agentleFS
Murtadha-NajemMurtadha-Najem/instagram-account-audit/skills/account-audit/SKILL.md

account-audit

Skill0 starsChanged 52 days agoMITView on GitHub

Full audit of any Instagram account from its link or handle, of any kind (company, brand, shop, creator, artist, memes, news, public figure) and any size. Collects every post, reel, carousel, comment, liker, tagged post, partner and highlight, recognises the people on screen, then produces a PDF report in two linked halves - what the audience engages with and why, and an expert review of the content craft (script, wording, delivery, filming, editing, design, caption) apart from engagement - plus before and after rewrites and a playbook. Use when the user asks to analyse, audit or evaluate an Instagram account or page ("حلل حساب", "سوي تقرير عن هذا الحساب", "audit @x", "/account-audit <link>"). Not for a single reel (use /reel) or for an account's reposts (use /repost-profile).

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. davidecac/pipeboard-skills-adv★ 2same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository2 other files it gives its agents.

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. About 0 are common, 14 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈14

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    The reel skill is required: If preflight says it is missing, ask the user to install /reel first, and stop.

    line 25
  2. Distinctive:

    Cookies: never ask the user to paste cookie contents.

    line 27
  3. Very specific:

    Read audit/recon/recon.md and look at audit/recon/sheet.jpg.

    line 36
  4. Distinctive:

    Private accounts: only what the login can see.

    line 43
  5. Distinctive:

    Ask in one message. Everything is optional: the audit runs from nothing, but each answer improves it.

    line 48
  6. Very specific:

    If it is theirs or a client's and a business account: ask for the owner's data (references/owner_data.md: the Meta Business Suite export, …

    line 53
  7. Distinctive:

    Pronouns: only as the user gives them.

    line 92
  8. Distinctive:

    Rate limit: stop and resume later; nothing collected is lost.

    line 111
  9. Very specific:

    Query ids changed: run PY SKILL/scripts/capture_doc_ids.py, then retry.

    line 112
  10. Distinctive:

    403 or a restricted login: stop at once and tell the user.

    line 113
  11. Distinctive:

    After the pipeline, run these:

    line 124
  12. Very specific:

    put the names in face_names in account.json.

    line 127
  13. Distinctive:

    Read both files yourself. If the brief contradicts what you saw or what the user said, fix it before stage 6.

    line 144
  14. Distinctive:

    If it lists questions for the user that would change the analysis, ask them now in one message.

    line 146
  15. Distinctive:

    Starting batches: Every time more posts are ready, run:

    line 155
  16. Distinctive:

    If an agent's output is blocked by content filtering, a transcript holds lyrics: relaunch that batch with a reminder of the lyrics rule.

    line 176
  17. Distinctive:

    brief mode: ask for half the length.

    line 201
  18. Distinctive:

    Check that every number you see is in a facts file, and that names are spelled as the user gave them.

    line 231
  19. Very specific:

    If the numbers differ a lot, correct cost_per_post in config.json so the next estimate is closer.

    line 241
  20. Distinctive:

    Scripts count; agents judge: No agent computes a number a script has.

    line 245
  21. Very specific:

    Run collect again with a new --since.

    line 256
  22. Very specific:

    For a comparison with an earlier audit: keep the old analysis/ folder under a dated name.

    line 258

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/account-audit

Everything runs from the project folder in scripts/config.json (project_root) with its Python (PY below). Skill files: ~/.claude/skills/account-audit\ (SKILL below). Scripts run from SKILL\scripts. Output for account <u>: profiles\<u>\audit\. The reel pipeline is used as it is: never edit reel/, reel.py, batch.py, look.py or the reel skill from here.

Speak the user's language in every message: questions, the cost table, progress, delivery. The report is in the user's language unless they ask otherwise (report_language in account.json).

The standard for every finding: it rests on posts you or the agents actually looked at, cites them by topic, date and ID, and never borrows general social-media advice that this account's posts do not show.

0. Ready check and the reel skill

PY SKILL/scripts/preflight.py
  • The reel skill is required. If preflight says it is missing, ask the user to install /reel first, and stop.
  • Other missing pieces: relay the fix the script prints. Face models missing only means stage 4 is skipped.
  • Cookies: never ask the user to paste cookie contents. They export them with the browser extension "Get cookies.txt LOCALLY" while logged in.

1. First look (a handful of requests, before anything is approved)

PY SKILL/scripts/recon.py "<link or handle>"

Read audit/recon/recon.md and look at audit/recon/sheet.jpg. From them, tell the user in a few lines:

  • what the account seems to be (type, language, what it posts);
  • its size;
  • how many posts fall in the window (default: the last 12 months).

The full understanding comes later, in stage 5, from a real sample.

Private accounts: only what the login can see. Someone else's account: fine when it is public. The report stays with the user.

