Research Instructions for Subagents
These instructions apply to every research subagent working in this repo. Read them before starting and follow them throughout.
1. Scope of the project
The repo is a personal research library on the food and drink a strength athlete actually buys in southeastern Wisconsin in 2026 — and where it comes from.
The audience is a lifter in the Milwaukee area who shops at Pick 'n Save, Metro Market, Trader Joe's, Woodman's, Costco, Aldi, Kwik Trip, and the rest, and who wants to know three things:
- What is actually in this — residues, contaminants, additives, and whether the testing data supports the worry or dismisses it.
- Where did it actually come from — not "California," but the growing region, the feedlot, the slaughter plant, the co-packer, and the corporation whose name is not on the box.
- What went into it upstream — what the animals were fed and dosed with, what the plants were sprayed with, what the processing added.
The third question is the reason this project exists. A page that describes a food without tracing its inputs has not done the job.
In scope: produce (romaine and leafy greens first), beef, chicken, pork, eggs, dairy (milk, Greek yogurt, cottage cheese, RTD shakes), seafood, carb staples (rice, oats, potatoes, bread), protein powder, creatine, pre-workout, micronutrients and fish oil, protein bars, and the retail channel itself.
Out of scope: meal plans, macro calculators, training advice, and anything that reads like coaching. We describe what the evidence says. The reader decides what to eat.
2. Eight sections every dossier must answer
For every topic, the writeup covers, in this order:
Intro — concrete facts. No fluff opener. Tell the reader what this category is in one short paragraph.
History — chronological narrative. Told as a story with dates, places, and people. Prose, not a bullet list. How did this food come to be produced the way it is now? The interesting history here is usually industrial and recent: the rise of the feedlot, the invention of the broiler chicken, DSHEA 1994, the bagged-salad category, the ultrafiltration patent.
Supply chain — inputs to shelf. The spine of this project. Do not shortchange it. Trace the item backwards from the shelf:
- The shelf — which of the surveyed stores carry it, under which brand, and who owns that brand. Distinguish national brand vs. private label (Kroger's Simple Truth, Trader Joe's own label) and, where findable, name the co-packer who actually makes the private-label version.
- The processor — the slaughter plant, the dairy, the mill, the powder blender. For USDA-inspected meat and poultry, use the
EST./P-establishment number off the package and resolve it against the FSIS Meat, Poultry and Egg Product Inspection Directory. Name the plant and the town. - The farm / growing region — and the season, where origin rotates (romaine flips between Salinas, CA and Yuma, AZ).
- The inputs. What was the animal fed (corn, soybean meal, distillers grains, alfalfa, silage)? What was it dosed with (implants, ionophores, beta-agonists, rbST, antibiotics — cite FDA CVM's Green Book and the annual antimicrobial-sales report)? What was the plant sprayed with (cite EPA tolerances at 40 CFR 180 and USDA's Pesticide Data Program for what was actually detected, not just what is permitted)? What did processing add?
- Where the trace goes dark. Say so explicitly, and say why. "Trader Joe's does not disclose its suppliers, so the chain ends at the label" is a finding. Guessing the supplier is not.
Studies & nuance — what the peer-reviewed and regulatory literature actually says. Residue monitoring data. Contaminant testing. Outbreak investigations. Efficacy claims for supplements. Distinguish what is permitted from what is detected from what is harmful at the detected dose — these are three different questions and the popular coverage collapses them constantly. Where a claim is marketing rather than science, say so. Where studies disagree, say so. Where a scare is not supported by the exposure data, say that too — this project is not in the business of manufacturing worry.
Preparation, safety, storage — food safety (cook temperatures, cross-contamination, the specific hazards of bagged and pre-washed greens), storage and spoilage, and for supplements, dosing as reported in the literature — never as a recommendation.
Cost — real numbers, date-stamped, with the store named. For any protein source, always compute cost per gram of protein. This is the number the audience actually optimizes and almost nobody publishes.
Further reading — books, papers, documentaries, databases. Annotate each with 1–2 sentences.
Sources — numbered bibliography matching the inline
[N]footnotes.
3. Citation rules — non-negotiable
- Cite as you gather, AND as you write. Every factual claim gets an inline
[N]footnote. - Use this format in the sources list:
[3] Author Last, F. (Year). *Title*. Publisher / Journal Vol(Issue), pp. URL or DOI if available. - If a fact is widely repeated but you cannot find a primary source, write "commonly stated, primary source not located" — do NOT fabricate citations, DOIs, URLs, page numbers, author names, or years.
