I want you to set up a daily literature digest for me that lives inside my Agent Context Folder and feeds my personal knowledge base. It should find new papers in my field, rank them by evidence and by relevance to my work, show me a short numbered list each morning as a web page I can open in my browser or on my phone, and file anything I tell it to keep. Work in six phases, in order. Don't skip ahead, and don't write anything until Phase 4. --- ## Phase 1: Find my folder and check what you can do 1. Ask me where my Agent Context Folder is. If I don't have one, help me pick a place you can read: my computer, or a cloud drive (OneDrive, Google Drive, Dropbox) if you can only read cloud files. If it's in a cloud drive, ask for the folder's name and where it sits in the drive (for example "top level of my OneDrive, named agent-context-folder"), not a file path. If you can't open the folder but I can attach files to this chat, ask me to attach my context files instead. 2. Find out what you can actually do here by trying it. Don't guess: - Can you read files in that folder? - Can you create, edit and move files in it? - Can you search PubMed: through a connector, by running code against NCBI E-utilities, or only through web search? - Can you run a prompt on a schedule, and will the scheduled run still have access to my folder? 3. Tell me the results in plain language, one line each. If you can't write files, say so: every phase still works, but you'll give me file contents to save myself. If you can't schedule, I'll run the digest by hand each morning. If you can't search PubMed or the web at all, stop here: tell me the digest can't run in this tool, and suggest a chat with web search turned on or a different assistant. Don't go on to the interview. 4. Read everything in `context/`. If it's empty or missing, tell me that the digest works much better with a context portfolio (the Context Portfolio Builder at contextportfolio.ai builds one in about an hour), then ask whether to continue with a short interview instead. ## Phase 2: Interview me Ask one or two questions at a time. Use what's already in my context files and confirm it rather than asking again. If I'm vague, offer two or three concrete options. Aim for about six turns. 1. **Topics.** 5 to 10 search topics, each narrow enough to be useful. "Carotid stenosis" or "AI in surgical education", not "surgery". Draft them from my context files and let me edit. 2. **Triage rules.** The rules only someone in my field would know. For example: "skip single-center retrospective series under 50 patients unless the topic is rare", "anything in [my flagship journal] makes the list", "no 'we asked ChatGPT' papers unless the evaluation is rigorous". Push me for at least three. 3. **Journals I never want to miss**, if any. 4. **How many papers a day.** Default 5. 5. **Domain tags.** 4 to 8 tags for sorting what I keep (for example: Clinical practice, Research methods, Education, AI in medicine). If my context files already have a list, use it. If not, we'll add one. 6. **Schedule.** What time and which days. Default weekdays, early morning. ## Phase 3: Confirm the plan Before building anything, show me: - My topics, triage rules, journals, number per day, and domain tags - Exactly which files and folders you'll create (Phase 4 list) - How you'll search (connector, code, or web search) and what that means for accuracy - How it will run: scheduled (when, and in what tool) or by hand Wait for my "go". ## Phase 4: Build it Inside my Agent Context Folder, create whatever is missing. Never overwrite or delete a file that already exists; if one is in the way, ask me. ``` agent-context-folder/ context/ (already there: my context portfolio) kept/ one file per thing I keep TEMPLATE.md inbox/ I drop things here to save them later processed/ you move originals here after filing them briefs/ one digest per day, as a web page (.html) seen.txt papers I've already been shown (starts empty) connections/ my Friday connections (another prompt fills this) my-daily-digest-prompt.txt ``` **kept/TEMPLATE.md**, exactly this: ``` --- title: source: (citation, DOI or PMID, or URL) date_kept: YYYY-MM-DD domain: (one tag from my list) evidence: (RCT, meta-analysis, cohort, case series, guideline, expert opinion, other) links: (file names of related items in kept/) --- ## Summary One or two sentences: what it found and how strong the evidence is. Written by the agent every time. ## Why I kept it Optional. My words, when I have them. Never replaces the summary. ``` **Domain tags.** If my context files don't already list them, add the list to the most fitting context file (or create `context/domains.md`). Every prompt must use that one list. **my-daily-digest-prompt.txt** is my filled-in daily digest prompt: the whole daily run, written out so any session (scheduled, or a chat I open later) can follow it. It holds my topics, triage rules, journals and number per day where I can edit them in any text editor, plus a copy of the kept-file template (so the keep step works even when `kept/TEMPLATE.md` can't be read), plus these steps in this order: 1. **Read every file in `context/`.** Who I am, what I work on, what matters right now. 2. **Don't run twice.** If `briefs/YYYY-MM-DD.html` for today already exists and has content, stop and tell me. Never replace a good brief with an empty one. 3. **File the inbox.** For each file in `inbox/` (not `inbox/processed/`): work out what it is and find the full citation if it's a paper; write one kept file from the template (rules below); then move the original into `inbox/processed/`. Never delete anything. If something can't be processed, leave it in `inbox/` and say why in the brief. 4. **Search.** Look for papers on my topics added since the date of my most recent brief (at least 2 days back, at most 7). Run a separate search for each topic; don't replace them with a few broad searches. On PubMed, search by the date the record was added (Entrez date), not publication date. If you're using web search instead of PubMed directly, say at the top of the brief that coverage may be incomplete. If you're using code, use NCBI E-utilities with my email in the request and pause about a third of a second between calls. 5. **Leave out what I've seen.** Drop anything listed in `briefs/seen.txt` or already in `kept/`, matched by PMID or DOI (or title, if a paper has neither). Don't read old briefs for this; `seen.txt` is the record. 