--- name: threadify-research-army description: Get a research post on your niche. An army of Haiku 5.5 agents and one Opus 5.5 race to research the most-liked recent Threads posts in your niche; a script checks every claim; Threadify writes the post from the winner and schedules it after one yes. Use when someone says "research army", wants a research post about their niche, asks what the most-liked Threads posts in their niche have in common, or wants to race cheap agents against a big model. --- # Research Army Get a research post on your niche. An army of Haiku 5.5 agents and one Opus 5.5 race to research the most-liked recent Threads posts in your niche; a script checks every claim; Threadify writes the post from the winner and schedules it after one yes. It is free: the race runs in Claude Code on this computer (it uses your Claude plan's usage, a few cents to about a dollar of model time). Two things need your own accounts: the corpus comes from ScrapeCreators (about 25-30 credits; new accounts get free credits) and the post step runs on Threadify. Follow [Threadify-001: setup and first-loop video](references/threadify-001.md) before the first connected call. A connection never authorizes scheduling. Only the owner's exact "yes" to the post card does. Resolve `scripts/` and `references/` paths against the directory containing this SKILL.md. The script is `node scripts/research-army.mjs` (Node 18+, no npm packages); every command prints JSON. The race needs the `claude` command line (Claude Code) on this computer. ## Rules - Follow [demo feedback rules](references/demo-feedback-rules.md) for the whole run. If `~/.threadify/owner-overlay.md` exists, read it too and follow it; it is private, so never quote it into posts, files you publish or receipts. - **Honest label, everywhere.** Say exactly what the corpus is, from the script's `label`: "the N most-liked recent Threads posts from K keyword searches (last 30 days)". Never say "most viewed" (there are no view counts) and never "the top posts on Threads" (keyword search is a sample, not a ranking). - **A fair race.** Both sides get the same brief (the script writes it once and prints its sha256), the same posts and Read-only tools. Neither side gets Bash or any code tool: the script runs both with `--tools`, `--disallowedTools Bash`, `--permission-mode dontAsk` and `--restricted` when the installed Claude Code has it, and then proves it from each run's own log (`bash_offered: false`, `bash_uses: 0`). Never change the brief, the prompts or the catalog to favour either side. - **The checker owns the tests.** Each side picks 8 tests from the script's fixed catalog and reports a count and 2 example quotes for each; the script recounts every one over all posts. A claim is backed only when the recount equals the claim. Winner: most backed claims, then lower cost, then faster. - **Numbers in the post come from the recount**, and only from the winner's backed claims. Never quote other creators' posts or handles in the post. - **The ScrapeCreators API key is never printed**, repeated, put in a command line or written anywhere but the owner-only header file from `scripts/scrapecreators-key.mjs save`. - **Posting.** Threadify writes the post with `generate_content`. Show it in full. It must pass `research-army.mjs check-post` (honest source line, only checked numbers). The agent may only change casing (lowercase, proper nouns kept) and proves `lower(original) == lower(final)` with `scripts/casing-guard.mjs`; edits are the owner's, made in the editable block. The tracked link goes in one Auto Plug the agent writes from the post's findings and the offer's approved facts, 50 minutes after the post. Never publish now. Schedule only after the owner's exact "yes", then read it back. - **Opening pages.** Use the host's browser tool when it has one; otherwise run `open -a "Google Chrome" ` (macOS). Confirm it opened. ## Steps 1. **Account and niche.** Call `get_connection_defaults` first. Say only "@handle · timezone" and the plan's quota if it's shown. Read the niche from the Threadify Brain (`get_brain_overview`, or `query_brain` "what is my niche") and confirm it in one line: "Niche: . Say a different one, or I'll go with this." Tell the owner in one line: "20 Haiku agents and one Opus race to research the most-liked Threads posts in your niche this month; a script checks every claim; Threadify writes the post." 2. **Set up ScrapeCreators once.** Run `node scripts/scrapecreators-key.mjs status`. When it says `missing` and the SCRAPECREATORS_API_KEY environment variable is already set on this computer, save it without showing it: `printenv SCRAPECREATORS_API_KEY | node scripts/scrapecreators-key.mjs save`. Otherwise send the owner one message in plain words: "Research Army reads Threads posts through ScrapeCreators, a data service you pay for yourself. One-time setup, about two minutes: 1. Sign up at https://app.scrapecreators.com. New accounts get free credits. 2. Copy your API key from the ScrapeCreators dashboard. 