Most IELTS practice fails in a boring way.
The learner writes an essay, gets a correction, nods, and moves on. Two days later they make the same grammar mistake again. A week later they still have the same weak Task 2 paragraphs. Speaking practice has the same problem: a mock answer gets feedback, but the pattern does not stick.
That is not a model problem. It is a memory problem.
For IELTS, the useful Agent is not the one that gives the longest essay correction. It is the one that keeps a mistake log and turns feedback into the next drill.
The loop I want
The Skill I use for this is the open IELTS Writing & Speaking Coach Skill. The direct file is SKILL.md.
The core loop is simple:
prompt → timed attempt → rubric feedback → error extraction → micro-drill → rewrite/retry → spaced review
That loop matters more than any single correction.
If the learner keeps missing articles, the next session should not pretend this is new. If their Task 2 body paragraphs are underdeveloped, the Agent should keep pushing paragraph development until it improves. If they speak fluently but avoid complex grammar, the next Part 3 practice should target that.
Random feedback feels helpful. A tracked loop is more useful.
What the Agent should ask first
Before reviewing anything, the Agent needs a small profile:
Academic or General Training?
Target band?
Current estimated band?
Test date or target window?
Weakest area: Task 1, Task 2, Speaking, grammar, vocabulary, fluency?
Weekly study time?
Preferred feedback style: strict, supportive, fast correction, detailed explanation?
This keeps the feedback grounded.
A candidate targeting Band 6.5 next month does not need the same coaching as someone trying to move from 7.0 to 7.5 in Writing. The first may need task response and sentence control. The second may need precision, cohesion, and fewer repeated errors.
Writing feedback should be criterion-level
For Writing, the Agent should not say "good essay" or "needs improvement" and stop there.
It should review against four criteria:
Task Achievement / Task Response
Coherence and Cohesion
Lexical Resource
Grammatical Range and Accuracy
For Task 1, it should check the overview, key features, comparisons, data accuracy, tone, and word count.
For Task 2, it should check whether the position is clear, whether all parts of the question are answered, whether paragraphs are developed, whether examples are specific, and whether cohesion is doing real work instead of just adding linking words.
The useful output is not a wall of comments. I prefer this shape:
Estimated criterion feedback
Biggest score limiter
Top recurring errors
Rewrite one paragraph
One drill before the next essay
The "biggest score limiter" is the part learners often need most. Ten corrections are overwhelming. One priority is actionable.
Speaking practice should not become memorized scripts
Speaking is trickier because bad AI coaching can push users toward memorized answers.
That is not the goal.
The Skill has a boundary against scripted speaking answers. The Agent should help the learner build flexible patterns, not fixed speeches.
For Part 1, direct and natural answers are enough.
For Part 2, the Agent can help the learner organize a two-minute answer from keyword notes.
For Part 3, it should push abstract reasoning:
comparison
cause and effect
advantages and disadvantages
speculation
examples
qualification
The speaking memory should track patterns like:
pauses too often after complex questions
answers Part 3 too briefly
uses simple grammar under pressure
repeats the same linking phrases
mispronounces a small set of frequent words
Again, the point is not one mock test. The point is noticing what repeats.
The mistake log is the product
If I had to keep only one feature, I would keep the mistake log.
A useful log is not just a list of corrections. It should connect error, example, fix, and review date.
Error type: article use
Example: "government should provide job for citizens"
Correction: "the government should provide jobs for citizens"
Pattern: missing article before singular countable/common institution nouns
Drill: write 10 sentences about public services using "the government", "a policy", "an employer"
Review: in 3 days
Now the Agent has something to do next time.
It can say: before we write another full Task 2 essay, do a five-minute article drill. Annoying, but useful.
What not to let the Agent do
IELTS coaching needs boundaries.
The Agent should not:
- guarantee a band score
- claim to be an official IELTS examiner
- encourage memorized essays
- write hidden scripts for speaking answers
- help with cheating or test misconduct
- punish accent instead of intelligibility and clarity
AI score estimates should be treated as estimates. Official results may differ.
That sentence is boring, but it belongs in the system. It keeps the Agent honest.
Using the Skill in your setup
If your agent client can install, import, or reference GitHub-based Skills, use the IELTS Writing & Speaking Coach Skill as the source.
For repo-based projects, keep it as:
skills/ielts-writing-speaking-coach/SKILL.md
Then wire your Agent to load it for IELTS Writing Task 1, Writing Task 2, Speaking Parts 1-3, mistake review, prompt generation, and study planning.
For Claude Code or another repo-aware coding agent, the Skill can be project context. Use it to update your agent instructions, practice templates, feedback schema, or mistake-log database.
For ChatGPT-style custom assistants, use the Skill as the source of truth if your setup supports imported instructions or external Skill references. Keep the learner profile and mistake log separate from the Skill itself.
The Skill is the coaching method. The learner history is memory.
Try the hosted Agent
If you want to use it directly instead of wiring up the Skill yourself, ClawMama has a hosted IELTS Writing & Speaking Coach Agent.
You can create the Agent and use it directly in Telegram or WhatsApp. The Skill is already attached, and the Agent can keep the learner's target band, recurring mistakes, vocabulary gaps, and practice schedule over time.
A quick test
Give your Agent this prompt:
Here is my IELTS Writing Task 2 essay. I want Band 7. Please review it and tell me what to practice next.
A weak Agent gives a long correction.
A better Agent finds the score limiter, extracts recurring errors, gives one rewrite task, and schedules the next drill.
That is the behavior worth building.
Top comments (0)