Kjell Tore Guttormsen
65b60337fe
Artikkelforslaget blir en persistert entitet (steg 2, A1-4) og godkjenningen får
et hjem (steg 3, A1-7) — før dette genererte agenten vinkel/begrunnelse som ble kastet.
Åtte additivt-valgfrie felt på TrendRecord/TrendInput/TrendItem:
- Forslag (A1-4/A1-5): angle · targetLevel (fritekst, brukerdefinert spenn — aldri enum) ·
rationale · relatedIds[] (flerkilde).
- F7 leser-side: actionability {formulated, note?} (N7-bånd-cap-gaten leser `formulated`) ·
verdict BÆRENDE/STØTTE/NYHET (lukket mekanisme-vokab).
- F9 leser-side: readerQuestion · painPoint · saturation (N7.5-sveipet fyller dem).
Ingen schema-bump: feltene er additivt-valgfrie, ingen record trenger migrering →
SCHEMA_VERSION forblir 4 (loadStore Math.max håndterer det). First-sight-persistering
(re-capture unionerer topics + re-scorer kun; klobrer aldri en triagert vinkel).
Validering: typede felt (verdict/actionability) feiler hardt ved malformert input;
fritekst/id-liste normaliseres lempelig (summary/topics-idiomet).
Livssyklus: TrendStatus += "selected" → new→selected→acted|skipped. Ny select-verb +
--ids-batch (act/skip/reset/select); partiell suksess = exit 0 + miss-rapport, all-miss = exit 2.
setStatusMany(): ren batch-mutasjon, per-id found/notFound.
Brief: forslagsfeltene rendres per kandidat (detaljert i topp-treff, kompakt token i bullets) +
ny «🚧 I produksjon»-seksjon (selected=valgt + acted=skrevet), pillar-uavhengig, deterministisk
sortert. selected/acted forlater arbeidskøen men vises i produksjons-boardet.
commands/trends.md Step 5 oppgradert: triage → Velg (select) / Skip, batchet per verb (--ids).
Trends-suite 245→266 (ny floor, +21 N6-tester). To eksisterende tester oppdatert for den
endrede brief-kontrakten (acted vises nå i I produksjon; ranking-descriptoren ekskluderer selected).
tsc rent. Roundtrip bevist: capture m/ alle felt → query → brief rendrer feltene → select → I produksjon.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S7SQpXJpBSNvpWaTNq1kWZ
142 lines
7.3 KiB
Markdown
142 lines
7.3 KiB
Markdown
---
|
||
name: linkedin:trends
|
||
description: |
|
||
Run a trend discovery pass over the user's own content pillars and source list:
|
||
delegate the scan to the trend-spotter agent, make sure kept candidates are persisted
|
||
to the trend store (dedup) and the dated morning brief is written, then return a
|
||
triage-ranked candidate list the user resolves per id (select/skip, batched). Default scoring mode
|
||
is long-form (chronicle/newsletter material); `--mode kortform` overrides for feed posts.
|
||
Use when the user wants a discovery pass, a trend scan, or a morning-brief refresh.
|
||
Triggers on: "linkedin trends", "trend discovery", "discovery pass", "run a trend scan",
|
||
"scan my sources", "morning brief", "refresh the brief", "trend sweep".
|
||
allowed-tools:
|
||
- Read
|
||
- Bash
|
||
- Task
|
||
- AskUserQuestion
|
||
---
|
||
|
||
# LinkedIn Trend Discovery
|
||
|
||
You are a thin discovery orchestrator. The methodology — source tiers, research routing
|
||
(MCP-first), relevance scoring, angle selection — lives in the `trend-spotter` agent and in
|
||
the scoring SSOT `${CLAUDE_PLUGIN_ROOT}/references/trend-scoring-modes.md`. Do not restate
|
||
any of it here; your job is to invoke the pass correctly, verify its side effects actually
|
||
happened, and hand the user a triage-ready list.
|
||
|
||
Data dir shorthand used below: `${DATA}` = `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}`.
|
||
Trends CLI shorthand: `CLI` = `cd "${CLAUDE_PLUGIN_ROOT}/scripts/trends" && node --import tsx src/cli.ts`.
|
||
|
||
## Step 0: Parse flags
|
||
|
||
All flags are optional, given after the command name:
|
||
|
||
| Flag | Meaning | Default |
|
||
|------|---------|---------|
|
||
| `--mode kortform\|long-form` | Scoring mode (see the SSOT for what each rewards) | **long-form** — no arguments means a long-form discovery pass |
|
||
| `--fresh-days N` | Freshness window for the morning brief | CLI default (7) |
|
||
| `--brief-only` | Skip the discovery poll entirely; render the brief from the existing store | off |
|
||
| `--dry-run` | Poll + score, but persist nothing: no capture, no brief, no status writes, no last-run marker | off |
|
||
|
||
Note the mode inversion deliberately: the **agent's** own default is kortform, this
|
||
**command's** default is long-form. That is why Step 2 must always pass the mode explicitly.
|
||
|
||
## Step 1: Load context
|
||
|
||
1. **Pillars:** Read `${DATA}/profile/user-profile.md` and extract the content pillars /
|
||
expertise areas. If the file does not exist, ask the user for their pillars before
|
||
proceeding (one question, comma-separated answer).
|
||
2. **Source list:** Resolve which list this pass will use — `${DATA}/trends/sources.md` if it
|
||
exists, otherwise the shipped defaults `${CLAUDE_PLUGIN_ROOT}/config/trends-sources.template.md`.
