Technical Writing Blog
الذهاب إلى القناة على Telegram
ClickHelp — online documentation tool for technical writers and teams. Create, translate, and publish documentation easily in one portal! Get your FREE trial 👉 https://goo.gl/NahsT2
إظهار المزيد1 254
المشتركون
لا توجد بيانات24 ساعات
+27 أيام
-730 أيام
جارٍ تحميل البيانات
القنوات المماثلة
سحابة العلامات
الإشارات الواردة والصادرة
---
---
---
---
---
---
جذب المشتركين
نوفمبر '26نوفمبر '26
نوفمبر '260
في 0 قنوات
أكتوبر '26
+14
في 0 قنوات
Get PRO
سبتمبر '26
+9
في 0 قنوات
Get PRO
أغسطس '26
+10
في 0 قنوات
Get PRO
يوليو '26
+5
في 0 قنوات
Get PRO
يونيو '26
+8
في 0 قنوات
Get PRO
مايو '26
+6
في 0 قنوات
Get PRO
أبريل '26
+14
في 0 قنوات
Get PRO
مارس '26
+11
في 0 قنوات
Get PRO
فبراير '26
+19
في 0 قنوات
Get PRO
يناير '26
+22
في 0 قنوات
Get PRO
ديسمبر '25
+18
في 0 قنوات
Get PRO
نوفمبر '25
+19
في 0 قنوات
Get PRO
أكتوبر '25
+17
في 0 قنوات
Get PRO
سبتمبر '25
+20
في 0 قنوات
Get PRO
أغسطس '25
+26
في 0 قنوات
Get PRO
يوليو '25
+40
في 0 قنوات
Get PRO
يونيو '25
+17
في 0 قنوات
Get PRO
مايو '25
+15
في 0 قنوات
Get PRO
أبريل '25
+16
في 0 قنوات
Get PRO
مارس '25
+20
في 0 قنوات
Get PRO
فبراير '25
+22
في 0 قنوات
Get PRO
يناير '25
+29
في 0 قنوات
Get PRO
ديسمبر '24
+23
في 0 قنوات
Get PRO
نوفمبر '24
+30
في 0 قنوات
Get PRO
أكتوبر '24
+21
في 0 قنوات
Get PRO
سبتمبر '24
+28
في 0 قنوات
Get PRO
أغسطس '24
+35
في 0 قنوات
Get PRO
يوليو '24
+30
في 0 قنوات
Get PRO
يونيو '24
+22
في 0 قنوات
Get PRO
مايو '24
+32
في 0 قنوات
Get PRO
أبريل '24
+21
في 0 قنوات
Get PRO
مارس '24
+35
في 0 قنوات
Get PRO
فبراير '24
+23
في 0 قنوات
Get PRO
يناير '24
+40
في 0 قنوات
Get PRO
ديسمبر '23
+36
في 0 قنوات
Get PRO
نوفمبر '23
+35
في 0 قنوات
Get PRO
أكتوبر '23
+51
في 0 قنوات
Get PRO
سبتمبر '23
+29
في 0 قنوات
Get PRO
أغسطس '23
+43
في 0 قنوات
Get PRO
يوليو '23
+35
في 0 قنوات
Get PRO
يونيو '23
+31
في 0 قنوات
Get PRO
مايو '23
+52
في 0 قنوات
Get PRO
أبريل '23
+28
في 0 قنوات
Get PRO
مارس '23
+40
في 0 قنوات
Get PRO
فبراير '23
+39
في 0 قنوات
Get PRO
يناير '23
+37
في 0 قنوات
Get PRO
ديسمبر '22
+41
في 0 قنوات
Get PRO
نوفمبر '22
+45
في 0 قنوات
Get PRO
أكتوبر '22
+35
في 0 قنوات
Get PRO
سبتمبر '22
+44
في 0 قنوات
Get PRO
أغسطس '22
+45
في 0 قنوات
Get PRO
يوليو '22
+43
في 0 قنوات
Get PRO
يونيو '22
+46
في 0 قنوات
Get PRO
مايو '22
+35
في 0 قنوات
Get PRO
أبريل '22
+56
في 0 قنوات
Get PRO
مارس '22
+47
في 0 قنوات
Get PRO
فبراير '22
+42
في 0 قنوات
Get PRO
يناير '22
+41
في 0 قنوات
Get PRO
ديسمبر '21
+32
في 0 قنوات
Get PRO
نوفمبر '21
+49
في 0 قنوات
Get PRO
أكتوبر '21
+42
في 0 قنوات
Get PRO
سبتمبر '21
+35
في 0 قنوات
Get PRO
أغسطس '21
+49
في 0 قنوات
Get PRO
يوليو '21
+57
في 0 قنوات
Get PRO
يونيو '21
+44
في 0 قنوات
Get PRO
مايو '21
+44
في 0 قنوات
Get PRO
أبريل '21
+41
في 0 قنوات
Get PRO
مارس '21
+49
في 0 قنوات
Get PRO
فبراير '21
+75
في 0 قنوات
Get PRO
يناير '21
+59
في 0 قنوات
Get PRO
ديسمبر '20
+873
في 0 قنوات
| التاريخ | نمو المشتركين | الإشارات | القنوات | |
| 02 نوفمبر | 0 | |||
| 01 نوفمبر | 0 |
منشورات القناة
Many teams that “use AI for docs” spend part of their day copying text from a chat window into the portal. At that point the writer is the one assisting the AI.