3. Open Terminal (Mac) or PowerShell (Windows), paste this line, press Enter, then paste your API key when it asks. It stays hidden and is saved only on this computer: `node "" save` Tell me when it says saved." Fill in the real absolute path. If the owner pastes the API key into the chat instead, pipe it to `node scripts/scrapecreators-key.mjs save` through stdin (a heredoc, never a command-line argument), never repeat it, and say once that it is now in this chat's history. Then run `node scripts/research-army.mjs balance` (free) and say "ScrapeCreators connected · credits left". On `scrapecreators_rejected`, ask the owner to copy the API key again and run save again. Never guess a balance. 3. **Pull the corpus.** Make a run folder `./research-army//`. Write `keywords.json` as `{"keywords": [...]}` with 25-30 short search phrases a creator in this niche would actually post about (the niche itself, its main topics, tools, outcomes and problems; 1-3 words each; no near-duplicates). Run `node scripts/research-army.mjs estimate --keywords /keywords.json` and say " searches, credits, you have ." Then `node scripts/research-army.mjs corpus --keywords /keywords.json --dir --niche ""` (searches only the last 30 days, at most 40 credits; reuses saved searches). Say the label and the counts from its output: posts pulled, duplicates, older than 30 days and non-English removed, posts kept, the likes range and the dates, and the credits before and after. Fewer than 40 posts kept: add keyword variants and pull again (new searches only) before racing. 4. **The race.** Show the live view: open a new terminal window in the run folder running `node "" watch --dir .` when this computer can (macOS: `open -a Terminal` or the owner's terminal app with that command), then start the race: `node scripts/research-army.mjs race --dir --army 20` (`--army N` for another size, 2-30). Run it in the foreground when your host streams command output; otherwise run it in the background and print `watch --dir --once` lanes every 15 seconds until it finishes. What happens, in one line each as it goes: - The script writes `race/brief.md` once (same text for both sides), splits the posts into 20 batches for the army and gives Opus all of them. - **Army:** a Haiku commander starts 20 Haiku scouts (`references/agents/research-army-scout.md`, `model: haiku`, `tools: Read`). Round 1: every scout reads its batch and nominates patterns from the catalog; the commander picks 8. Round 2: every scout counts those 8 in its batch; the commander adds the counts up, so no batch's count is lost. - **Opus:** one Opus reads all the posts alone, picks 8 from the catalog and counts them. - Each side has 5 minutes of silence and 12 minutes in total before it's stopped. It ends with the scoreboard: backed claims (exact, and within ±1), cost (`total_cost_usd`), wall time and the models read back from each run's `modelUsage`, plus the Bash proof and the winner. 5. **Check every claim.** Say the scoreboard in a few lines, then each side's claims as ": claimed X, recount Y, backed/unbacked". If `bash_blocked` is not true, or a side's `models` are not the expected Haiku/Opus models, say so plainly and do not call the race fair. If a side failed or was stopped, say so; its claims count as unbacked. Open the report: serve the run folder on this computer only (for example `python3 -m http.server 8050 --bind 127.0.0.1 --directory ` in the background) and open `http://127.0.0.1:8050/scoreboard.html` (browser tool, or `open -a "Google Chrome" http://127.0.0.1:8050/scoreboard.html`), because some browsers refuse `file://` pages. Stop the server when the run ends. `node scripts/research-army.mjs check --dir ` rebuilds `check.json` and `scoreboard.html` from the saved findings without re-racing. 6. **Threadify writes the post.** Run `node scripts/research-army.mjs facts --dir ` for the winner's backed findings, then follow [the post step](references/post.md): one `generate_content` call, `check-post`, one Auto Plug with the offer's exact link (an offer named "Research Army" is used without asking), both in full in chat, the editable blocks with **Done editing**, and the casing guard. 7. **One yes, then schedule.** Follow [the post step](references/post.md): one post card at the best time tomorrow, `schedule_post` with `auto_plug: {trigger: "time", delay_minutes: 50, content}` only after the owner's exact "yes", then the readback of the post and its stored Auto Plug. Outputs stay in `./research-army//`: `keywords.json`, `raw/`, `corpus.json`, `race/` (brief, batches, run logs, `race.json`), `findings-army.json`, `findings-opus.json`, `check.json`, `scoreboard.html`, `post.txt`, `plug.txt` and `post.md`. ## When something is missing No `claude` command line: say the race needs Claude Code on this computer and stop after the corpus. No ScrapeCreators access: give the step 2 setup message and wait. Out of credits: say how many are missing and that the owner buys credit packs at https://app.scrapecreators.com (never buy for them). No Threadify connection or no generation access: stop after the scoreboard and say "Connect Threadify at threadify.app to have the post written in your voice and scheduled." Never write the post yourself instead. A scoreboard is not a post, and a scheduled post is not a published post.