Find Quotes

Before you start: read this whole install prompt, tell me in a few lines what this skill does and what installing it involves, and wait for my go-ahead before you do anything else. Prerequisites — an external MCP service is required to USE this skill, and a paid-plan feature is involved: installing the skill file itself needs no account, no key and nothing else first, and this install does NOT connect anything, create an account or ask for a key. But the skill is instructions only — it runs no local code and does nothing at all unless the remote Arcmira MCP server is already connected to my agent at https://mcp.arcmira.com/mcp (that connection is NOT part of the skill folder and is NOT set up by this install). To use it I need my own Arcmira account, signed in one of two ways (repo README, "Authentication options"): (a) sign in through the host — add the MCP URL with no key, the server answers with an OAuth challenge, the host opens arcmira.com for my sign-in and consent, and connects with a token carrying the permissions I allowed; or (b) a key — a client that cannot do that sign-in sends an account key (arc_sk_..., from arcmira.com) as a Bearer header instead. Reads are metered in credits from my account's plan, then my on-demand budget set in the Arcmira dashboard (see point 2 for the exact rates: search, caption transcript reads, and Premium transcript reads); a Premium transcript (quality: "premium", speaker ids added) requires a paid plan. The skill file itself is free: it sits in plugins/arcmira/, which the repo licences under Apache-2.0 (the repo-root LICENSE is proprietary and covers the server, which this install does not install or copy). Install "find-quotes" for me. It finds exact spoken quotes and clip-ready moments on podcasts and YouTube: the verbatim words, the speaker, the date, a timestamped watch link, and a clip start and end in seconds. It resolves the name I give to an id first (filters take ids only — ent_..., UC... channel ids, 11-character video ids; a name where an id belongs throws id_required): the exact name I said, with my own words about it as context, comes back as best (use it), suggested (an assumption the answer must name, with suggested.evidence as the reason) or ask (several fit; the options go back to me). A speaker resolves with type "person" and goes in speakerIds (who said it); a person or brand the passage is about goes in about; a show goes in channelIds; the topic words go in query, never in about. Then search finds the passage and transcript with start and end returns the exact lines around it with second offsets, which give the verbatim quote and the clip boundaries. SKILL.md's worked program searches what one person said about a topic (limit 5), then reads a roughly 60-second transcript window (start minus 10 seconds to start plus 50 seconds) around each of the top 2 hits. The answer quotes the words exactly as the transcript lines give them, trimmed to whole sentences, never paraphrased inside quotation marks and with nothing fixed inside the quotation marks (caption lines are machine text); it names the speaker or show searched, with the id, gives the show, episode title, date and speaker when named, links the watch_url (which starts at the moment), and gives the clip start and end in seconds. If a filtered search returns no chunks, it reruns with fewer filters before saying nothing was found, and when nothing matched it says so, with as_of, instead of offering a quote from memory. Repository: https://github.com/arcmira/mcp (the skill is the whole plugins/arcmira/skills/find-quotes/ folder at the pinned commit: SKILL.md only, 1 file) Pinned version: install the code at commit 4598d894ad1a1856cbc10473b374d2f3aa7bc6c6 exactly. Do not substitute a newer version. If this commit is unreachable, stop and tell me. 1. Clone https://github.com/arcmira/mcp into a temporary folder, check out commit 4598d894ad1a1856cbc10473b374d2f3aa7bc6c6, and copy the plugins/arcmira/skills/find-quotes/ folder from that checkout. Then compare every file in the folder with the fingerprints recorded for this listing on 2026-10-19: [FINGERPRINTS] - SKILL.md: SHA-256 6d4431b66a043e8c0696f60ea9b3e826646a2fa6ade8ebd2379e92e09f7ca36c If a file does not match, it changed since the listing was prepared: do not install it yet, tell me what you found and let me decide. 2. Read SKILL.md in full, and tell me in plain words: what the skill does when used, which services it connects to, whether it reads or sends any personal information or keys, and anything that seems unusual. Stop on any red flag and tell me before going further. Confirm these points from the files: - The skill folder holds 1 file only: SKILL.md (5,602 bytes). There is no script, no code, no configuration and no data file in it. The skill is NOT self-contained in the way a code skill is: it is a procedure plus one worked JavaScript program, and every step of that program runs through tools that live outside the folder — arcmira_execute_read (runs the program), arcmira_describe (the method reference it points to) and, after an answer, arcmira_feedback. The MCP connection itself is configured outside the skill folder (in the repo: plugins/arcmira/.mcp.json and plugins/arcmira/mcp.json, both naming https://mcp.arcmira.com/mcp; neither file is part of this install). SKILL.md does not reference the other five skills in the plugin, the repo README or the server source; the one cross-reference it makes is a note that the separate company-watch skill is what sets up a monitor (see below). If the Arcmira MCP tools are not connected, installing this file changes nothing and the skill cannot run. - Nothing is sent anywhere during installation itself; the only download at install time is the skill folder from GitHub at the pinned commit. No program is run during installation, no MCP server is connected, no account is created, no key is asked for, and there is no telemetry in the folder. - What leaves this machine at use time, disclosed here before it ever happens — using the skill is a separate step that only starts when I ask for a quote or a clip moment, and every send below goes to Arcmira's own servers on my Arcmira account: (a) The program itself: each question is sent as one JavaScript program through the arcmira_execute_read tool to the remote MCP server at https://mcp.arcmira.com/mcp, which