|
||
Tell the user which one applies. Do not read research-tooling or route research yourself —
|
||
the agent owns Research Routing (its own "Research Routing" section reads the profile's
|
||
`### Research Tooling` block); duplicating it here would drift.
|
||
|
||
## Step 2: Run the discovery pass
|
||
|
||
**If `--brief-only`:** skip the agent entirely — go to Step 3 and render the brief from the
|
||
existing store.
|
||
|
||
Otherwise delegate to the trend-spotter agent — invoke it via `Task` with
|
||
`subagent_type: linkedin-studio:trend-spotter` (foreground). The prompt MUST state explicitly:
|
||
|
||
- **The scoring mode** from Step 0 (default long-form). Never omit it — the agent falls back
|
||
to kortform when the caller is silent.
|
||
- The pillars and the resolved source-list path from Step 1.
|
||
- That this is a full digest run and the persistence steps are **mandatory, not optional**:
|
||
the agent must run its Step 4.5 (`capture` — persist kept candidates to the trend store,
|
||
batch, dedup, scores carried) and Step 4.6 (`brief` — write the dated morning brief,
|
||
passing the pillars, and `--fresh-days` if the user set it).
|
||
- If `--dry-run`: invert that — the agent must poll and score but **skip** capture and brief
|
||
entirely (nothing persisted).
|
||
- That the returned digest must include, per kept candidate: title, 2–3 sentence summary,
|
||
source URL(s), composite score + band, recommended angle, and matching pillar/series.
|
||
|
||
## Step 3: Verify the side effects (skip on `--dry-run`)
|
||
|
||
Trust but verify — the pass is only done when its artifacts exist:
|
||
|
||
1. **Store:** `CLI status --json` — confirm the store mutated (captured count reflects the
|
||
run; on a no-new-trends day `{added: 0, merged: N}` is a fine outcome, not a failure).
|
||
2. **Brief:** confirm today's file exists: `${DATA}/trends/morning-brief/<today YYYY-MM-DD>.md`.
|
||
If the agent captured but failed to render the brief, render it directly — the brief is
|
||
deterministic: `CLI brief --pillars "<pillar1,pillar2,…>"` (add `--fresh-days N` if set).
|
||
3. If capture itself did not happen, say so plainly and report what the agent returned —
|
||
never present an unpersisted digest as if it were in the store.
|
||
|
||
## Step 4: Present the triage-ranked list
|
||
|
||
Present the candidates ranked highest composite first (the agent's digest already carries the
|
||
ranking — do not re-rank). Per candidate, the contract is:
|
||
|
||
```
|
||
N. [Title] (id: <trend-id>)
|
||
Score: X.X — [Band] | Pillar: [pillar/series]
|
||
[2–3 sentence summary]
|
||
Source: [URL(s)]
|
||
Angle: [recommended angle]
|
||
```
|
||
|
||
Include each candidate's store id (shown in the brief and via `CLI list --json`) — the triage
|
||
step below resolves per id. On `--dry-run`, present the same list but say clearly that nothing
|
||
was persisted and there are no store ids to triage.
|
||
|
||
## Step 5: Triage (skip on `--dry-run`)
|
||
|
||
Resolve the top of the queue now instead of leaving it as homework. For the candidates in the
|
||
top bands (Immediate + High; cap at 8), use AskUserQuestion — one question per candidate, up
|
||
to 4 candidates per call, options:
|
||
|
||
- **Velg** — you'll write about this: mark `selected`, moving it onto the brief's "I produksjon"
|
||
board (valgt) so the queue stops re-surfacing it while it's in progress
|
||
- **Skip** — not for me: mark `skipped` (dropped from the queue)
|
||
- **Leave** — keep it in the queue untouched
|
||
|
||
Then apply the decisions through the store CLI. **Batch by verb** — collect all the "Velg" ids
|
||
and all the "Skip" ids and resolve each set in ONE call (ten candidates ≤ two calls):
|
||
|
||
```bash
|
||
CLI select --ids <id1,id2,id3> # everything chosen this pass
|
||
CLI skip --ids <id4,id5> # everything rejected this pass
|
||
```
|
||
|
||
Single-id form (`--id <id>`) still works for a one-off. "Leave" means no call. A partial batch
|
||
(some id unknown) still applies the matches, reports the misses, and exits 0. Finish with a
|
||
one-line summary: N valgt, N skipped, N left in queue. (`act` — already written — and the
|
||
auto-`act` when an edition reaches scheduling arrive with the N7 trend→newsletter bridge.)
|
||
|
||
## Step 6: Write the last-run marker (skip on `--dry-run`)
|
||
|
||
On completed runs (including `--brief-only`), stamp the marker so other surfaces can tell when
|
||
discovery last ran:
|
||
|
||
```bash
|
||
mkdir -p "${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/trends" && \
|
||
date -u +"%Y-%m-%dT%H:%M:%SZ" > "${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/trends/.last-run"
|
||
```
|
||
|
||
## Reference Files
|
||
|
||
- `${CLAUDE_PLUGIN_ROOT}/references/trend-scoring-modes.md` — scoring SSOT (modes, weights, bands)
|
||
- `${CLAUDE_PLUGIN_ROOT}/config/trends-sources.template.md` — shipped source-list defaults (user override: `${DATA}/trends/sources.md`)
|
||
- `${CLAUDE_PLUGIN_ROOT}/agents/trend-spotter.md` — the discovery methodology this command invokes
|