- Home
- Skills
- Prompt & Skill Engineering
- Content Forge
Content Forge
Short-form video, social post and long-form article skeletons, with headline formulas, section structure, hard writing rules and a five-point pre-publish check. Runs offline; no model call, no invented numbers.
Free
Works with the AI Agents you already use
Content Forge
Example session with this skill installed
I post about industrial maintenance for a living -- short videos mostly, some carousels. The
problem is not ideas, it is that my drafts come out sounding like everyone else's. I open with
"in today's fast-paced world" without noticing, and my headlines could be about anything. I use an
AI to draft but it produces the same flat shape every time unless I hold its hand through the
structure.
- Read your context and instructions
- Compiled the content forge
The structure is the part you are holding its hand through, so let the pack hold it instead:
python compose.py --tab video \
--topic "Using AI to cut nightly reporting time" \
--aud "maintenance technicians" --aim "subscribe"
It returns three headline candidates, each betting on a different motive -- mistakes, a threshold,
an identity -- rather than three rewordings of one idea:
## 1. Headline candidates (3 versions, each betting on a different motive)
1. Using AI to cut nightly reporting time: three mistakes I made, the second one was expensive
2. Before you start on Using AI to cut nightly reporting time, settle this one thing
3. The three things maintenance technicians get wrong about Using AI to cut nightly reporting time
Then the parameters it locked from your arguments, so the model cannot quietly change the audience
or the goal:
## 2. Writing parameters
- Topic: Using AI to cut nightly reporting time
- Audience: maintenance technicians
- What this content should make them do: subscribe
- Form: Short-form video script (30-60 seconds spoken, one camera, captions burned in)
Then the beat skeleton (hook 0-3s through call to action in the last 5s), the four hard rules, the
deliverable shape, and the five-point check as the last block -- so the draft comes back already
failing or passing something explicit, instead of waiting for you to notice it opened with a
greeting.
Connects securely to your tools. The creator never sees your data.
About this skill
Content does not fail because the writing is bad. It fails twice, in the same two places, almost every time: the headline carries no specific number and no identity label, so nobody clicks; and the first three seconds are throat-clearing, so the people who did click leave. Both are fixable with a rule, which is cheaper than fixing them with talent.
This pack writes those two as rules and puts a fixed five-point check at the end of every output. Three tabs, each with six headline formulas, a section skeleton, four hard rules and a deliverable shape.
THE THREE TABS
video Short-form video script -- 30-60 seconds spoken, one camera, captions burned in. Hook 0-3s, tension 3-10s, evidence 10-40s, turn 40-50s, call to action in the last 5s. xhs Social post (Xiaohongshu / carousel) -- headline in three versions, first two lines (they decide whether the post expands), bulleted body with real numbers, where people get it wrong, tags. Deliverable adds a cover-text suggestion under 10 words. wechat Long-form article (newsletter / blog) -- five headline versions, standfirst under 40 words, opening hook, 4-6 sections, evidence to gather, closing hook.
THE HARD RULES ARE THE PRODUCT
Four per tab, non-negotiable, stated to the model as rules rather than advice. The video tab's:
- A concrete number or a counter-intuitive claim must land in the first three seconds. No greeting.
- Keep spoken lines under 12 words so they fit one breath.
- Every 8 seconds, give a visual cue: shot, caption, or overlay.
- Never open with "Hey guys" or "Today I want to share".
Rule 1 is the one that changes the draft. "No greeting" is not style -- a greeting is the single most reliable way to lose the first three seconds, and it is also the first thing a model writes when nobody stops it.
HEADLINE FORMULAS HAVE TWO BRANCHES, ON PURPOSE
Each tab ships six formulas: three for a short topic and three for a long one. The branch is chosen by topic length, and the long branch drops the added qualifier rather than appending it. A long topic plus a long suffix is how you get a three-line headline that gets truncated on a phone -- the formula that reads best on a desktop performs worst where it is actually seen.
The audience name is substituted verbatim into the formulas, which is why the argument help says to
pass a plural or a mass noun: "the five traps maintenance technicians hit" is a sentence, "the five
traps maintenance technician hit" is a translation. The tool does not guess plurals -- English
plural rules would get staff, equipment and lead wrong, and a wrong plural in a headline is
worse than a rule in the help text.
THE FIVE-POINT PRE-PUBLISH CHECK
Fixed, appended to every output, titled "one miss means do not publish":
- Would the first two lines stop someone? Read them alone and ask whether you would scroll past.
- Is there one specific number or scene? If not, add a real one.
- Is any sentence speaking for the reader rather than about yourself?
- Does it promise something you cannot deliver? Delete it -- that matters more than writing well.
- Is the last line one clear action? Subscribe or message, not both.
WHAT IT DOES NOT DO
It does not write the content. It writes the prompt that makes a model write it to a skeleton. It does not generate images, hashtags, thumbnails or posting schedules. (Tags appear as a section to fill, not as generated hashtags.) It does not measure performance. No analytics, no network, no account access. It does not know your product. What it cannot know becomes a slot in the skeleton, marked as a slot, so the model asks instead of inventing a case study.
REQUIREMENTS
Python 3.8 or newer, standard library only. One file, compose.py. Prints to stdout and writes
nothing. python compose.py --list prints the three tabs.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Prompt & Skill Engineering.
Generates prompts, not final content. Requires manual pluralization of audience inputs for correct headline grammar. Runs strictly via CLI with no GUI.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 21 days ago
- Free with an Agensi account
Permissions
Terminal -- usage, not behaviour. This is a script skill: the operator runs `python compose.py --tab video --topic "..." --aud "..." --aim "..."` in their own shell and reads the prompt from stdout. Nothing in the package starts a process, opens a shell, evaluates a string or downloads anything; `compose.py` imports argparse, json and sys and nothing else. Read Files -- off, and correctly off. `compose.py` opens no path at all: the headline formulas, the skeletons and the checklists are string constants inside the file, all four inputs arrive through argv, there is no path literal and `os` is not imported. The word sits inside the product text: one headline formula in this pack is "{topic}: read these three lines before you start", and the format rules ask for "five headlines, each labelled with the reader it suits" -- English inside the generated prompt, not a call. Verify in one line: `grep -n "open(" compose.py` returns nothing. The submit form's own auto-detect produced Terminal and nothing else on this ZIP, which is the same reading. Write Files -- off. The only output is stdout -- no `open()`, no temp file, no draft file, no cache. Nothing is saved between runs. Browser -- off. Nothing opens a page, drives a browser, posts anything or imports a browser library. Network Access -- off. No socket, no HTTP client, no DNS, no telemetry, no update check. The u.org.cn URL in the prose is text for a human reader; nothing requests it. Environment Variables -- off. No `os.environ`, no `getenv`, no `process.env`; `os` is not imported.