Write Benchmark Thread

Write benchmark thread

Turn benchmark/eval results into a high-signal X/Twitter thread in a credible human voice. Optimized for reach without slop: specific numbers, honest scope, no AI prose tells, no over-claiming.

Before you write anything — load context (do not skip)

1. Read tweet-craft/FEEDBACK_LOG.md in full. Every entry is a real mistake a reviewer caught. Treat them as hard constraints. A fresh agent that skips this file repeats these mistakes.
2. Read tweet-craft/THREAD_STRATEGY.md for the algorithm weights, hook data, METR template, and anti-slop research.
3. Read tweet-craft/HERO_IMAGE_PROCESS.md before making or requesting a hook image. The hero conclusion changes per analysis; preserve the reusable conclusion-first structure, not Ponytail-specific wording.
4. Load the clear-writing skill and apply Strunk's rules (active voice, concrete language, omit needless words). Load its references/elements-of-style/03-elementary-principles-of-composition.md for depth.
5. Gather the actual results data — paired deltas, difficulty buckets, solve counts, token/cost/patch sizes. Never invent a number. If a number isn't in the source data, say so or omit it.

The process

1. Find the surprise. What would make a technical reader stop scrolling? State the tension (often "X and Y are both true"). This is the hook.
2. Draft using the anatomy (see THREAD_STRATEGY.md):
   • Hook (surprise + tension + specific number, no question, no list)
   • Preview (one-line payoff — the most-skipped tweet; do not skip)
   • Setup (the numbers, stated plainly, no "headline N ways" preamble)
   • Mechanism (why the result, not just what)
   • Cost / honest counter-result
   • Evidence (one concrete named example, never a wall of task names)
   • Synthesis (the generalizable lesson, scoped to what was tested)
   • Close (quotable standalone line + repo link + real question = reply bait)
   • Credit reply (tag every creator you built on)
3. Run the pre-flight checklist (below). Fix everything before review.
4. Render for human review over Tailscale (see "Review loop" below).
5. Apply feedback → append to FEEDBACK_LOG.md → redraft. This is the loop. If feedback is not logged, the next agent can't benefit and the iteration fails its purpose.

Pre-flight checklist (must pass all before review)

Accuracy:

• No superlatives or unverifiable claims ("most popular", "the best").
• Every number is in the source data. None invented.
• Control and treatment named precisely (e.g. "unmodified Pi agent" vs "the Ponytail Pi extension"), never vague like "bare agent".
• Scope claims to what was measured. No generalizing one skill to all.

Voice (anti-slop — each is a logged failure mode):

• No setup phrases ("The headline, three ways", "Let me show you", "Here's the thing", "Worth noting").
• No self-referential closers ("in one sentence", "that's the takeaway").
• No strained metaphors. If a phrase needs parsing, rewrite it plainly.
• No false certainty. Use "can" where you mean possibility, not "does".
• No em-dashes as a connective in every clause.
• Active voice, concrete nouns, no needless words (Strunk §III).

Distribution:

• Hook has personal pronouns and a real question or tension (not a yes/no).
• Hook tweet has a conclusion-first hero image (images boost engagement +30-120%; the hook is the highest-value slot). It must be a dedicated scroll-stopping hero visual, NOT a dense analytical chart. Use HERO_IMAGE_PROCESS.md: preserve the reusable anatomy, but choose a new headline/conclusion/metrics for each analysis.
• No links or @s in the hook tweet.
• Each image is referenced by the tweet it attaches to.
• Primary metric is labeled inline so readers know what to weight.
• Credit tweet tags every creator; handles verified on X.
• Repo link only in the close, never the hook.

Review loop (Tailscale, no Pi Annotate extension)

The pi-annotate extension's MV3 service worker sleeps and breaks auth on remote machines. Use Lavish bound to the Tailnet IP instead:

cd <repo>
mkdir -p .lavish && cp <rendered>.html .lavish/   # + any relative assets
# lavish serves the file's directory; relative asset paths only, no leading /
LAVISH_AXI_HOST=100.112.72.93 LAVISH_AXI_LINK_HOST=100.112.72.93 \
  npx -y lavish-axi .lavish/<rendered>.html
# share the printed session URL with the reviewer's browser
# in another window, poll for feedback:
LAVISH_AXI_HOST=100.112.72.93 LAVISH_AXI_LINK_HOST=100.112.72.93 \
  npx -y lavish-axi poll .lavish/<rendered>.html

Lavish feedback posts over HTTP to the served origin, so it reaches the desktop poll directly — no daemon pairing needed. Verify laptop→desktop fetch is HTTP 200 (session + artifact + first image) before asking the reviewer to act.

After feedback arrives: apply each annotation, append every lesson to FEEDBACK_LOG.md as a (mistake → fix → why) entry under the right category, then --agent-reply and re-poll.

Images and charts

Use tweet-craft/HERO_IMAGE_PROCESS.md for the hook image. The hero should be conclusion-first and reusable across analyses: different headline, different facts, same disciplined anatomy.

Use the benchmark-social-graphics skill for all chart data. Deterministic overlay only — never let an image-gen model place chart text or numbers (it hallucinates them). Image-gen is for styled base cards or craft references only; final benchmark facts must be rendered by code.

What NOT to do

• Don't ship a draft that hasn't passed the pre-flight checklist.
• Don't apply reviewer feedback without also appending the lesson to FEEDBACK_LOG.md — unlogged feedback is lost to the next agent.
• Don't repeat a mistake already in FEEDBACK_LOG.md.
• Don't generalize claims beyond the experiment.
• Don't optimize for likes; replies (+27×) and bookmarks (+10×) dominate the ranking. Likes are nearly worthless by comparison.

References

• tweet-craft/FEEDBACK_LOG.md — the growing lessons ledger (read first)
• tweet-craft/THREAD_STRATEGY.md — algorithm weights, hook data, METR template
• tweet-craft/HERO_IMAGE_PROCESS.md — reusable conclusion-first hero image process
• tweet-craft/PONYTAIL_THREAD_REWRITE.md — a worked example + its review notes
• clear-writing skill — Strunk + AI-writing-tell reference
• benchmark-social-graphics skill — deterministic chart overlays
• x-thread-fetch skill — extract existing threads for reference/templates