agentleFS
TechNickAITechNickAI/hermes-skills/skills/productivity/fieldy/SKILL.md

fieldy

Skill0 starsChanged 3 months agoMITView on GitHub

Use when reading conversations, transcripts, summaries, or action items captured by a Fieldy AI wearable note taker, or when the user mentions Fieldy, their wearable, or wants the record of something said in person rather than on a video call. Covers API key setup, the time-window query model, pagination, and the async-processing and speaker-labelling gotchas that make naive reads report incomplete data.

  • Reads credentials

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 2 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use when the user mentions Fieldy, their wearable, or wants a real-world (non-video-call) conversation transcript, summary, or the action …

    line 28
  2. Distinctive:

    If the user runs several capture devices (a meeting notetaker, a second pendant), coverage differs by device.

    line 38
  3. Very specific:

    Verify: python3 scripts/fieldy.py whoami → returns {"email": "..."}.

    line 50
  4. Very specific:

    Only then summarize. If the user asked what was actually _said_, do not answer from summary alone — that field is model output about the …

    line 151
  5. Very specific:

    Default pageSize on conversations is 6.

    line 167
  6. Very specific:

    Before any speaker profile exists, segments come back with speaker: "Unknown" and the conversation's speakers[] is empty.

    line 188

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Fieldy

When to Use

Use when the user mentions Fieldy, their wearable, or wants a real-world (non-video-call) conversation transcript, summary, or the action items the device captured. Also use when asked what was said in a specific in-person meeting, hallway conversation, or errand.

Fieldy is a wearable AI note taker (pendant or wrist, Bluetooth to a phone) with a companion desktop app that captures Zoom/Teams/Slack calls. It transcribes conversations and generates summaries, keywords, quotes, speaker labels, and action items. This skill reads that data over Fieldy's Public REST API.

If the user runs several capture devices (a meeting notetaker, a second pendant), coverage differs by device. A gap in Fieldy is not evidence the conversation never happened — check the other sources before concluding anything is missing.

Setup

  1. Open the Fieldy mobile app → Settings (gear icon) → Developer Settings → copy the API key. It starts with sk-fieldy- and is long-lived. The same screen exists in the desktop app.
  2. Export it as FIELDY_API_KEY, or put it in a 0600 dotenv file and pass --env-file /path/to/.env. It is a bearer token for the user's entire personal conversation history — treat it like a password.
  3. Verify: python3 scripts/fieldy.py whoami → returns {"email": "..."}. A wrong key returns HTTP 401 {"code":"UNAUTHORIZED"} with exit 1.

The client reads only the environment variable and an explicitly-passed --env-file. It deliberately does not search config directories for a key: an implicit scan on a multi-account machine can hand the caller a different person's credential and silently return someone else's conversations. If several agents share a host, give each one its own FIELDY_API_KEY.

API facts

Verified against the live API and its published OpenAPI 3.1.1 spec.

  • Base URL: https://api.fieldy.ai/api/public/v2
  • Auth header: Authorization: Bearer sk-fieldy-<key>
  • Rate limit: the API reference states 100 requests / 60s followed by a 60s cooldown; the vendor's launch post says 30/min. Treat 30/min as the safe budget and back off on 429. The bundled client sleeps between pages and retries 429 with escalating delay — but only for GET (see Pitfalls).
  • The API never legitimately redirects, and the client refuses redirects outright. This is a security control, not politeness: urllib replays request headers on a redirect, so a 302 to another host would hand that host the bearer token. Verified by test.
  • There is no search endpoint. Every read is time-window based — pull a range and filter locally. Do not look for a ?q= parameter; it does not exist.
  • Interactive docs: https://api.fieldy.ai/docs (Scalar). The raw OpenAPI JSON is not served at /openapi.json, /docs/json, or /docs/openapi.json (all 404). It is embedded in the docs HTML as const scalarConfig = {...}, where the content key holds the whole spec as a JSON string. Extract with json.JSONDecoder().raw_decode on the text after that assignment, then a second json.loads on content. A cached inventory of every endpoint, parameter, and response schema is in references/openapi-endpoints.md.

