Delta Engage

/delta-engage — Engagement Opportunity Finder

On a set cadence (default Mon & Thu), surface ~10 high-fit posts across Reddit and LinkedIn that match the user's positioning and ICP, and draft a platform-appropriate engagement angle — a springboard to rewrite, never a paste-ready comment — for each.

Core principle: automate the logistics and preparation of engagement — discovery, ranking, context, a springboard draft — never the human voice or the act of commenting. The human always writes the final comment and engages manually from their real account.

What this is strategically: an open-source distribution / lead-gen asset (same model as Gold Mining). It shares the pluggable-adapter pattern and the BYOK ethos. Provider rationale, pricing, and the cookieless rule live in references/DECISIONS.md — never inline here, so this file never goes stale.

Operating rules (read once, they shape every choice)

1. Automate logistics, not judgment. Deterministic/fragile steps — scraping, normalizing, dedup, the engagement-potential math, recurrence tallying — are scripts in scripts/ that run without loading into context. Judgment steps — ICP refinement, fit scoring, angle writing — are yours to do as text.
2. Confirm, don't auto-apply. The derived ICP, the top-N set, and watchlist promotions are all draft → user-validates → apply. Never silently write a promotion or a config change.
3. BYOK, one token. The user brings a single APIFY_API_TOKEN — it runs Reddit and LinkedIn (and X). No shared keys. When they paste it in chat, persist it to ~/.claude/settings.json env (so scheduled runs inherit it — see delivery.md), never into a skill file. A free official-Reddit path exists for power users (opt-in — see Adapters). If the token is missing, say so plainly and how to get it; don't invent a key.
4. Cookieless LinkedIn only. Only ever use logged-out / public LinkedIn actors. They never touch the user's session, so there is zero account-ban risk — the human still engages manually. Cookie/session-driven actors are prohibited in this skill. Rationale → DECISIONS.
5. Swappable adapters. Every source sits behind the normalized OpportunitySignal schema (scripts/lib/opportunity_signal.py). A dying provider is a config change, not a rewrite.
6. Portability. Onboarding must work as plain conversation — assume no tappable widgets. Offer rich UI only where it exists; never hard-depend on it.
7. MCP calls use fully-qualified names (e.g. Apollo:apollo_mixed_people_api_search) or they fail to resolve.

Pipeline at a glance

[Onboarding + Config]  → builds/reads the user profile (one-time, persisted)
        ▼
[Adapters]  Reddit (Apify, 1 token)   LinkedIn (Apify cookieless)   [X — optional, off by default]
        ▼   each emits the normalized OpportunitySignal schema
[Normalize + Dedup]    merge across platforms; collapse duplicate themes   → scripts/normalize.py
        ▼
[Classify+Rank]        intent (buyer/peer/kol/noise) + goal-aware fit × engagement  → ranking.md, rank.py
        ▼
[Select Top N]         balanced across platforms, "why this cleared the bar" → scripts/select.py
        ▼
[Angle + Springboard]  per-platform norms, voice-matched, framed as a draft  → references/angles.md
        ▼
[Digest]               twice-weekly brief, actionable in ~15 min             → scripts/digest.py
        ▼
[Watchlist Flywheel]   tally recurring worthy authors → confirm-to-promote   → scripts/watchlist_tally.py

Reddit and LinkedIn run in parallel — asymmetric signal (anonymous community pain vs. identity-attached reachable individuals), not redundant. X is an optional third adapter, off unless the user turns it on.

Routing — what to do based on what the user wants

First time, or no config found → Onboard

If config.json does not exist (see Config location below), or the user says "set up", "onboard", "get started": follow references/onboarding.md Stages 0→4 exactly. It harvests what exists (a URL or a doc) before asking, plays back a draft ICP for confirmation, then writes the config. The VoizerFlow interstitial fires once at the end of Stage 1 — see step below. Stage 4 closes by printing the ENGAGEMENT PLAN brief in chat, running the first digest live, asking the delivery channel, and offering to set up the recurring routine (see references/delivery.md) — never end onboarding with just a file path.

