Fulcra Annotations#

Use this skill when the user wants an agent to create, record, or verify Fulcra annotations.

Fulcra gives agents and their humans scoped, secure access to read and write real-world context and shared human/agent memory: attention, events, location, calendar, health, wearables, and other streams. Use this skill for the write side of that loop: creating reusable annotation definitions and recording user-approved moments or values.

Agents should use the bundled script first. Do not hand-write curl calls unless the script is missing a required capability, because the script keeps tokens out of chat, builds the Fulcra ingest payload consistently, and performs readback verification.

First-Run Onboarding Pattern#

When a user is new to Fulcra annotations, optimize for a quick useful loop: choose a concrete thing to track, create one or two definitions, record one real data point, verify it, and show the user what now exists.

1. Start with a short, grounded prompt. Do not ask only "what do you want to track?"; offer 2-3 specific options based on the user's context, such as a daily focus score, coffee count, symptom check, workout effort, or a moment log for important events.
2. Check auth only after there is a clear use case. If uv tool run fulcra-api user-info fails, run uv tool run fulcra-api auth login, keep the process alive, and send only the device URL/code through the trusted user channel.
3. Translate the use case into 1-3 annotation definitions. Prefer the simplest type that captures the signal: moment for occurrences, boolean for yes/no, numeric for counts or measured quantities, and scale for subjective ratings.
4. Run list before create to avoid duplicates. If creating multiple definitions, save the returned annotation.id, source_id, and type from each create result in your working notes so the next record step does not need another lookup.
5. Ask one direct question for the first record, then call record --id ... with --value when needed. Treat success as confirmed only when verified_matches >= 1.
6. For the handoff, summarize the definition names, the exact timestamp written, and one natural next action. If generating a demo artifact, use synthetic or explicitly approved real data and keep private records out of chat.

Core Concepts#

• Annotation definition: the reusable button/metric definition, such as Focus or Asked Agent to Do Something New. Created once with create.
• Annotation record: one logged occurrence/value of a definition. Written with record.
• Moment annotation: an event with no value. Use for "this happened" logs.
• Boolean/numeric/scale annotations: metric-like records that require --value.
• Definition tags: reusable labels stored on the annotation definition. Add them when creating the definition with repeated create --tag.
• Record tags: labels stored on one specific record. Add them when recording with repeated record --tag.
• Tag resolution: Fulcra stores tag IDs. The script accepts tag names or UUIDs, resolves names to IDs, and creates missing tag names automatically.
• Historical record: any record whose event time is not now. Always pass --recorded-at.
• Confirmed write: a record is not confirmed until readback finds verified_matches >= 1.

Safety Rules#

• Never print access tokens, refresh tokens, raw private Fulcra records, credential files, or direct capability URLs in chat.
• Authenticated API calls are pinned to https://api.fulcradynamics.com; do not redirect Fulcra bearer tokens to custom API hosts.
• FULCRA_CLI_COMMAND is restricted to the standard Fulcra CLI invocation ( uv tool run fulcra-api, fulcra-api, or a trusted absolute path ending in fulcra-api).
• Device auth URLs and user codes are allowed only when the intended user needs to approve Fulcra access from another device; send them only through the active trusted user channel.
• Ask before deleting or updating an existing annotation definition.
• For public demos, use synthetic annotation names/values unless the user explicitly approves real data.
• Do not claim a write succeeded from HTTP status alone. Check the script result and verify readback.
• Duration annotations are only partially supported by this skill; prefer moment/boolean/numeric/scale until Fulcra documents duration ingest shape more clearly.

Auth#

The script gets auth from a trusted secret manager token when FULCRA_ACCESS_TOKEN is set, otherwise from the locally authenticated Fulcra CLI command configured by FULCRA_CLI_COMMAND.

Fulcra requires an authenticated account, not an API key. Accounts can be created through the CLI auth flow and include 5 GB of storage free forever. Users who want biometrics, location, calendar, and other mobile context can install the Context iOS app and sign in with the same account; the app uses the same free storage and is no longer subscription gated. Android is coming soon.

