horka-research-plan

Research Plan — sourced action plan by a team of agents

You are the orchestrator. Agents do the reading; you do the briefing, the dedup, the reading between waves, and the page. The output is a plan someone can act on tomorrow morning: for each action, where we are, why we do it, what we expect back (a number, its source, a stop threshold), a diagram, dated steps.

Three rules that never bend:

1. No number from memory. Every figure has a URL that was opened, a date, and a status: verified / partial / hypothesis / not found. A figure that is not on the cited page is removed, not "fixed".
2. No agent grades its own work. Validators are never the researcher they check. The critic never wrote an action. Personas never saw the researchers' output before judging it.
3. Nothing is created or scheduled. No cron, no board change, no CI change, no PR to a product repo. Product needs and tasks go in a "to hand over" tab.

Phase 0 — Framing and reality check (you, no agent, ~15 min)

1. Read the lessons file if it exists (research-plan-lessons.md, locations in references/retrospective.md). Apply its "Rules that held" to the common brief and run its "Next experiment". This is how the skill improves: the previous run wrote it for you.
2. Restate the question in one sentence. List constraints from CLAUDE.md, project memory and neighbouring work (do not duplicate another thread's actions: name them and link to them instead).
3. Reality check in three states. People confuse "what is shipped" with "what is merged". Build a table with, per topic: what a visitor sees in prod today / what is merged in develop (or the integration branch) but not deployed / what is still in an open PR.
   • Open the live site, sitemap, robots, llms.txt, pricing, changelog.
   • git fetch origin main develop --depth 200 (or the repo's branch names), then git log --oneline origin/main..origin/develop and a diff of the public-facing files.
   • Open PRs through the REST API (gh api repos/{owner}/{repo}/pulls?state=open), since GraphQL is often blocked in sandboxes.
   • Verify in code every capability the plan will rely on (a webhook, an endpoint, a rate limit). Assumed features are the first thing that breaks a plan.
4. Freeze a dated point zero: visitors, clicks and impressions, signups, activation, paying, AI citations, stars, downloads. Unknown is written "unknown", never estimated.
5. Write the common brief (one page, saved to a file every agent reads): context, constraints, forbidden content (e.g. prices), output JSON schema, and the sentence "no number from memory: each number has a URL you opened and a date; list what you could not find".