- Label trade/industry sources:
[12] (industry source — Tyson Foods),[13] (industry source — Glanbia). Corporate sourcing and sustainability reports are legitimate evidence about what a company says it does. They are not evidence that it does it. - Label advocacy sources too:
[14] (advocacy source — Environmental Working Group). EWG's Dirty Dozen is a real, citable artifact and also methodologically contested for ignoring dose. Both facts go in. - Date-stamp anything perishable: prices, brand recommendations, ownership, regulatory status. Write "as of [Month Year]".
Evidence hierarchy — prefer these primary sources by category
- Pesticide residues actually detected: USDA Pesticide Data Program (PDP) annual summaries; FDA Pesticide Residue Monitoring Program; FDA Total Diet Study.
- Pesticides legally permitted: EPA tolerances, 40 CFR 180; EPA registration-review dockets on regulations.gov.
- Drugs given to animals: FDA CVM Green Book (approved animal drug products); FDA's annual Summary Report on Antimicrobials Sold or Distributed for Use in Food-Producing Animals; USDA FSIS National Residue Program results.
- Who processed it and where: USDA FSIS Meat, Poultry and Egg Product Inspection Directory (establishment numbers); USDA NASS Census of Agriculture; Wisconsin DATCP licensing records.
- Who owns it: SEC filings (10-K, proxy statements). Ownership claims should cite the filing, not a news article about the filing, where possible.
- Outbreaks and contamination: CDC and FDA outbreak investigation reports; FDA recall notices.
- Supplements: 21 CFR 111 (cGMP); FDA warning letters; FDA CAERS adverse-event data; FDA's tainted-products database; the published work of Pieter Cohen et al. on adulterated supplements; NSF Certified for Sport and Informed Sport product databases; Clean Label Project and Consumer Reports testing (label these as the third-party testing organizations they are).
- Efficacy of a supplement: Cochrane reviews, systematic reviews and meta-analyses, and International Society of Sports Nutrition position stands. Not forums, not influencers, not the manufacturer.
4. Tone and voice
- Curious, plain-spoken, slightly skeptical — in both directions. Skeptical of the manufacturer, and skeptical of the wellness scare. The failure mode of this genre is a page that treats every detection as a poisoning.
- Numbers > adjectives. "PDP detected residues of X on N% of romaine samples, at a mean of Y ppm against an EPA tolerance of Z ppm [4]" beats "lettuce is covered in pesticides."
- Always give the reader the denominator. A detection at 1/1000th of the tolerance is a fact; it is not a hazard. Say which one you are reporting.
- Avoid fluff openers. Start with a concrete fact.
- It is encouraged to call out where popular wisdom is wrong — in either direction. "Hormone-free chicken" describes federal law, not the product. "Uncured" bacon is cured, with celery-powder nitrite.
- No moralizing. Do not tell the reader what to eat, what to avoid, or what is "clean."
- No medical advice. Report dosing as it appears in the literature. Never prescribe.
- No emojis.
5. Deliverables per topic
Each research subagent produces TWO files for its topic:
/research/<topic>/dossier.md— the full research notes, including leads that were chased and rejected, and an explicit list of every claim that could not be sourced. This is the long version./site/<topic>.html— a polished, standalone HTML page derived from the dossier.
HTML page requirements
- Semantic HTML5 (
<article>,<section>,<nav>,<aside>,<figure>,<table>). - Readable typography: serif body (~18px), line-height ~1.6, max-width ~760px for prose.
- Inline
<style>block. No external CSS, no Google Fonts, no CDN scripts, no tracking. System font stack for nav UI; serif for prose. - Section order matches the dossier: Intro → History → Supply Chain → Studies & Nuance → Preparation/Safety/Storage → Cost → Further Reading → Sources.
- Inline footnote markers:
<sup><a href="#src-3">[3]</a></sup>linking to the sources list. - Sources list at bottom: each item gets
id="src-N"and a back-link. - A small top
<nav>linking back toindex.htmland to sibling topic pages. - For comparison tables (supply-chain maps, cost-per-gram-of-protein, residue data), use semantic
<table>with<thead>/<tbody>/scope="row|col". Wrap wide tables in a div withoverflow-x: auto. - No emojis.
6. What NOT to do
- Don't pad. A 3,000-word dossier with real sources beats a 7,000-word one full of filler.
- Don't fabricate citations, even plausible-looking ones. This is the cardinal sin.
- Don't guess at a supply chain. If you cannot establish who makes a private-label product, write that you could not. An inferred co-packer stated as fact is a fabrication wearing a lab coat.
