agentsclimarketplace

Performance optimizer

Skill moses607/socialforge/skills/performance-optimizer

SocialForge — 17 open-source Claude skills that turn any capable model into an AI social-media growth agency: competitor analysis, hooks, viral scripts, repurposing, trends, engagement, and an orchestrator. MIT.

Install
npx -y skills add moses607/socialforge --skill performance-optimizer

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 24 days oldThe repository was created 24 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Post-mortems a piece of content and produces the next iteration, not just an explanation. It isolates the weakest link, proposes single-variable A/B tests, and applies a double-down rule when something wins. Use when someone asks "why did this flop", "why did this win", "help me optimize/improve my content", "set up an A/B test", or wants a post-mortem. Works with any capable model.

SKILL.md

4.5 KB, ~1.0k tokens by cl100k_base, as published. Nobody here has run it

Performance Optimizer

A post that flopped is a free experiment — but only if you extract one lesson and ship the next test. Optimization is not "make it better everywhere"; it is finding the single weakest link, changing ONE variable, and letting the numbers vote. Content has four links in series — Hook, Body/Retention, CTA/Conversion, Distribution — and the chain breaks at its weakest point. Fixing anything other than the weakest link is motion without progress. Winners are not luck to admire; they are formats to industrialize. Every result routes to one of three verbs: KILL, ITERATE, or SCALE.

1. Post-mortem — isolate the weakest link

  1. Pull the funnel: hook rate, retention/avg watch, saves+shares per view, follows-per-view, reach.
  2. Find the FIRST metric below the account's median — that is the weakest link. Attribute the outcome to it, not to a vibe.
  3. Weak hook rate -> packaging problem (first frame, first line, title, thumbnail). Good hook + retention cliff -> body problem (pacing, payoff, structure). Good retention + low saves/follows -> CTA/value problem. Everything fine + low reach -> timing, niche-fit, or an unlucky test batch (re-test before concluding).
  4. State ONE root cause in a sentence. If you can't, you're guessing — get more data.

2. Design single-variable A/B tests

  1. Change exactly ONE variable per test so the result is attributable. Multi-variable "improvements" teach nothing.
  2. Highest-leverage variables in order: hook line, first frame/thumbnail, first 3 seconds, format/structure, topic angle, CTA, length, posting time.
  3. Write the hypothesis as: "If I change [X], then [metric] improves, because [reason]." Keep everything else identical.
  4. Run 3-5 posts per variant before judging — a single post is noise; the algorithm's test audience varies wildly.
  5. Judge on the diagnostic RATE tied to the change (hook test -> hook rate), not on total views.

3. Double-down, and decide KILL / ITERATE / SCALE

  1. Double-down rule: when a post beats your median by ~2-3x, immediately make 3 more in the same format/angle/hook pattern while it's hot. Winners cluster.
  2. KILL: below median on hook AND value after 3+ attempts — the concept doesn't land. Stop; free the slots.
  3. ITERATE: mixed signals (strong hook, weak body, or vice versa) — keep the strong link, run one test on the weak link.
  4. SCALE: clear winner — replicate the pattern, vary only surface topics, and push volume. Turn the one-off into a series/template.
  5. Iteration loop: Ship -> read the one weakest link -> change one variable -> re-ship -> compare to median -> route to Kill/Iterate/Scale. Repeat weekly.

Output template

POST-MORTEM
- Result vs median: [win / flop / average] ([metric], [n] vs [median])
- Funnel read: hook [n]% | retention [n]% | saves+shares [n]% | follows/view [n]
- Weakest link: [hook / body / CTA / distribution]
- Root cause (one sentence): [...]

RANKED FIXES (highest leverage first)
1. [fix] — targets [metric]
2. [fix]
3. [fix]

NEXT 3 EXPERIMENTS (one variable each)
1. Change [X] -> hypothesis: [metric] improves because [reason]
2. Change [Y] -> ...
3. Change [Z] -> ...

DECISION: KILL / ITERATE / SCALE — [why]

Platform variants

  • Short video: iterate hooks and first-frames fastest; retention curve is the truth serum.
  • YouTube long-form: A/B thumbnail+title first (CTR), then intro (30s retention); tests take days, not hours.
  • Carousels/LinkedIn: test slide 1 / opening line and the save-worthy payoff; saves and dwell decide.
  • X: test the first line and the format (hook+list vs story); reposts and profile clicks judge it.

Rules

  • Isolate ONE weakest link before proposing any fix — no shotgun changes.
  • One variable per test, always — attributable or worthless.
  • Never conclude from a single post; require 3-5 before you kill or scale.
  • When something wins, make 3 more immediately — ride the pattern while the algorithm favors it.
  • Judge each test on the rate it targets, not on vanity totals.
  • Kill decisively. Dead concepts steal the volume your winners need.
  • Compare to the account's own median, never to an absolute or to someone else's numbers.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Gives 0 of the 12 instructions most test skills give in ~1.0k tokens

Counted across 964 of the 1,571 authors here whose files we hold, read 2026-08-07

  • Close the browser when donein 55 of 964, across 12 files
  • Wait for network idle statein 51 of 964, across 6 files
  • Launch Chromium in headless modein 49 of 964, across 6 files
  • Use descriptive selectors for elementsin 49 of 964, across 6 files
  • Run provided scripts with help flag firstin 49 of 964, across 6 files
  • Add appropriate explicit waitsin 48 of 964, across 5 files
  • Use bundled scripts as black boxesin 46 of 964, across 3 files
  • Do not read script source codein 46 of 964, across 3 files
  • Use sync playwright for scriptsin 46 of 964, across 3 files
  • Inspect dom before executing actionsin 46 of 964, across 3 files
  • Run the full test suitein 37 of 964
  • Write the failing test firstin 29 of 964, across 23 files

Said here and by no other author read

  • Isolate one weakest link before proposing any fix
  • Write the test hypothesis explicitly
  • Run 3-5 posts per variant before judging
  • Judge tests on diagnostic rates not vanity totals
  • Make 3 more posts immediately when a post wins
  • Route every result to kill iterate or scale

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 326,970. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.