In our article, this is L1 on a five-level scale of AI autonomy in documentation. The scale starts at L0, where AI only suggests phrasing as you type, and ends at L4, where it monitors the product and publishes updates without approval. Almost no one is at L4 today.
The middle of the scale is where it gets harder. Moving from L1 to L2 is an integration task: the AI gets into the portal and sees the structure, the existing topics, and the workflow statuses. Moving from L2 to L3 is a governance task. Someone has to decide which content can update without a human, what counts as a safe change, and how to roll back a bad one.
Release notes can run at L3 while compliance docs stay at L1. That’s a deliberate setup.
Read the full breakdown of all five levels
| 2 |  Your docs chatbot gives a wrong answer, and the usual fix is a bigger model. We tested that. It didn’t help.
We ran 50 questions our users had actually asked through 14 configurations, all on the same model.
Often the model simply never got the article with the answer. No model, however big, can answer from an article it never saw. A reranker didn’t help either: no improvement in four comparisons.
What did help was the work around the model: how the docs are prepared for search, how much context the model gets, and what it does when it isn’t sure.
Read the full test and a five-step check for your own docs | 112 |
| 3 |  🤖 Your AI support assistant has a 78% deflection rate. Is that a success? Not necessarily.
Some users get the answers they need. Others receive an unhelpful response, give up, and open a support ticket the next day.
On your dashboard, both scenarios can look exactly the same.
That’s why deflection rate alone isn’t enough to evaluate an AI documentation system.
📊 Here are 5 metrics that really matter:
✅ Accuracy
📚 Groundedness
🔗 Citation quality
💬 True deflection
💰 Cost per accurate answer
The last point is especially important: a cheap AI response can become expensive if it creates more work for your support team.
The goal isn’t to make conversations end. It’s to help users find reliable answers.
We’ve put together a practical guide to evaluating AI documentation systems, measuring what matters, and looking beyond impressive dashboard numbers.
📖 Read the full article | 115 |
| 4 |  Imagine reading a single page torn out of the middle of a manual. The instructions make sense, but you can’t tell which product they’re for, which version, or who they were written for.
An AI assistant gets your documentation in the same state. Before indexing, every article is cut into fragments of a few hundred tokens, and each fragment is searched on its own. The “Enterprise only” line stays at the top of the article. The steps end up in another chunk.
So the useful question isn’t “is the text correct?” It’s “does every fragment know where it came from?” Product, version, audience, task type and environment have to live in structured fields and be copied to every chunk, not mentioned once in the text.
We wrote about which fields to start with, where to keep them, and what to do when the system doesn’t know who’s asking.
Read the article on AI-ready metadata | 191 |
| 5 |  AI makes writing documentation faster. It doesn’t make it more consistent. Ask an assistant to explain the same step in five topics, and you get five slightly different versions of one instruction.
Single-sourcing fixes that: a fragment lives in one place and is reused everywhere. The setup used to be the hard part, because someone had to find every copy by hand.
On October 15 we’ll hand that job to AI. An assistant connected to ClickHelp through MCP will go through a project with repeated content, turn those passages into snippets and replace the copies in each topic. We’ll also show how variables keep product versions in check.