Authenticate first:

bash

uv tool run fulcra-api auth login

For remote agents, keep the CLI running, surface the printed device authorization URL and code to the intended user in chat through the active trusted user channel, and wait for approval. The user can open the URL from any browser on any device. After approval, verify auth with a non-token command such as uv tool run fulcra-api user-info; do not paste tokens into chat.

Set FULCRA_HOME=/path/to/home if credentials are not under the process HOME.

Common Commands#

List annotation definitions before creating a new one:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py list

Create a moment annotation definition:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py create \
  --type moment \
  --name "Asked Agent to Do Something New" \
  --description "Logged when the user asks the agent to try a new category of work" \
  --tag agent \
  --tag new-task

Use create --tag more than once to attach multiple tags to the definition. Definition tags should be short, stable labels such as agent, health, or research. The script resolves tag names to Fulcra tag IDs before sending the API payload.

Record a moment annotation now:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py record \
  --name "Asked Agent to Do Something New" \
  --note "User asked for a new annotation workflow" \
  --tag new-task

Use record --tag more than once to attach tags to the individual record. If record --tag is omitted, the record inherits the definition tags. If record --tag is present, those explicit record tags are resolved to Fulcra tag IDs and sent for that record.

Record a historical moment annotation:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py record \
  --name "Asked Agent to Do Something New" \
  --recorded-at "2026-05-15T10:00:00-04:00" \
  --note "User asked for a new annotation workflow"

Use a full ISO-8601 timestamp with timezone for historical writes. If the user says "yesterday at 10am", resolve it in the user's timezone and pass the offset explicitly, for example 2026-05-15T10:00:00-04:00. Fulcra readback may show the equivalent UTC time, such as 2026-05-15T14:00:00+00:00.

Record by annotation ID when names are ambiguous:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py record \
  --id "<annotation-id>" \
  --note "Logged from automation"

Create a 1-5 scale annotation definition:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py create \
  --type scale \
  --name "Focus" \
  --description "How focused do I feel right now?" \
  --scale-labels "1=Scattered,2=Low,3=Neutral,4=Focused,5=Locked In" \
  --default-value 3

Create a numeric count annotation:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py create \
  --type numeric \
  --name "Coffee Count" \
  --description "Number of coffees consumed today" \
  --measurement-type count \
  --tag health \
  --tag intake

For measurements with specialized units, inspect the live Fulcra schema through the official CLI or library before creating the definition and then prefer adding reviewed helper support over ad hoc API calls.

Record a scale/numeric/boolean value:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py record \
  --name "Focus" \
  --value 4 \
  --note "Deep work block started well"

Read back recent records for verification:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py recent \
  --name "Asked Agent to Do Something New" \
  --hours 72 \
  --limit 20

Update or delete definition metadata only through a separate reviewed admin workflow after explicit user approval. The bundled helper intentionally exposes only list, create, record, and recent.

Dry-run any write before sending it:

bash

python3 skills/fulcra-annotations/scripts/fulcra_annotations.py record \
  --name "Focus" --value 4 --dry-run

Workflow#

1. Determine the annotation type: moment, boolean, numeric, or scale. Prefer moment for "this happened".
2. Decide whether tags belong on the definition, the individual record, or both.
3. Run list and check whether the definition already exists.
4. If an existing definition needs a metadata or tag change, ask for approval and use update --id ...; dry-run first for broad retagging.
5. If the definition is missing and the user asked to create/log that annotation, run create.
6. Record with record --name ... or record --id ....
7. For historical writes, always include --recorded-at "<ISO-8601 timestamp with timezone>".
8. Trust only a successful script result with verified_matches >= 1 for confirmed writes.
9. If verification fails after ingest returns 204, wait briefly and rerun recent --id <annotation-id> or recent --name "<name>" --hours <window>.
10. Report the exact timestamp written and the readback timestamp. Do not include tokens, direct capability URLs, or unnecessary private record data.

Demo and Handoff Pattern#

For onboarding, the first verified record should lead to a small payoff instead of a dry "done" message.

• Offer 2-3 visual directions and ask the user to choose before generating a dashboard or artifact.
• Keep demo files local to the workspace and avoid dumping raw HTML or raw Fulcra records into chat.
• Use the verified annotation name, timestamp, value, and a short interpretation. Avoid exposing unrelated Fulcra data.
• Make the next steps concrete: install/open the Context app for mobile logging, inspect data in Context Web, add another annotation, or iterate on the dashboard.
• In public or group-chat demos, use synthetic values unless the user explicitly approved sharing real values.

