twitter-strategy - SKILL.md Agent Skill

name: twitter-strategy description: Use to analyse Oak's X account activity vs GROWTH.md targets and surface drift flags — post cadence, engagement patterns, who's replying, reply rate from priority audiences, reach drift, voice drift. Triggers on "how am I doing on X", "twitter strategy review", "am I on track", "GROWTH metrics check", "is the teach loop holding", "account-level review", "weekly X review", "is the strategy working", "are we on track for mainnet operator gate". Reads canonical docs (GROWTH, THESIS, BRAND); writes the review inline in chat. Distinct from x-algorithm-grader (per-post grading) — this is account-level over a 7-day or 30-day window.

Twitter strategy

Account-level activity vs GROWTH targets. Not per-post grading. Catches drift between what Oak is doing on X and what the GROWTH bet calls for.

Read first

GROWTH.md — the bet (§1), the bottleneck (§2), the current target cluster (§3), the GTM sequence (§4), the daily loop (§5), what we will not chase (§6 — split into §6.1 permanent rules and §6.2 tactical deferrals), metrics (§7), channel canon (§8).
THESIS.md — the structural argument the teach loop is supposed to teach.
BRAND.md — voice canon. "Lead from structure, not from fear."

This skill points at GROWTH; it does not restate it.

What this skill does

One mode: review. Three windows by default — last 7 days, last 30 days, and rolling. Outputs drift flags per GROWTH §5 loop bucket, per §6 will-not-chase rule (split §6.1 permanent vs §6.2 tactical), per §7 metric, and per §8 channel-canon claim.

When to run

Weekly, ideally Sunday evening or Monday morning before the Teach loop fires.
On-demand when drift is suspected (engagement dropped, voice slipped, reply rate fell off).
Before any strategic change to the daily loop.

Procedure

Step 1 — Pull account activity

Run:

bird user-tweets tannedoaksprout -n 200 --plain > /tmp/oak-recent-30d.txt
bird mentions -u tannedoaksprout -n 200 --plain > /tmp/oak-mentions-30d.txt

-n is a count, not a window. Post-filter both corpora to tweets with timestamp >= now - 30 days. If the earliest tweet in either corpus is still less than 30 days old after -n 200, paginate with --cursor until the cutoff is reached (cap at 5 pages).

Step 2 — Bucket posts by GROWTH §5 loop

Classify each original post (not retweet) into one of the four buckets named in GROWTH §5:

Understand — listening reply on someone else's substantive post; replies that surface a methodology question or extend the cluster's argument toward Jinn's frame.
Teach — public artefact on the thesis or the bridge model. Threads, essays, talks, recorded walkthroughs, originals carrying the canonical pitch.
Engage — direct offers and warm-list moves visible from public data (named asks, public DM-equivalents, quote-tweet outreach to specific candidates). Most synchronous Engage activity (DMs, calls) is invisible from public-only data; treat as [unknown — out of scope for this skill].
Refine — invocations of growth-refine proposing canon amendments. Visible from this account only when the user posts about a refine output; otherwise tracked in chat history, not on X.

Quote tweets are Teach if the QT itself adds substance; otherwise Engage. Replies are Understand by default; reclassify to Teach if the reply is a substantive standalone artefact.

Step 3 — Compute targets vs actuals

Derive the actual window from the corpus: span_days = first_tweet_ts − last_tweet_ts (in days, after the 30-day post-filter from Step 1). Working-day target = span_days × 5/7. Do not assume a literal 30d / 5-per-week — use the span the corpus actually covers.

GROWTH §5 implicit targets:

Teach: ≥1 per working day. Compute: count(Teach posts) / (span_days × 5/7).
Understand: ≥1 per working day. Same denominator.
Engage: weekly cadence of direct offers / warm-list moves to the named warm list. Cross-check against growth/.local/jinn-warm-contacts.csv last-contact dates and rung advancement if present.

GROWTH §7 metrics:

External testnet operators — surface count if known (manual input — out of scope here).
Inbound interest — count DMs / unsolicited mentions from §3 target-cluster handles in window. Use mentions corpus.