Endpoints

MethodPathNotes
GET/user/mereturns {email} only — the cheapest auth probe
GET/conversationsstartTime+endTime required, ISO 8601; mode=starts-in-range|intersects-range, pageSize max 50 (default 6), cursor, recordingSource=wearable|phone|desktop
GET/PATCH/DELETE/conversations/{id}GET returns bare null for an unknown id, not a 404
POST/conversationscreate
GET/transcriptionsstartTime required, endTime optional; conversationId, recordingSource, order=asc|desc, inclusive, pageSize max 1000, limit max 2000
GET/POST/PATCH/DELETE/tasks, /tasks/{id}GET requires status (new/approved/completed/rejected/skipped/cancelled/expired)
GET/POST/PATCH/DELETE/speaker-profiles, /speaker-profiles/{id}
GET/POST/PATCH/DELETE/memory-templates, /memory-templates/{id}prompt templates that shape summaries
GET/POST/PATCH/DELETE/sharables, /sharables/{id}public share links; also GET /sharables?conversationId=, GET /sharables/resolve?idOrUrl=

Response shapes

  • Conversation: id, title, summary, content, startTime, endTime, type(FULL|BRIEF), keywords[], speakers[], quotes[{text,context}], location{address,coordinates,city,...}, locationId, templateId, calendarEventId, updatedAt. memorySpeakers and memoryTemplateId are DEPRECATED aliases kept for old clients — use speakers and templateId.
  • Transcription segment: id, text, timestamp, speaker, speakerProfileId, start, end, createdAt, source, recordingSource. start/end are seconds offset within the recording; timestamp is the absolute time.
  • List responses are {items: [...], nextCursor: string|null}. Tasks return {items} with no cursor.

Usage

S=path/to/scripts/fieldy.py

python3 $S whoami                                    # auth probe
python3 $S conversations --days 1 --text             # last 24h, readable
python3 $S conversations --start 2026-05-01T00:00:00Z --end 2026-05-07T23:59:59Z
python3 $S conversation <id>                         # one conversation, full JSON
python3 $S transcript --conversation-id <id> --text  # speaker-labelled transcript
python3 $S transcript --days 1 --source wearable --text
python3 $S tasks --status new
python3 $S speakers
python3 $S templates
python3 $S raw /sharables --param conversationId=<id>   # any endpoint

--text gives readable output on every subcommand; the default is JSON for piping. --text, --verbose, and --env-file work on either side of the subcommand.

--limit and --page-size exist only on conversations and transcript, which are the paginated commands, and must be placed after that subcommand. Passing them elsewhere is a hard argparse error rather than a silent no-op. Pagination is automatic for conversations and transcript — those are the only cursor-paginated endpoints. tasks, speakers, and templates return a single page and do not accept --limit or --page-size. Within the paginated commands the client follows nextCursor until the range is exhausted and never requests a larger page than --limit needs. If it cannot complete the walk — a repeated cursor, or the page cap — it exits nonzero rather than printing a plausible-looking partial list. Pass --allow-partial to accept a truncated result with a warning.

Mutating calls through raw require an explicit --yes.

Workflow: "what did I discuss about X this week"

  1. conversations --days 7 as JSON, then filter locally on title, summary, keywords, and quotes. There is no server-side search, so the filtering is yours to do.
  2. For each hit, transcript --conversation-id <id> --text to get the raw record.
  3. Only then summarize. If the user asked what was actually said, do not answer from summary alone — that field is model output about the conversation, while the transcript is the conversation. Quote the transcript.

MCP alternative

Fieldy also serves MCP at https://api.fieldy.ai/mcp (HTTP transport), offered as a one-tap connector in Claude and ChatGPT. That path uses a browser-driven OAuth handshake, and there is no documented way to authenticate it with a bare sk-fieldy- key. For an agent writing code, prefer the REST API. Reach for MCP only when the goal is Fieldy inside a chat client's connector UI.

