HTTP wrapper). Project .env provides LC_API, LC_USER, LC_PASS, and table/column IDs. The bot logs in and derives LC_AUTH_HEADER at runtime. The bot itself contains no LatticeCast-specific code beyond the lc_* function calls.

When to use this skill

• You have 3 source dimensions and want every cross-product article

written without context leak between articles.

• You want deterministic, idempotent, rerunnable batch generation.
• You want cheap parallelism: each claude -p is independent.

For single-article writing: just use claude directly in the seo-system project — pilot Claude reads CLAUDE.md, runs get_context.sh, and writes ad-hoc.

Reference scripts

[example-scripts/run.sh](example-scripts/run.sh) [example-scripts/orchestrator.sh](example-scripts/orchestrator.sh) [example-scripts/worker.sh](example-scripts/worker.sh) [example-scripts/audiences.csv](example-scripts/audiences.csv) [promotions.csv](example-scripts/promotions.csv) [products.csv](example-scripts/products.csv)

Copy these into /seo-bot/, replace CSVs with your real sources, ensure /.env defines LC_API, LC_USER, LC_PASS, ARTICLES_TABLE_ID, TITLE_COLUMN_ID, and that lc_api.sh is available to source. Then bash seo-bot/run.sh.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: LatticeCast
• Source: LatticeCast/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.