Find who to pitch. Never send it for you.
Give your agent one plain request, such as "find podcasts worth pitching and draft the pitch". A small team of agents finds who is worth approaching, narrows the list, studies the best fits, drafts the pitch and has the draft reviewed. You get a short summary to accept or reject, and an email draft that only you can send.
Pitching is mostly research: who is a fit, what they have covered lately, how they like to be approached. This blueprint does that research in the open. Fifteen targets become five, five become three studied in depth, and one is picked. Every claim about a target carries the page it came from and the date it was read, so you can check it before you trust it. The research is kept, so the next run starts ahead instead of from nothing. The worked example pitches podcasts; the same steps fit newsletters, partners, events and press.
Three research dossiers, a one-page pitch packet for the strongest target, and a pitch email saved as a draft file. Plus the working papers, a checker, prompts for every role, and a finished example run for an invented business.
No link, no claim. What could not be found is written down as not found. A pitch with one unsupported fact is held, however good the rest of it is.
A lead plans, a researcher finds and scores, a voice study reads how the host talks, a writer drafts, a reviewer scores the drafts. One agent can play every role in turn; run the review in a fresh chat.
Answer the brief's six questions, including where the evidence may be kept. Accept or reject at the end. If the email goes out, you send it.
The funnel, fifteen to one.
- Longlist. Fifteen targets worth a closer look, each with a link and a reason.
- Shortlist. The five highest scores. The criteria and their points are set before anything is scored, and add up to 100. Anything not found scores 0.
- Dossiers. The three strongest: what it is, two recent items close to your topic, how they talk, how to reach them.
- The pick. The strongest target that is ready, on one screen with the other two.
- The drafts and the review. A pitch packet and an email. The reviewer scores each: a pass needs 80 or more out of 100, and full marks on factual support. A draft that fails goes back once; if it fails again, it is held for you.
A show is not taking guests. A host has too little on record to study. A draft will not pass review. Each is written down as a hold, with the reason, and shown to you as a hold. The worked example has two.
What happens on the first run?
Your agent copies the blueprint to work/research-to-marketing, then proves the checker works on the finished example before touching your brief.
python3 scripts/research_check.py example/run python3 scripts/research_check.py run --up-to brief
On the example, the last line starts with READY FOR THE OWNER and the two holds are printed. On your run, each step should end OK SO FAR; anything marked problem: gets fixed before the next step.
If the r2c skill is in your workbench, the researcher uses it for every research step and it does the saving. Without it, your agent uses its own web search and keeps the evidence in a private folder you approve. Either way, nothing needs a new account.
What you can ask your agent, once it is set up
PROMPTS.md has one prompt per role and a single-agent version. These are for along the way.
Run the research checker on my run folder and tell me, in plain sentences, what is missing or does not agree.
Show me the top three and why the pick won. For each, name the one fact that mattered most and the page it came from.
Read corrections.md and tell me, one line each, the rules you will follow on this run.
List every claim in the pitch email with its source link and retrieved date, and mark any that rest on a single source.
Words you now own
- Longlist, shortlist. Fifteen candidates, then the five that scored highest.
- Dossier. One target in depth, every claim with its link, date and evidence card.
- Pitch packet. One page, 500 words or fewer, for the target you picked.
- Evidence card. One per source: what it says, its link, the date, what it is good for. The original page is kept untouched beside it.
- Source index. The running list of every saved source. The researcher reads it first, so the next run only fetches what is new.
- Corrections file. Everything you have corrected, written as one rule each. The lead reads it before planning every run.
Before you rely on it
- The checker counts, it does not judge. It checks every claim has a link, a date and a card. It cannot tell you a link is right or a pitch is good. You click the links, and you decide.
- Shared or synced workbench? Name a private folder. The suggested evidence folder sits beside your run. If other people can see your workbench, keep the evidence somewhere they cannot.
- Sensitive run? Write "no saving". The research still runs with links and dates; nothing is kept, and the roll-up says so.
The email is a draft file for you to read. If it goes out, you send it.
Fair questions
No. You say what kind of target in the brief: podcasts, newsletters, partners, events or press. The steps are the same.
This one finds who to approach and drafts one pitch. The campaign review workflow turns one brief into a week of posts. Both use r2c for research when you have it.
Yes. It is files and prompts, and the checker is plain Python. Paste the same prompts into either app.