- Don't collapse permitted / detected / harmful. See section 4.
- Don't moralize, don't prescribe, don't do nutrition coaching.
- Don't endorse brands in topic pages.
buying-guide.htmlis the one exception and it says so on its own face. - Don't use emojis.
7. Topics and ownership
One subagent per topic. Waves run in order; foundations first, because the later topics cite them.
Wave 1 — Foundations
retail-map— who owns Pick 'n Save, Metro Market, Trader Joe's, Woodman's, Costco, Aldi, Festival, Sendik's, Meijer, Kwik Trip, Outpost. Banner consolidation, distribution networks, and the private-label decoder. Two poles anchor this page: Kwik Trip, which is vertically integrated and owns its own dairy, bakery, and commissary (a short, largely disclosed chain), and Trader Joe's, which is almost entirely private-label and discloses no suppliers (a chain that goes dark at the label). Establish what each one actually makes versus buys.traceability— the field protocol: how to read anEST./P-number, a PLU, and a lot code off a package and resolve it to a named plant and region. The methodological centerpiece. Structure the page around the spectrum of traceability the stores actually offer, from Kwik Trip's vertically integrated and largely disclosed chain at one end to Trader Joe's undisclosed co-packers at the other, with the Kroger banners in between.label-decoder— USDA Organic, "Natural," Non-GMO Project, grassfed, Certified Humane, Global Animal Partnership, "Product of USA," NSF Certified for Sport, Informed Sport, Prop 65. What each one audits, and what it does not.contaminants— cross-cutting: pesticide residues, heavy metals, PFAS, glyphosate, microplastics, nitrites, mycotoxins. Dose, exposure, and where the headline outruns the data.
Wave 2 — The basket
romaine-and-leafy-greens— the flagship. Salinas/Yuma seasonality, the E. coli O157:H7 outbreak history and the cattle-adjacency hypothesis, the Leafy Greens Marketing Agreement, FSMA produce and traceability rules, and what PDP actually detects on lettuce.other-produce— berries, bananas, avocado, apples. PDP data, and an honest treatment of the Dirty Dozen methodology fight.beef— cow-calf to feedlot, corn and distillers grains, growth implants, ionophores, beta-agonists, the grassfed label after USDA withdrew its standard, and the share of ground beef that is cull dairy cattle (a Wisconsin story).chicken— the lifter's staple protein. Broiler integrators, feed, antibiotic classes and the "no antibiotics important to human medicine" claim, and why hormones are not the issue.pork— ractopamine, the hog belt, processor consolidation.eggs— cage-free vs. pasture-raised vs. nothing, avian influenza and price shocks.dairy— milk, Greek yogurt, cottage cheese, and ultrafiltered RTD shakes. Dairy-cow feed, rbST, and the bulk-tank antibiotic residue testing every tanker gets. Kwik Trip's own dairy is the best available case study: a retailer that processes its own milk from a defined set of Wisconsin farms is the shortest verifiable chain in the whole project. Establish how short it actually is, and what it does and does not disclose about the farms and their feed.seafood— canned tuna and mercury, farmed vs. wild salmon, species mislabeling.carb-staples— rice and inorganic arsenic, oats and glyphosate, bread, potatoes.
Wave 3 — Supplements
protein-powder— whey, casein, plant. The Wisconsin whey chain (whey is a cheese byproduct, and Wisconsin is the cheese state — trace it). Heavy metals in third-party testing. Amino spiking and how the nitrogen test is gamed. What NSF Certified for Sport and Informed Sport actually certify.creatine— sourcing, Creapure vs. commodity, DCD and DHT contaminants, and the unusually strong efficacy literature.pre-workout-and-stimulants— undeclared stimulants, the Cohen research program, FDA's tainted-products list.micronutrients-and-fish-oil— vitamin D, magnesium, fish-oil oxidation and rancidity.bars-and-rtd-drinks— protein bars and ready-to-drink shakes; the co-packer map.
Wave 4 — Synthesis (built after the topic pages exist)
cost-per-gram-protein— every protein source ranked by dollars per gram of protein, per store, date-stamped.buying-guide— the practical tiered basket, store by store. Acknowledges on its own face that brand-specific guidance is a deliberate exception to the no-endorsement rule in section 6. Honor the spirit: present tradeoffs, list multiple options per tier, name marketing-vs-evidence mismatches.index— landing page.
8. Sources master list
Each subagent appends new sources to /sources/bibliography.md under its topic section. Deduplicate by title where possible. The orchestrator does a dedup pass after each wave.