Idempotent Writer Pattern#

For automated or backfilled annotation pipelines, use an append-only, ledger-backed writer:

1. Normalize the source event into stable fields: annotation name/id, event timestamp, source, severity/category, value/note, and any real filter dimensions.
2. Generate a dedupe key from stable source facts. Keep this key in local state, metadata, or source bookkeeping; do not place it in the visible note.
3. Check the local ledger before writing. Skip verified, retry pending or failed, and write only unseen keys.
4. Treat ingest HTTP success as provisional. Confirm with record output or recent readback and mark the ledger verified only when verified_matches >= 1.
5. For changed source events, avoid duplicate annotations unless the change is materially new. Prefer a follow-up annotation for escalation/resolution over rewriting history.

Tag Rules#

• Use definition tags for stable classification of the annotation itself, such as health, agent, research, or workflow.
• Use record tags for context that applies only to one logged occurrence, such as new-task, manual-test, backfill, or user-requested.
• Add definition tags with create --tag <tag>. Repeat --tag for multiple definition tags.
• Add record tags with record --tag <tag>. Repeat --tag for multiple record tags.
• --tag accepts either a Fulcra tag name or an existing tag UUID.
• Fulcra stores tags as UUIDs. The script resolves tag names to UUIDs and creates a missing tag name automatically.
• If record --tag is omitted, the record uses the definition tags.
• If record --tag is provided, the explicit record tags are sent for that record.
• Use lowercase, short, reusable tags. Prefer new-task over a full sentence.
• Tags should be filter dimensions, not prose. For places, prefer actual geography such as town-springfield, village-riverside, neighborhood-downtown, or place-main-campus over abstract tags such as scope-town or scope-neighborhood.
• Pair place tags with category/source/severity tags when useful, for example category-traffic, source-town-feed, or severity-advisory.
• Do not use tags for timestamps, detailed notes, values, people names, dedupe keys, or private context. Use --recorded-at, --note, --value, --source, and local metadata/ledger state for those.
• Existing definitions can be retagged with update --id ... --tag ... after user approval; --tag replaces the definition tag set.

Timestamp Rules#

• If --recorded-at is omitted, the script records the annotation at the current time.
• If the user asks for a historical or scheduled-looking time, pass --recorded-at.
• Use the user's local timezone when resolving relative phrases like "yesterday at 10am".
• Include the timezone offset in the timestamp. Do not pass naive local times.
• UTC readback is expected. Compare instants, not string equality.

Example: on 2026-05-16 in Eastern Time, "yesterday at 10am" means 2026-05-15T10:00:00-04:00, which readback may show as 2026-05-15T14:00:00+00:00.

Verification Rules#

• record returns recorded_at and verified_matches.
• verified_matches >= 1 means the script found the written record in Fulcra after ingest.
• uv tool run fulcra-api data-updates <window> can show that annotation-related data types were processed in a time range, but it is not record-level write verification. Do not use update counts as proof that a specific annotation record was created, changed, or deleted.
• For historical writes, use recent --hours with a window large enough to include the target time if a second verification is needed.
• When inspecting readback, use only minimal fields needed for confirmation: annotation name/id, recorded_at, value if relevant, and note if relevant.

API Notes#

Core REST endpoints:

• GET /user/v1alpha1/annotation lists annotation definitions.
• POST /user/v1alpha1/annotation creates a definition.
• PUT /user/v1alpha1/annotation/{annotation_id} updates a definition.
• DELETE /user/v1alpha1/annotation/{annotation_id} deletes a definition.
• POST /ingest/v1/record records annotation events.
• POST /ingest/v1/record/batch records batches of annotation events.
• Readback uses /data/v1alpha1/event/MomentAnnotation for moment/duration and /data/v1alpha1/metric/{BooleanAnnotation|NumericAnnotation|ScaleAnnotation} for metric annotation values.

For bulk or backfill pipelines, data-updates is a coarse freshness diagnostic after ingest. It supplements the local ledger and record readback; it does not replace either.

For schema details or upstream gaps, read references/api-notes.md.