October 15, 18:00 CEST / 12:00 PM EDT
Register for the webinar | 193 |
| 6 |  A new product edition usually starts with a copied manual. It's fast, and for one copy it works fine.
The trouble comes later. The support address changes, and now it has to be found in four manuals and two languages. One copy gets missed, and a user calls the wrong number.
Single-sourcing keeps one content base and builds every manual from it. Shared procedures are written once as reusable topics. Product names, versions, and contacts go into variables. Enterprise-only sections are handled with conditional content, and warnings or legal notices live in snippets.
The approach has limits. Too many conditions make a topic hard to read, and if two products share less than 50–60% of their content, separate projects may be simpler.
Read the article | 252 |
| 7 |  A company ships its second product, and the docs team starts arguing: one site for everything, or a separate site for each product?
Both camps have a point. Split the docs, and legal pages, security policies, and shared how-tos get maintained twice. Merge them, and readers of one product keep landing on another product's pages.
The argument goes in circles because layout alone guarantees nothing. A more useful test: do readers always know where they are, and can the team avoid maintaining the same content twice?
Seen that way, "separate for readers" and "shared for the team" are two independent settings, and most of the trade-off disappears.
In the article, we look at four setups teams usually start from, three levels where separation is actually required, and why reuse should be in place before the second documentation site.
Read the article | 258 |
| 8 |  Your documentation calls it "SSO configuration." Your reader types "can't log in with my work account."
It's the same problem. But keyword search sees two unrelated phrases, and the reader leaves without an answer.
In the September release, AnswerGenius started searching by meaning as well as by keywords. When readers describe a problem in their own words, AnswerGenius still finds the right topic and answers from it. The answers themselves are now written by a newer AI model.
Also in this release: Documentation Sites on a ClickHelp subdomain with nothing to set up on your side, Restricted publications in the Published Docs MCP Server, and five features out of beta.
Read the September Release Notes → | 209 |
| 9 |  Most documentation is written for people the author never talks to. Nobody interviews the admin who opens a release note at 2 a.m. or the developer skimming an API reference between meetings.
That doesn't mean the reader is unknown. Product type tells you who uses it. The search query tells you how urgent the problem is: "resolve 503 error" and an onboarding portal visit call for very different pages. Role terminology and compliance requirements fill in the rest.
The hard part is keeping the profile honest. A persona built on one signal, usually product type, drifts from reality. So does one written for "Jenkins-familiar DevOps engineers" after the team moves to GitLab CI/CD.
The article lays out a method: a one-sentence reader profile, a five-stage knowledge model, and a 10-item checklist before publishing.
Read the full article | 244 |
| 10 |  Documentation generation with AI takes under a minute. It reads naturally, matches the tone of your docs — and is wrong in three places, because the ticket behind it was overridden by a decision no one wrote down.
That's not an AI problem. It's a timing problem. The decisions that shape documentation (terminology, defaults, edge cases) get made during requirements and design. Documentation usually gets attached to testing and release, long after those decisions are already settled.
AI removed the friction of writing. It did nothing about access to reliable information. If accurate info only exists after release, AI just produces plausible-sounding misinformation faster — and support bots reproduce that error at scale.
We laid out a practical model: where writers need early access to specs, and where it isn't worth the disruption.
Read the full article | 225 |
| 11 |  Release notes usually get written last, under time pressure, by whoever's left in the room. A dev tries to remember what shipped. A PM fills gaps from memory. Someone turns "fix: null pointer in auth" into something a user can read (five minutes before the release goes out.)
None of that needs to come from memory. It's already in your commits, PRs, and issue tracker. A commit like fix(auth)!: handle expired tokens tells you exactly what changed and whether it's breaking. Tools like semantic-release can turn a history of these into a changelog automatically.
What automation still can't do: decide if a "feat" commit is something users should hear about, or whether five performance commits are really one story worth telling.
👉 Full story here | 195 |
| 12 |  Two dictionaries just picked "slop" as their word of the year. In documentation it looks different: a portal that doubled in size over a year while support tickets stayed flat.
New on our blog: why cheap writing killed scoping, what happened at Snowflake after they cut ~70 tech writing roles right after a 30% revenue year, and the 6 questions nobody's left to ask anymore.
Full piece here | 205 |
| 13 |  Ask machine translation to handle the same product term twice, and there's no guarantee it comes back the same way. Multiply that across a documentation set, and one button ends up with three different names in three different languages.
