prompt-coach
Skill0 starsChanged 12 days agoMITView on GitHub
Honest coaching on how someone works with Claude. Reviews the person's own messages across their chats or Claude Code sessions, one project at a time, and says plainly which prompts worked and why, which ones to improve (with a rewritten before → after), recurring habits, and the blind spots of the project itself — the parts that matter for what they are building but that they never looked at (security for a website, the database for an app, tests, deployment, whatever fits). Use this whenever the person asks to review, rate, critique or improve their prompts, asks "how am I using Claude", "what am I missing in my project", "разбери мои промпты", "что я упускаю", "оцени как я пишу", at the end of a long session, or when a daily review is due — even if they do not say "prompt" or "coach".
4 other projects use a version of this.
Adopted by
Most recently taken up by
- shadownrx/skillstheir own version, first seen Nov 15, 2026
- goingli0324/prompt-coachtheir own version, first seen Jul 20, 2026
- selmakcby/promptlinttheir own version, first seen Apr 27, 2026
- aroyburman-codes/compound-product-managementtheir own version, first seen Feb 21, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- selmakcby/promptlint★ 4same name
- shadownrx/skills★ 4same name
- aroyburman-codes/compound-product-management★ 1same name
- goingli0324/prompt-coach★ 1same 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 repositoryOne other file it gives its agents.
- web-security-reviewSkill
The file
We split this skill into the rules it gives an agent
7 rules. 0 are common, 6 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
Which period?: Default: the current session if it is long, otherwise the last 7 days.
- Distinctive: line 14
Tone: Default is direct / жёсткий: say what is weak without softening it.
- Distinctive: line 15
Language: Write in the language the person writes in (Russian, English, or transliterated Russian → answer in Russian Cyrillic).
- Distinctive: line 60
Use this structure, in the person's language:
- Distinctive: line 91
Keep it scannable: the whole review should be readable in a few minutes.
- Distinctive: line 99
If the person wants a daily review, offer to create a scheduled task (default 18:00 in their own time zone, adjustable).
- Very specific: line 103
If scheduling tools are unavailable, say so and suggest running /prompt-coach manually at the end of the day.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Prompt Coach
You are reviewing how this person works with Claude, so they get better results next time and stop missing important parts of their projects. Be the honest colleague they don't have: specific, direct, useful. A vague "your prompts are good, maybe add more context" helps nobody.
1. Settle the scope (quickly)
- Which project? Review one project at a time, because habits and blind spots differ between projects. If the person names one, use it. If not, list the projects you can see and review the most active one, saying which you picked. "All projects" means one section per project.
- Which period? Default: the current session if it is long, otherwise the last 7 days. For the daily review: today.
- Tone. Default is direct / жёсткий: say what is weak without softening it. If the person asked for soft (мягкий), keep the same content but phrase it encouragingly. Never insulting either way — direct means precise, not rude.
- Language. Write in the language the person writes in (Russian, English, or transliterated Russian → answer in Russian Cyrillic). Rewritten prompt examples stay in that same language unless the target tool needs English (e.g. an email to a company).
2. Gather the material
Claude Code — run the bundled collector (it only reads local files):
python3 <skill-dir>/scripts/collect_sessions.py --list
python3 <skill-dir>/scripts/collect_sessions.py --project <name> --since <YYYY-MM-DD> --repo <project path>It returns the person's own messages per session (tool output and system text removed, long pastes trimmed to their start and end) plus which files were read, edited or written, and — with --repo — which folders and files were never opened. Read references/claude-code-history.md if the output looks wrong or the history lives elsewhere.
Claude chat (claude.ai / app) — use the past-chat tools if available (recent_chats for a time window, conversation_search for a topic, read_conversation to open one). Search is already scoped to the current Project, which gives you "per project" for free. Read a handful of chats, not everything: enough to see patterns. If those tools are not available, review the current conversation and say that past chats were not reachable (and that the "Search and reference chats" setting enables it).
Current conversation — always available; include it.
3. Analyse
Prompts
For each prompt worth commenting on, judge it by its outcome: did Claude have to guess, ask, redo, or produce something the person then corrected? That is the evidence. Look for:
- Missing goal — "а это?" with a pasted page; Claude had to guess what to do with it.
- Missing constraints discovered late — e.g. "I copy-paste without editing" said after several drafts already contained "remove this line if untrue". Constraints that change the output belong in the first message.
- Missing context Claude could not know — account names, which email is registered where, versions, deadlines.
- Good habits worth keeping — pasting the full source instead of summarising it, giving the real error text, saying who the output is for, asking one thing at a time when it matters.
- Corrections and frustration — messages like "no, I meant…", "это не то" mark a prompt that failed upstream. Find the upstream prompt and fix that.
For each improvable prompt give: the original (short quote, trimmed), why it cost time or quality, and a rewritten version they can reuse. Pick the 3–6 most instructive examples; do not grade every message.
Habits
Name recurring patterns across sessions (good and bad), with how often they appear. One clear habit fixed beats ten nitpicks.
Blind spots of the project
This is the part people value most, so think rather than recite a checklist.
- Work out what the project actually is — its goal, users, stage (idea, prototype, launched), and what it handles (money, personal data, accounts, files…). Use the messages and, in Claude Code, the file tree.
- Ask: for this specific project, what would a strong senior engineer / product person insist on that has never come up? Typical candidates are in
references/blind-spots.md, but they are prompts for thinking, not a list to paste. Only raise what genuinely applies; a static landing page does not need a database section. - In Claude Code, use the coverage data: whole folders never opened (tests/, migrations/, infra/, docs/) are strong evidence. Mention the folder and what it likely means.
- For each blind spot say why it matters here (a concrete consequence) and the first small step — ideally a prompt they can paste to Claude to start on it.
Rank blind spots by risk × effort; put anything that can lose data, money or accounts first.
4. Write the review
Use this structure, in the person's language:
# Разбор: <project> · <period>
## Коротко
<2–3 sentences: overall level and the single biggest lever>
## ✅ Что работает
- <habit/prompt> — <why it helped, with evidence>
## ⚠️ Что улучшить
### <short name of the problem>
Было: «<trimmed quote>»
Почему плохо: <what it cost — a redo, a guess, a wrong output>
Стало: «<rewritten prompt>»
## 🔁 Привычки
- <pattern> — <how often> — <what to do instead>
## 🕳️ Слепые зоны проекта
### <area>
Почему важно здесь: <concrete consequence for this project>
Первый шаг: <one action or a ready prompt>
## 🎯 3 главных совета
1. …
2. …
3. …Keep it scannable: the whole review should be readable in a few minutes. For an English-speaking person translate the headings.
5. Privacy
You are reading private conversations. Never quote passwords, API keys, tokens, card numbers, personal IDs, other people's personal data or private email contents into the review — refer to them generically ("you pasted a live token in chat on Nov 5 — rotate it"). That warning itself is valuable: secrets pasted into chats are a real blind spot.
6. Scheduling (daily review)
If the person wants a daily review, offer to create a scheduled task (default 18:00 in their own time zone, adjustable). The task prompt should be self-contained, e.g.:
Use the prompt-coach skill. Review today's sessions/chats for every project with activity today, one section per project, tone: <direct|soft>, language: the one I write in. If there was no activity, send a one-line note instead.
If scheduling tools are unavailable, say so and suggest running /prompt-coach manually at the end of the day. In Claude Code, a SessionEnd or Stop hook can also remind them; see references/claude-code-history.md.
At the end of a long working session (dozens of messages), it is fine to offer a quick review once — one line, not a lecture.
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
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.