Pitfalls

  • startTime and endTime on /conversations are required. A bare GET /conversations is a 400, not "everything".
  • Default pageSize on conversations is 6. An unpaginated call will happily report six conversations as though that were the whole week. Always follow nextCursor.
  • mode defaults to starts-in-range, so a conversation that began before your window and ran into it is excluded. Use intersects-range when the question is "what was happening at 3pm" rather than "what started today".
  • Transcripts are fetched by time range, not by conversation id alone. The spec describes conversationId as "legacy client input resolved to canonical recording source". Resolve the conversation first and pass its real startTime/endTime; do not assume the id filters on its own. Verified: a conversation-scoped fetch and a raw time-window fetch over the same interval returned identical segments.
  • Processing is async after a recording stops ("Sending to Private Cloud" → "Transcribing" → "Generating Title"). A just-ended conversation can return a null title, null summary, empty speakers[], and zero transcript segments. That is not an API failure and not an empty conversation — retry later before reporting nothing was captured.
  • A recording has a 3-hour hard cap, so a long day is many conversations rather than one.
  • The device only captures while transcription is running. A gap in the data means it was not recording, not that the API lost anything.
  • Before any speaker profile exists, segments come back with speaker: "Unknown" and the conversation's speakers[] is empty. Once a profile is created, segments label as Speaker 1, Speaker 2, and so on. These are positional labels, not identities — a fresh account's only profile was named User. Do not promise "who said what" beyond what the labels support.
  • GET /conversations/{id} with an unknown id returns HTTP 200 with a bare null body rather than a 404. Check for None explicitly or a downstream conv["startTime"] raises an opaque TypeError.
  • Transcripts are verbatim, including profanity, false starts, and whatever was said near the device by people who did not know it was recording. Treat the content as sensitive by default and do not echo more of it than the task needs.
  • Partial results fail loudly. A truncated list printed with a normal count is the worst outcome this client can produce: an agent reads it as the complete record of a week and reports that something was never discussed. Any incomplete pagination walk exits nonzero unless --allow-partial is given. Relatedly, an empty page mid-range is a hole rather than the end — the walk follows a live cursor past it.
  • Ids are percent-encoded before they reach the URL. Unencoded, a conversation id of x/../../user/me traverses to a different endpoint — verified: it returned the account profile instead of a conversation. Ids get interpolated from model output and transcript text routinely, so treat any id as untrusted.
  • The raw path argument must be a bare API path. ?, #, ://, and a leading // are rejected, and the rejection message deliberately does not echo the offending value, since the reason for rejecting it is that it may carry a secret. Query values go through --param.
  • Timestamps keep sub-second precision. Rounding --start down and --end up to whole seconds silently widens or narrows the window at the boundary, which changes which transcript segments come back.
  • DELETE and PATCH mutate the user's personal record, and POST /sharables mints a publicly accessible link to a private conversation. They are reachable only through raw --method ... --yes; there is no convenience subcommand for them by design. Never call one without an explicit instruction naming the target.
  • Only GET is retried. A retried POST/PATCH/DELETE that actually succeeded before the connection broke repeats the side effect — for POST /sharables that means several public links to one private conversation. A failed mutation reports Outcome UNKNOWN and stops; reconcile state before trying again.
  • Error output prints the endpoint path without the query string and redacts anything matching sk-..., because both can carry the key and API error bodies can quote transcript text. --verbose widens the echoed body; use it when debugging, not in normal operation.
  • A time window is anchored to --end, so --end <date> --days 7 means the seven days before that date. An earlier version anchored the lookback to now, which silently produced start > end — the API answers an inverted window with zero rows, and an agent then reports "nothing was discussed" when the query was simply malformed. Inverted or unparseable windows now exit 1.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.