We sat down with Alconost to talk through where this actually comes from — and why the fix isn't a smarter MT engine. It's a maintained termbase with an owner, real review, and a process that keeps up when the product changes.
Real incidents included: a core product term that came back as a compound word that doesn't exist in the target language, and a single bad string that spread through a translation memory badly enough to break a client's build.
👉 Full story | 210 |
| 14 |  Wrote up a comparison of 10 docs-as-code tools: free and commercial.
The short version: MkDocs if you want something running in an afternoon, Docusaurus if you need versioning, Antora if your docs are spread across multiple repos. For managed hosting, Read the Docs is the most mature option. Swimm is in a different category — it's less about publishing and more about keeping docs accurate as code changes.
Full breakdown is here | 189 |
| 15 |  New in ClickHelp this August 👇
Readers can now plug their own AI agent (Claude, Cursor, ChatGPT, Gemini) straight into your published docs — the Published Docs MCP Server reads your current content and answers only from what's actually there, not a guess or an outdated training snapshot.
Search also got smarter: AI Overview answers questions right in the results, and you can keep asking follow-ups without leaving search.
Smaller but useful: Watch + version history are now in the Next Reader UI, and Quick Publish shows you the full list of topics before you hit publish.
✍🏼 Full Release Notes | 168 |
| 16 |  Another one on content reuse — this time about snippet management.
Short version: if the same warning or install step is copy-pasted across 20 topics, and every product change means editing all 20 by hand — that's the sign it should be a snippet instead.
We go through the difference between variables (values: version number, email, company name) and snippets (blocks: procedures, warnings, legal text), what's actually worth turning into a snippet vs. what should stay inline, and how to keep a snippet library from turning into a pile of note1, temp_block, and snippet_new.
👉 Read the full article | 154 |
| 17 |  Technical documentation has a text problem — not too little text, but text doing jobs it's not good at.
Describing a multi-tab configuration wizard in prose takes 200 words and still leaves users guessing what the screen looks like. A 90-second screencast answers the same question and gets out of the way.
The article breaks down when each format earns its place:
- Screencasts work when the workflow is short, stable, and UI-heavy.
- Automated capture (Scribe, Tango, iorad) wins when the product changes often ( updating one step is faster than re-editing a full video.)
- Full video makes sense for onboarding and training where production quality matters.
The part worth noting: video without a transcript is a dead end for search. Users type questions into search boxes, not video players. Text stays the primary layer — multimedia just fills in what text can't show.
👉 Read the full article | 166 |
| 18 |  Claude writes well. But without access to your actual data, it's writing in a vacuum.
MCP fixes that. One server, and any compatible client (Claude, Cursor, Windsurf) can reach your tools and data without a custom connector for each one.
We wrote a step-by-step guide to building a local MCP server in Python:
— project setup with uv
— a working tool with @mcp.tool() in 30 lines
— local testing with MCP Inspector before touching Claude Desktop
— the claude_desktop_config.json file and where to find it
— the most common failures: wrong path, missing uv, vague docstrings
For technical writers: the guide ends with a pointer to ClickHelp's MCP server — which gives AI agents direct read/write access to your documentation portal.
👉 Read the full guide here | 178 |
| 19 |  Docs aren't just read by people anymore. So we tested how AI handles them.
We ran a quick test: gave ChatGPT a link to one of our documentation pages and asked it to summarize what it was about.
It opened the link, read the page, and answered in seconds. No extra steps.
Small result, but a good reminder — docs today get opened by more than just people.
Search engines, AI assistants, link previews in Slack and LinkedIn all try to do the same basic thing: open a link, get something useful back.
Read the full story | 158 |
| 20 |  DITA is powerful. But does your team actually need it?
Topics, maps, reuse, single-source publishing — it all works. Just comes with an XML editor, a publishing engine, a CCMS, and months of setup.
For a 3-person team with one product: almost always overkill. For 20+ authors, five languages, and regulatory requirements: almost always worth it.
Between "Word doc chaos" and "full enterprise CCMS" there's a whole spectrum of tools that don't get talked about enough.
We broke it all down — structured authoring explained, tools compared (DITA OT, Paligo, Heretto), and a practical guide to picking the right setup for your team size and workload.
👉 No fluff | 177 |