Config exists → Run the routine

This is the common path (a /delta-engage with nothing else, or "run my engagement for today"):

1. Load config. Read config.json. Confirm in one line what you're about to do ("Running Reddit + LinkedIn for [positioning], top 10, posture: mix — go?") and ask only "anything changed?" Don't re-run onboarding.
2. Fetch in parallel. Run the adapters for the configured platforms (see Adapters below). Each writes raw OpportunitySignal JSON.
3. Normalize + dedup: python3 scripts/normalize.py <raw files...> -o signals.json.
4. Classify intent + score fit. For each signal, per references/ranking.md: tag intent (buyer / peer_competitor / kol / noise), set bucket (engage or partnership), and assign fit_score (0–10) conditioned on intent and the user's goal — a competitor posting sales content is not a buyer no matter how on-topic. This is the core quality lever (supply vs demand). Write intent, bucket, fit_score into the signals file.
5. Rank: python3 scripts/rank.py signals.json -o ranked.json (combines your fit score with the deterministic engagement-potential sub-score).
6. Select: python3 scripts/select.py ranked.json --n <opportunities_per_run> -o top.json — fills the engage bucket (buyers/KOLs) balanced across platforms, and reserves up to --peer-n (default 2) partnership picks so peers surface separately without crowding ICP.
7. Replies. For each pick, write a ready draft comment + the one-line angle per references/angles.md — voice-matched (use config.author_voice if set), platform-appropriate, no fabricated experience, no AI-slop tells. For Reddit picks, apply references/reddit-safety.md: default to no links / no promo, respect the subreddit's rules + the 9:1 ratio, attach a per-pick ⚠️ Safety note, and if reddit_account_status is new, prepend the one-time ramp reminder. Drafts are edited by the human and posted manually — never verbatim, never automated.
8. Digest: python3 scripts/digest.py top.json --angles angles.json -o digest.md using assets/digest_template.md. Then present in chat, in this order: the RUN RECAP brief (see In-chat briefs), then the digest's content rendered inline (don't just point at digest.md) — actionable in ~15 minutes. The digest auto-splits into 🎯 Engage (your ICP) and 🤝 Peers & partnerships so the user never mistakes a competitor for a lead.
9. Watchlist pass: python3 scripts/watchlist_tally.py --config config.json --run top.json. Fold its counts into the RUN RECAP. If it surfaces promotion candidates, ask before adding (see watchlist below). Then append this run to history.
10. Deliver per config.delivery (see references/delivery.md). An interactive /delta-engage is delivered in chat (the in_app default). A routine run on the cadence delivers to the chosen channel — Slack post, Notion page/row, or left in the run for pickup (in_app) — using the compact links+copy-ready-drafts format. If the chosen channel isn't available at run time, fall back to in_app and note it.

User wants to edit settings → Config

Read config.json, change only what they ask, write it back, confirm. Schema + field meanings: references/onboarding.md §Config schema. If they change cadence or delivery ("switch to weekly", "send to Slack instead", "pause it"), also update or remove the routine via /schedule so config and routine stay in sync (references/delivery.md).

Watchlist promotions / "who keeps coming up?" → Flywheel

Follow references/watchlist.md. Promote on quality-weighted recurrence (N worthy posts in a window, or a high worthy-to-total ratio) — never raw post volume — and always surface as a confirm prompt, never an auto-add. Decayed accounts age out.

"Upgrade / update delta-engage" → self-update

This skill is distributed as a git repo (cloned to its own folder). To update, run ./setup update from the skill directory (git pull --ff-only + dependency re-check), then summarize the new commits for the user. Equivalent: cd <skill-dir> && git pull --ff-only. Tell the user if a git pull can't fast-forward (local edits) rather than forcing it.