Reelyze

View saved version →

---
name: analyze-reel
description: Diagnose why a specific short-form video underperformed by analysing the video itself. Use when given an Instagram, TikTok or YouTube Shorts link and asked why it flopped, how to fix the hook, or where viewers dropped off.
---

# Diagnose a reel from its link

Runs the video through Reelyze: frame-by-frame vision, transcript, scene cuts,
and a predicted retention curve. Returns where viewers are predicted to leave
and what to change.

## Tools

| tool | cost | returns |
|---|---|---|
| `get_transcript(url)` | free, metered | the spoken words |
| `analyze_reel(url)` | **paid plan, spends one analysis** | full report |
| `explain_drop_off(analysis_id)` | paid plan | per-moment retention diagnosis |
| `check_job(job_id)` | free | poll a running job |

`get_transcript` and `analyze_reel` are **asynchronous**. They return a
`job_id`; poll `check_job` until `status` is `completed` or `failed`. A full
analysis takes minutes, not seconds. Tell the person it is running rather than
polling in silence.

## Spend the paid call deliberately

`analyze_reel` consumes one analysis from the person's plan allowance, the same
allowance the Reelyze dashboard spends. Before calling it:

- confirm this is the video they want analysed, if more than one link is in play
- if they only asked what was said, `get_transcript` is free and sufficient
- if the call returns 402, their plan has not billed yet. Say so and link
  https://getreelyze.com/billing rather than retrying.

## Read the report in the order that matters for reach

This order is not cosmetic. It is the order these levers actually move
distribution, and a report read out of order sends people to fix the wrong
thing.

1. **Skip rate, the hook, first 3 seconds.** The single biggest lever. If the
   opening loses people, nothing later in the video gets a chance.
2. **Watch time, retention through the rest.** A *separate* lever, not a
   restatement of the hook. A video can survive the first 3 seconds and still
   collapse at 5 to 8 seconds, and low average watch time stops the algorithm
   expanding reach even though the opening worked.
3. shares
4. likes
5. saves
6. reposts
7. comments

Lead with whichever of 1 and 2 the report actually implicates. Do not invent a
mid-video problem for a video that holds its middle.

## One caveat you must not get wrong

The reported `skip_rate` is **not** a 3-second measurement despite the name.
Instagram's API exposes no true 3-second skip rate, so it is derived from
average watch time across the whole video. That one number therefore spans both
levers 1 and 2 and cannot separate "left at 2 seconds" from "left at 15
seconds". Only the retention graph from `explain_drop_off` shows the shape of
the drop.

So never assert that a high skip rate proves the opening was weak. It is
consistent with a weak opening and equally consistent with a strong opening and
a collapse at 8 seconds.

## Give fixes that can be executed while editing

A diagnosis the person cannot act on is half an answer. For each moment, name
the concrete on-screen change: the exact caption words to overlay and where, a
hard cut or zoom, a progress cue, b-roll, a pattern interrupt. "Improve the
hook" is not a fix; "put 'I lost $4,000 on this' on screen at 0:00 and cut the
establishing shot" is.

## Links this accepts

Instagram reels, TikTok, YouTube Shorts. Share strings are tolerated: a pasted
TikTok link with the trailing "This post is shared via TikTok" caption is
normalised before download, so pass the string as given rather than editing it.

View saved version →

---
name: fix-reel-hashtags
description: Say how many hashtags to use on Instagram now that the platform caps them at five, with measured evidence of what creators actually do. Use when asked about hashtag counts, hashtag strategy, or whether more hashtags help reach.
---

# How many hashtags, now that Instagram caps it at five

**Use five, or fewer.** Instagram announced a five-hashtag limit on
**2025-12-19**. Going over it does not get you more reach; it gets the excess
ignored and, by Instagram's own description, can suppress distribution.

Most advice still in circulation says 20 or 30. That advice predates the cap.