forwards it as GET requests only to https://api.arcmira.com/v1/* (resolve, search, transcript). The program text contains exactly what the question needs: the name I gave, resolved first, with CONTEXT set to my own words about that name (never a guess — a bare name is sent with no context), the topic words in query, the resolved filter ids (speakerIds for who said it, about for what it is about, channelIds for a show), any date window built from arcmira.today() / arcmira.daysAgo(n), and, for the transcript reads, the video id and the start/end window in seconds. (b) My credential: the program itself never holds my key — the server's outbound proxy adds my caller's credential to what it forwards (repo README, "The sandbox"). That credential is the OAuth token from my host sign-in (carrying the permissions I allowed at consent), or my arc_sk_ account key sent as a Bearer header if my host uses a key; it goes to Arcmira only. (c) Credits, charged to my account's plan first, then my on-demand budget: a search uses 4 credits per passage it returns past the first 5 (the worked program asks for limit 5); a captions transcript read uses 4 credits per started 15 minutes of its start/end window, or of the whole episode without a window (the worked program's window is about 60 seconds, so 4 credits per read, for up to 2 reads); a Premium transcript read (quality: "premium", paid plans, it adds speaker ids into speakers[]) uses credits for the whole video whatever the window — 300 credits per started 15 minutes of the whole video in the repo's own reference — and a Premium read of a video not yet transcribed starts a transcription. (d) Arcmira's call logging, stated in the repo README ("What we log"): each tool call is logged to Arcmira's product analytics (PostHog), linked to my account — one record per call holding the tool name, my host's name and version, the server version, how long the call took, the input (the program text of an execute tool, first 4,000 characters; the arcmira_describe topic; or the feedback category and note), the optional intent when the agent sends one (at most 300 characters: my request in a few words), and the outcome, not the result (ok or the error code, the result size, whether it was truncated, and the API routes made, e.g. GET /v1/search, not its query string). Before storage, anything that looks like a credential or an email address is replaced with [redacted]; results, transcript text and logs printed by a program are never logged; a call with no valid credential is not logged. (e) Feedback, after an answer only: if anything was wrong, slow or missing for me, or I want a source or capability Arcmira lacks, SKILL.md says to send one arcmira_feedback — one POST /v1/feedback of type experience to Arcmira, with a category (wrong_entity, bad_data, missing, slow, confusing or other) and a note saying what happened, plus the call_id / request_id when one is named. That is a stored record at Arcmira, sent only in that situation, never during installation. (f) What this skill does NOT do at use time: its worked program uses arcmira_execute_read only — it creates no monitor and makes no write. SKILL.md only notes, after the answer, that monitors can follow what a research turn found (each monitor alert uses 25 credits from the account's plan, then its on-demand budget) and that the separate company-watch skill sets one up; setting one up is a different skill and a different step, not part of this one. The only other address in SKILL.md is a documentation link, https://arcmira.com/docs/mcp-server, which is a link to read, not a call the skill makes. - The quoting rules from SKILL.md, which are part of what I am installing: quote only the named speaker's lines when a chunk found with speakerIds includes other people's lines (lines start with "Name: " when the speaker is known); when no line names a speaker, say the attribution covers the passage, not each line; some videos have no readable caption track yet (transcript_unavailable, transcript_fetching) — for those, quote the search chunk text and say the words come from the search passage; search ranks by the words in query, so the distinctive words of the phrase go there, not a paraphrase. - Do not connect the Arcmira MCP server, do not sign in to Arcmira, do not create an Arcmira account, do not set or ask for an arc_sk_ key or an OAuth token, do not run the worked program or any other program, do not search or read any transcript, and do not send any arcmira_feedback during this install: connecting the MCP server and using the skill are my decisions, later, when I actually ask for a quote. 3. When I say "yes, install", install it as a skill: copy the whole find-quotes folder into the agent's skills directory, in a folder named "find-quotes", keeping its 1 file (SKILL.md at its root). A manual clone followed by copying the plugins/arcmira/skills/find-quotes folder into ~/.claude/skills/ is the route described here; any other route is fine only as long as every installed file matches its fingerprint. 4. Verify with no network calls: frontmatter valid (name: find-quotes), the 1 file in place with matching fingerprint. Do not run any program, connect any MCP server or call arcmira_describe to verify. 5. Report what was installed, where, and what I still need to do myself (for example: connect the Arcmira MCP server to my agent myself at https://mcp.arcmira.com/mcp — it is not in the skill folder and this install did not connect it; sign in to Arcmira through my host's OAuth flow, or get an arc_sk_ account key from arcmira.com and configure it as a Bearer header if my host cannot sign in; make sure my Arcmira plan has credits for the metered reads disclosed above, and a paid plan if I want Premium transcripts; then, when I ask for a quote, give the exact name and my own words about it as context, plus the topic words — knowing the program text, the name, the context and the topic are sent to mcp.arcmira.com / api.arcmira.com on my account, that each tool call is logged to Arcmira's PostHog analytics as disclosed above, and that one arcmira_feedback (category and note) is sent to Arcmira after an answer if something was wrong, slow or missing; using the skill is a separate step that only starts when I ask for it, and the use-time sends disclosed above happen only then). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.