2. Ask, then show the cost, then wait

Ask in one message. Everything is optional: the audit runs from nothing, but each answer improves it. Ask only what applies:

  1. Window: the last 12 months, or another range?
  2. Whose account: the user's own, a client's, or a third party's?
    • If it is theirs or a client's and a business account: ask for the owner's data (references/owner_data.md: the Meta Business Suite export, boosted posts, audience and reel screenshots). Say it is the most useful single addition.
  3. Group accounts: the account's own other accounts (a sister brand, a magazine, the founder's page), so posting with them is not counted as a partnership.
  4. Team and faces: who works on the account, with their Instagram handles and one or two posts where each person clearly appears. This names the faces and separates team likes from the audience. Without it, recurring faces become "person 1, 2 ..." and team likes are found by behaviour alone.
  5. Brand: colours and a logo for the report, if they want the account's own look.
  6. Anything else the posts cannot show: goals, campaigns, known events.
  7. Large accounts only (recon size_class large or huge):
    • Collection is never reduced. It runs at the slow pace, and takes longer.
    • Recommend a spare Instagram login for collecting, so the user's main account carries no risk. Its cookies go in a separate file, passed with --cookies.
    • Offer review of all posts, or of a stratified sample, with the cost of each (modes_sampled_150 in recon.json).

Then show the modes as a table, with this account's numbers from recon.json:

  • the time each takes, and how long collection alone takes;
  • the approximate tokens each uses;
  • a one-line "what you get" for each.
modewhat you get
fullboth halves, the link between them, rewrites, errors, playbook, appendix
engagementwhat the audience responds to, who engages (genuine against internal), timing, partners: no craft review
contentthe expert craft review of every post and the panel verdicts, apart from the numbers: no likers or comments collection
briefengagement on every post, craft review on 20 to 25 posts, about six pages

State the assumptions in one line: the estimate scales with post count and type mix, and Gemini free keys are used for audio. Do not start until the user picks a mode and says go. Write what they said to audit/account.json:

{"username": "", "mode": "full", "since": "YYYY-MM-DD", "report_language": "ar", "relationship": "own | client | third_party",
 "utc_offset_hours": 3, "group_accounts": [], "display_name": "",
 "team": [{"name": "", "display": "name as written in the report language", "handles": [], "reference_posts": [], "pronoun": null}],
 "face_names": {}, "brand": {"colors": {}, "logo_light": null, "logo_dark": null},
 "cookies": null, "pace": "normal | slow", "house_style_no_dashes": true, "notes": ""}
  • Pronouns: only as the user gives them. Never infer one from a name.
  • Names: spelled exactly as the user writes them.

3. Collect (background, resumable; slow pace for large accounts)

PY SKILL/scripts/collect.py posts <u> [--since YYYY-MM-DD] [--cookies PATH] [--pace slow]

Then start the rest in the background, as one command, while stage 4 runs. The layers depend on the mode:

  • full, engagement, brief: comments, likers, tagged, partners, highlights, slides.
  • content: highlights and slides only.
PY SKILL/scripts/collect.py <layer> <u>

COLLECT ERROR: relay it plainly.

  • Cookies expired: the user exports them again.
  • Rate limit: stop and resume later; nothing collected is lost.
  • Query ids changed: run PY SKILL/scripts/capture_doc_ids.py, then retry.
  • 403 or a restricted login: stop at once and tell the user. Never push on.

4. Pipeline, profiler sample, faces, delivery

  1. PY SKILL/scripts/batches.py <u> sample, then run the pipeline on the sample first: PY SKILL/scripts/run_pipeline.py <u> --codes <audit>/profile_sample.txt.
  2. Then the full pipeline in the background: PY SKILL/scripts/run_pipeline.py <u>. It is resumable, uses the account's cookies and pace, and stops by itself if Instagram refuses.
  3. Faces run alongside it, unless the face models are missing:
    • faces.py <u> extract, repeated as posts finish (a loop every few minutes is fine);
    • faces.py <u> register when the user gave reference posts.
  4. After the pipeline, run these:
    • faces.py <u> extract, then faces.py <u> unknowns;
    • show the user faces/unknowns.jpg and ask who the recurring people are (optional; unnamed stay numbered);
    • put the names in face_names in account.json;
    • faces.py <u> match.
  5. PY SKILL/scripts/delivery.py <u> after the pipeline (Gemini listens to every video with speech).
  6. PY SKILL/scripts/owner_data.py <u> if the owner sent files.

Posts the pipeline could not transcribe (Gemini 429 or 503):

  • Retry them later with PY reel.py "https://www.instagram.com/p/<code>/". The download is cached, so only the analysis reruns.
  • batches.py holds them back until they are transcribed, or until you pass --allow-untranscribed.

5. Understand the account (one agent, before any coding)

As soon as the sample is through the pipeline, start one general-purpose agent:

Read SKILL/instructions/profile.md and follow it exactly. Account folder: <audit>. Report language: <lang>. Write <audit>/analysis/account_brief.md and <audit>/analysis/taxonomy.json.