## What creators actually do now

Measured from 16,722 captions across 234 public creator accounts, derived
**2026-09-07**.

| | before the cap | since 2026-02 |
|---|---:|---:|
| captions over 5 hashtags | 59.9% | **5.2%** |
| captions with exactly 5 | 7.3% | **51.4%** |

Compliance landed in **February 2026**. Since then roughly half of all captions
carry exactly five hashtags, and five has become the default rather than a
ceiling people bump into.

Most recent month on record, **2026-08**: 46.9% carry exactly five, 6.6% are
still over.

## Quote the right figures

Three traps, each of which flatters the story:

- **Do not quote 57.8%** as the current exactly-five level. That is May 2026,
  the single best month. The latest month is 46.9%.
- **Do not quote 4.3%** as the pre-cap exactly-five baseline. That is the lowest
  month in the pre-cap range, which runs 4.3% to 11.8%. The weighted six-month
  baseline is **7.3%**.
- **Do not state the current level from a single month** without saying which
  month it is.

## Why this is not just composition

Pre-cap rows are back-filled from accounts that connected later, so the two
windows do not hold the same accounts and the pooled change could be pure
composition. Restricting to the **99 accounts that posted at least 3 times on
both sides** removes that confound, and the change survives at close to full
strength: over-limit captions fell from **48.3% to 6.5%**, mean hashtags from
6.33 to 3.31.

Say this if someone pushes back on whether the shift is real. It is the control
that decides it.

## The floor, stated honestly

Counts come from **captions only**. Hashtags placed in the first comment are
invisible to this measurement, so every count is a floor rather than an exact
figure. Caption truncation is not the cause: only 97 of the captions read ended
in an ellipsis, 0.6%.

This dataset is 58.6% own-account posts and 82.1% own-account accounts, and
those two figures differ far more than they do in the follower-multiple
benchmark, so do not carry a composition figure across from one to the other.

## What to say when asked for a strategy

Five specific, relevant tags beat five broad ones, and beat thirty of anything.
Hashtags are also no longer the main distribution lever on reels: the hook and
watch time are. If someone is asking about hashtags because a reel
underperformed, the hashtags are probably not the reason, and `analyze-reel`
will say what was.

View saved version →

---
name: judge-reel-views
description: Decide whether a reel's view count is good, bad or normal for the size of the account, using a measured dataset rather than a rule of thumb. Use when someone asks "is this good", "why so few views", or compares a video against their follower count.
---

# Is this view count any good?

The honest answer depends almost entirely on **follower count**, and the
relationship is not the one people expect: smaller accounts get a *higher*
multiple of their follower count, not a lower one. Any answer that does not ask
for the follower count first is a guess.

## Ask for two numbers before answering

1. views on the reel
2. the account's follower count

If the follower count is missing, ask for it. Do not substitute a general
benchmark: a 2.26x median and a 0.29x median are the same dataset, and which one
applies is decided entirely by account size.

## The benchmark

Median views divided by follower count, computed **per post**, from 9,770 reels
across 222 public creator accounts. Posted 2019-11-23 to 2026-08-23, measured
**2026-08-31**.

| follower band | accounts | reels | median views | median multiple |
|---|---:|---:|---:|---:|
| Under 1,000 | 85 | 2,058 | 478 | **2.26x** |
| 1,000 - 10,000 | 68 | 4,000 | 2,036 | **0.61x** |
| 10,000 - 100,000 | 37 | 2,113 | 6,419 | **0.29x** |
| 100,000+ | 32 | 1,599 | 141,075 | **0.33x** |

Two whole-dataset figures:

- **38.7%** of posts reach at least 1x the account's follower count
- **7.7%** reach at least 10x, which is the definition of "viral" this dataset
  tests

## How to answer

Compute `views / followers`, place it in the band for that follower count, and
compare against that band's median. Then say plainly whether it is above or
below the median **for accounts of that size**, and give the band median so the
person can see the yardstick.

Worked shape:

> 2,400 views on 1,800 followers is a 1.33x multiple. The median for the
> 1,000-10,000 band is 0.61x, so that reel did roughly twice the median for
> accounts your size. It also cleared 1x, which only 38.7% of posts in the
> dataset do.