Step 4 — Detect §6 violations and §8 channel-canon drift

Permanent-rule violations (GROWTH §6.1). Hard-fail items, flagged unconditionally:

Fear-bait, empowerment-bait, or marketing register. Heuristic: posts containing "the future is", "we are so early", "this changes everything", "bullish doesn't even cover it", "if you don't do X you'll miss Y", or scare quotes around opponents. Flag each as a candidate violation; do not auto-fail without context.
Retired framings. Heuristic: literal substring match against Own What You Know, become a founder, your AI's experience is worth something, desired obsolescence, launch a token. Flag.
Founder framing. Posts that pitch from a separate-status position to the reader. Heuristic: posts using "we" referring to Oak + Ritsu without including the reader.

Tactical-deferral surfacing (GROWTH §6.2). Soft signals — surface as informational, do not flag as violation. Posts that touch a §6.2 tactical deferral (e.g. crypto-native cluster pitch language while Phase 1 is current) are noted in the review as deferral drift but only escalate to violations if §6.2 explicitly ratchets the deferral to permanent in a future spec.

Channel-canon drift (GROWTH §8). Surface when:

Premium-mandatory claim contradicted by a non-Premium post on the account being reviewed.
Reply-to-reply engine claim contradicted by a sustained pattern of replies received with no reply back from the author.
Cluster-fit claim contradicted by Teach posts written for a different cluster than GROWTH §3's current target.
Weekday cluster-peak claim contradicted by Teach posts shipping consistently outside the 09:00–14:00 cluster-peak window.
Constructive-tone overlay contradicted by negative / combative register in the corpus.

§8 drift signals point to a growth-refine candidate, not a §8 amendment direct from this skill.

Step 5 — Detect engagement drift

For each Teach post in the 30-day window, record reach (impressions if visible, replies, QTs). Compare to baseline (median of the prior 30 days, or absolute baseline ~10k impressions for a healthy Teach post in this cluster).

Flag drift in either direction:

Falling reach + same voice → cluster shifting; refresh cluster-model.
Rising reach + voice drift → recruiting wrong audience; re-anchor against BRAND.md.

Step 6 — Output

Print the structured review (format below) inline in chat. Do not write the review itself to disk — this skill is a lens, not a state file. The freshness stamp in Step 7 is the only persisted artefact.

If growth-day invokes this skill, the review feeds the day's drift flag.

Step 7 — Write the freshness stamp

After printing the review, write a freshness stamp to growth/.local/twitter-strategy-last-run.md. Overwrite each run; do not append. Format:

# Twitter strategy — last run

YYYY-MM-DDTHH:MM:SSZ
Window: 7d / 30d
Drift flags: N

This file is read by growth-day Step 0 to detect freshness without re-running the lens.

Output format

TWITTER STRATEGY REVIEW — YYYY-MM-DD (window: 7d / 30d)

§4 LOOP — actuals vs target
  Teach          [N posts]  vs target ≥5/wk      [PASS / DRIFT]
  Understand     [N replies] vs target ≥5/wk     [PASS / DRIFT]
  Direct offer   [N asks]   vs target ≥1/wk      [PASS / DRIFT]
  Interact       [unknown — out of scope]

§5 VIOLATIONS
  [N flagged] — [list each with URL + heuristic match]
  (or) (none in window)

§6 METRICS
  External operators (manual)     [if known]
  Inbound interest (mentions)     [N from priority audiences]
  Cluster signal (RTs, QTs)       [N priority-audience boosts]

ENGAGEMENT DRIFT
  Recent Teach post reach         [median impressions, replies, QTs]
  Trend                           [stable / falling / rising]
  Likely cause                    [voice / cluster / neither]

DRIFT FLAGS (top-3 surface for growth-day)
  1. [single sentence with action]
  2. [single sentence with action]
  3. [single sentence with action]

Voice constraints

British English. No emoji. Plain prose. Decision-shaped.
"DRIFT" and "PASS" labels only — do not editorialise pass/fail.
Action sentences are imperatives: "Ship one Teach post on outcome mining today", not "Consider posting more".
If a metric is genuinely unknowable from public data, say so: [unknown — out of scope]. Do not invent.

Composition

Inputs: canonical docs (GROWTH, THESIS, BRAND), bird CLI.
Outputs: chat review.
Consumed by: growth-day (incorporates drift flags into the day's brief).

What this skill does not do

Grade a specific draft. (That is x-algorithm-grader.)
Generate posts. (That is oak-content-strategy or x-post-builder.)
Modify GROWTH.md. (Changes to canonical docs go through spec/2026-04-28-canonical-docs.md proposal flow.)