6. **Rank** what's left by (a) strength of evidence: meta-analysis and RCT first, then prospective cohort, retrospective cohort, case series, expert opinion; (b) relevance to my projects and goals in `context/`; (c) my triage rules. Journals on my never-miss list always make it. 7. **Write the brief** as one self-contained web page, `briefs/YYYY-MM-DD.html`, that opens in any browser and reads well on a phone: one column, large readable text, no external files or scripts. Inside it: - First line: how many papers you found, how many were already seen, how many you ruled out. - Then the top papers, **numbered 1 to N**, with big visible numbers. For each: citation with a PubMed or DOI link that opens in a new tab, one-sentence finding, study design, and one line on why it matters to me specifically. - If nothing clears the bar, say "Nothing worth your time today." Don't pad. - **Filed from inbox:** the kept files you wrote in step 3, and anything you couldn't process. - **Also screened:** every other paper you looked at, one line each, citation and PMID or DOI only, in a section that starts collapsed, so I can check what got ruled out. - Last line: `Reply "keep" with numbers, e.g. "keep 2, 4" or "keep 2 because ..."` 8. **Show me the brief**: open it or display it if you can, and tell me where the file is. 9. **Update `briefs/seen.txt`, last.** Add one line for every paper you looked at today (picked and screened): `YYYY-MM-DD PMID DOI`, leaving a blank where an ID is missing. If a paper has neither, write its title instead. Then delete lines dated more than 14 days ago; the search never goes back further than 7, so older lines aren't needed. Do this only after the brief is saved, so a run that fails partway never marks papers as seen. **The keep step** (also written into `my-daily-digest-prompt.txt`). When I reply "keep" with numbers, here or in any later chat ("keep 3 from Tuesday"), use the matching brief in `briefs/` (today's unless I name a day). For each number: - Write one file to `kept/` named `YYYY-MM-DD_first-author_short-title.md`, using the kept-file template. If you can't save files to my folder, give me each kept file as text with its file name, ready for me to save into `kept/`. - Always write the Summary yourself. If I gave a reason, put my words, not a paraphrase, under "Why I kept it". If I didn't, leave that section empty. - Pick one domain tag from my list. - Read the summaries in `kept/` (ignore `TEMPLATE.md`) and list up to 3 genuinely related items under `links`. Add this new file's name to the `links` of each of those items too. If you can't read `kept/`, leave `links` empty and say so; the Friday prompt can find connections later. - Tell me the file names you wrote. The same kept-file rules apply when filing the inbox. **Accuracy rules for every run:** - Never invent a citation, PMID, DOI, author, or finding. If you didn't retrieve it, you don't have it. - If you're searching the web rather than PubMed, include a paper only if you found its PMID or DOI on a real page. Otherwise drop it. - Report errors (a search that failed, a file you couldn't read) at the top of the brief. Don't hide them, and don't save anything to a fallback location. ## Phase 5: Test it on real papers 1. Run `my-daily-digest-prompt.txt` once, now, with today's real results. Show me the brief. 2. Walk through the list with me. For each paper ask: should it be here? Then look at the Also screened list: did anything get ruled out that I'd have wanted? Turn my answers into sharper triage rules, update `my-daily-digest-prompt.txt`, and re-run if the changes are big. Repeat until the list looks like what I'd pick myself. 3. Test the keep step: have me keep one paper with a reason. Show me the kept file and the backlinks you added. 4. Test the inbox: have me drop one item into `inbox/` (a PDF, a saved web page, or a text file with a link). File it the way the morning run will, and show me the kept file and that the original is now in `inbox/processed/`. 5. Show me `briefs/seen.txt` so I can see today's papers were added. Don't call it done until I've seen a real brief, agreed with the ranking, seen a kept file from each route, and seen `seen.txt` update. ## Phase 6: Schedule it Use this tool's own scheduling, at the time and days I picked. If the scheduled run can read my folder, the scheduled prompt can be short: read and follow `my-daily-digest-prompt.txt` in my Agent Context Folder (give the full path). If the scheduled task can't read that file, put the complete daily instructions into the task itself. Copy them in full; don't summarize. They must include: - the search window: only papers added since the last run (at least 2 days back, at most 7), never "recent" in general - how to skip papers I've already seen: `briefs/seen.txt` if the run can read it; if not, skip anything posted in earlier runs of this task that it can see - the rule to drop any paper without a PMID or DOI it actually found - the count line at the top (found, already seen, ruled out), the numbered list, and the Also screened list - the keep step, with the kept-file template written out After you create the task, show me its full text so I can check nothing was dropped. Read the schedule back to me in one plain sentence ("Every weekday at 5:00 AM, this runs your digest and saves it to briefs/"). Tell me where I'll find each morning's brief (the file, and whether it also shows up in this app) and where to reply "keep". If this tool can't run on a schedule with access to my folder, say so plainly. Then give me the one line to paste each morning instead: "Read and follow my-daily-digest-prompt.txt in my Agent Context Folder." Don't set up system-level scheduling (cron, Task Scheduler) unless I ask for it. If you can only read files I attach to the chat (not my folder), don't tell me to "read the folder". Give me a morning routine that actually works here, in three short steps: which files to attach (`my-daily-digest-prompt.txt`, my context files, and `briefs/seen.txt`), the one line to paste ("Follow the attached my-daily-digest-prompt.txt"), and that you'll give me the brief and any kept files as text to save into my folder myself. --- ## Don't - Don't start building before I say "go". - Don't make me fill out a form. Talk to me. - Don't delete anything, ever. Move it to `inbox/processed/`. - Don't overwrite my files, my context portfolio, or an existing brief. - Don't fill in "Why I kept it" for me. - Don't add sections I didn't ask for (news, weather, task lists). The digest is papers and my inbox. I can add more later. Start with Phase 1, step 1.