Read both files yourself. If the brief contradicts what you saw or what the user said, fix it before stage 6.

If it lists questions for the user that would change the analysis, ask them now in one message. Carry on without an answer if the user does not reply.

6. Two review waves (agents, streamed while the pipeline runs)

The wave depends on the mode:

  • Content wave: full, engagement and brief.
  • Craft wave: full, content and brief.

Starting batches. Every time more posts are ready, run:

  • PY SKILL/scripts/batches.py <u> content for the content wave;
  • PY SKILL/scripts/batches.py <u> craft [--sample 25 | --basics] for the craft wave.

Then run the same command with --flush once the pipeline is done.

  • brief mode: --sample 25.
  • content mode: --basics, because there is no content wave.
  • Sampled large accounts: --sample N.

The craft wave for a post starts after its content coding exists, except in content mode.

One background general-purpose agent per batch file:

Read SKILL/instructions/code_content.md (or code_craft.md) and follow it exactly. Batch file: <path>. Account brief: <audit>/analysis/account_brief.md. Taxonomy: <audit>/analysis/taxonomy.json. Account: <audit>/account.json. Project folder: <project_root>. Python: <python>. Report language: <lang>.

Checking the waves:

  • Count the output files against the batch.
  • --reassign hands out again posts that were assigned but never finished.
  • If an agent's output is blocked by content filtering, a transcript holds lyrics: relaunch that batch with a reminder of the lyrics rule.

7. Numbers (scripts only)

PY SKILL/scripts/build_dataset.py <u>
PY SKILL/scripts/genuine.py <u>          # not in content mode
PY SKILL/scripts/script_metrics.py <u>
PY SKILL/scripts/craft_facts.py <u>      # not in engagement mode
PY SKILL/scripts/images.py <u>           # exhibits named by the craft reviewers
PY SKILL/scripts/appendix.py <u>         # also writes post_index.md

Rerun build_dataset.py after genuine.py only if you changed account.json. Read facts.md, genuine_facts.md, script_facts.md and craft_facts.md yourself before briefing the writers, so you can catch anything odd.

8. Expert panel, then writers (agents, in parallel where they can be)

Panel (full, content and brief): one background agent per entry of taxonomy.json "panel":

Read SKILL/instructions/panel.md and follow it exactly. Your key: <key>. Role: <expert>. Dimensions: <dimensions>. Focus: <focus>. Account folder: <audit>. Project folder and Python: <...>. Report language: <lang>. Write <audit>/analysis/panel/<key>.md.

  • brief mode: ask for half the length.
  • Placing the verdicts: copy each panel file into sections/ as 07_<key>.md to 10_<key>.md, in the order the taxonomy lists them.

Writers: one background agent per role in instructions/write_report.md that the mode needs. The synthesis role goes last, after all the others have finished.

Read SKILL/instructions/write_report.md and follow it exactly. Role: <role>. Mode: <mode>. Account folder: <audit>. Report language: <lang>. Write into <audit>/analysis/sections/.

9. Build, check, fix, deliver

PY SKILL/scripts/build_report.py <u>

The build must print CHECK: clean. Otherwise fix the section files and rebuild. The check covers:

  • a heading left at a page bottom;
  • dashes or arrows;
  • a post cited without its date;
  • an ID that is not in the dataset;
  • missing charts or images.

Tables and paragraphs never split across pages. The stylesheet already prevents it.

Then read it yourself:

  • the cover, the contents page and the summary;
  • one engagement page and one craft page with images;
  • one rewrites page and the playbook.

Check that every number you see is in a facts file, and that names are spelled as the user gave them.

Deliver:

  • Send report.pdf with SendUserFile.
  • Then, in the user's language, give:
    • the page count;
    • the five or six most specific findings;
    • the wall-clock time of each stage;
    • the subagent tokens by wave, from the task notifications, against the estimate you gave in stage 2.

If the numbers differ a lot, correct cost_per_post in config.json so the next estimate is closer.

Efficiency rules

  • Scripts count; agents judge. No agent computes a number a script has.
  • Only two waves watch the posts. The panel and the writers read text and the cut exhibits, not videos.
  • Stages overlap. Collecting the other layers, the pipeline, faces and both waves run together. Never wait for the whole pipeline before starting a wave.
  • Batches of about 35 weight units, at most 10 agents per wave.
  • One download per post: the pipeline's cache serves every later look.
  • Normal-sized accounts (like a company page with a few hundred posts) need none of the large-account handling: normal pace, all posts reviewed.

Updating an earlier audit

Run collect again with a new --since. The pipeline, the faces and the waves skip what is done.

For a comparison with an earlier audit: keep the old analysis/ folder under a dated name. Then add a section that compares the old facts files with the new ones.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.