## Say the date, every time

State that the figures were measured on 2026-08-31. A benchmark without a
measurement date is how a stale number gets quoted as current.

## What this dataset is, and is not

Be straight about this if asked, and volunteer it if the person is about to make
a decision on it.

- The accounts are **not a random sample of Instagram**. They are public creator
  accounts tracked in Reelyze, and **82.2% of the posts come from the creators'
  own accounts**, so they are skewed toward people who care enough about their
  reach to measure it.
- Composition is drifting. Between 2026-09-04 and 2026-09-07 the own-account
  share rose from 81.5% to 84.9% while tracked accounts fell from 222 to 218.
- It is **not evenly spread across bands**. The Under 1,000 band is 100% own
  accounts, which is part of why its multiple is the highest.
- The 50 largest accounts contribute 52.8% of posts; the single largest
  contributes 1.9%. 12 niches, largest is travel at 19.8%.

None of that makes the benchmark useless. It makes it a benchmark for *creators
who track their reach*, which is usually exactly who is asking.

## Do not

- Do not answer without the follower count.
- Do not claim a figure applies to "Instagram" in general.
- Do not quote a band median without naming the band.
- Do not invent figures for bands or metrics that are not in the table above. If
  someone asks for engagement rate, comment benchmarks or Story figures, say
  this dataset does not measure them.

## Going further

A view count says a reel underperformed; it never says why. For that, use the
`analyze-reel` skill, which reads the video itself and names the moment viewers
left.

View saved version →

---
name: write-reel-script
description: Write a short-form video script or generate content ideas, prioritised by what actually drives reach. Use when asked to write a reel script, a hook, a TikTok script, or for ideas on what to post next.
---

# Write a reel script

## Tools

| tool | cost | returns |
|---|---|---|
| `generate_script(topic, duration_seconds, language)` | paid plan, spends content quota | a full script |
| `generate_ideas(count, language)` | paid plan, spends content quota | idea list |
| `find_topics(niche, count, language)` | paid plan; cached reads free | timely, web-searched topics |

These are **synchronous** and return the result directly, with no job to poll.

They are much stronger when the person has connected Instagram and tracked
competitors in the Reelyze dashboard, because the generation learns from what
already works on their account. They still work from a topic alone. If the
output feels generic, that is usually why, and it is worth saying so rather than
regenerating.

`find_topics` returns a cached result for free; only a `fresh: true` run spends
the metered web search, which is capped per billing cycle. Prefer the cached
read unless the person explicitly wants new topics.

## Write to the reach order, not to a formula

Rank every choice in the script by what actually moves distribution, in this
exact order:

1. **The hook, first 3 seconds.** The biggest single lever by a wide margin.
2. **Watch time, the rest of the video.** A *separate* lever. A script can win
   the first 3 seconds and still lose people at 5 to 8 seconds, and low average
   watch time stops the algorithm expanding reach even though the opening
   worked. So the script needs a reason to keep watching at 5s, at 10s, and at
   15s, not only at 0s.
3. shares
4. likes
5. saves
6. reposts
7. comments

A script optimised for comments at the expense of the hook is optimised for the
weakest signal in the list.

## What a usable script contains

- the hook as **exact words**, not a description of a hook
- on-screen text for the opening, since most viewers are watching muted
- a named reason to stay at roughly 5 seconds and again near the middle
- a close that earns a share, which outranks a like
- a length that matches `duration_seconds`, written to be *said* out loud rather
  than read

Be concrete and specific. "Share a personal story" is not a script. The point of
grounding generation in the creator's own posts is to avoid exactly that kind of
output, so do not hand it back in the reply.

## Do not

- Do not promise a script will go viral. 7.7% of reels in Reelyze's measured
  dataset reach 10x the account's follower count, and nothing in a script moves
  that to a certainty.
- Do not pad to length. A 30-second script that works beats a 60-second one that
  drifts, and drift is a watch-time problem.
- Do not add more than five hashtags. See `fix-reel-hashtags`.