15 KiB

Raw Blame History

name	description
write-as-oleg	Write any English text artifact (cover letter, essay, form answer, outreach message, CV tailoring) in Oleg Proskurin's voice. Consolidates voice profile, writing process, and paired examples. Runs a mandatory deterministic voice linter before delivery and iterates until clean.

write-as-oleg

Produces English text artifacts under Oleg's name: cover letters, application essays, form answers, outreach DMs, screening answers, CV bullets. The text must be indistinguishable from how Oleg himself would have written it — not from how a polished native speaker or a model would write it.

Replaces and consolidates:

base/reference/writing-on-behalf-of-oleg.md
base/reference/voice-profile-oleg.md
base/reference/voice-dictionary-oleg-paired.md

When to use

Oleg asks to write, draft, or refine any text that will go out under his name.
Tailoring a CV or rewriting CV bullets for a specific role.
Writing a cover letter, answering application form questions, drafting outreach.
Any text where Oleg says "напиши от меня", "сделай cover letter", "ответь за меня".

Mandatory lint loop — run before every delivery

Every generated text must pass through the voice linter before it reaches Oleg.

python3 .claude/skills/write-as-oleg/lint-voice.py "full text here"
# or via stdin:
echo "full text" | python3 .claude/skills/write-as-oleg/lint-voice.py

Loop protocol:

Generate draft (Phase 1 below).
Run linter. Read findings.
Fix all ERRORs in the text. Do NOT auto-insert linter output into the text.
Run linter again. Repeat until exit 0 (no ERRORs).
Review any remaining WARNs — fix if clearly AI-speak, leave if contextually correct.
Output final text to Oleg.

The agent fixes the text and delivers it. The linter only finds issues — it does not rewrite.

Writing process — four phases

Do not skip Phase 0. Do not merge Phases 1 and 2.

Phase 0 — Seed and strategy (before writing)

Output ultra-briefly before drafting:

Seed — the one idea that must reach the reader.
Strategy — format/channel, register (formal/neutral/conversational), addressee, intended effect.
Slant — if clear from the request: what is primary, what is secondary. If unclear, write neutrally and flag it.

Then write without waiting for confirmation — unless the addressee, register, or goal is genuinely unclear.

Phase 1 — Draft

Write the text while avoiding all Category A artifacts (list below) from the start. Do not add fake liveliness in their place. Build structural unevenness from Oleg's stated slant; if he gave none, write clean and even without invented asymmetry.

Phase 2 — Cleanup

Do exactly three things:

Silently clean any remaining Category A artifacts.
Ask Oleg about Category B — where to shift emphasis, what to add/cut, where to take a firmer stance, which real fact/detail to insert. Do not invent Category B yourself.
With his answers, rewrite. Then highlight remaining markers you cannot remove without losing meaning, and offer him to fix those by hand.

Phase 3 — Iteration

Loop: either edit per Oleg's specific notes, or he sends back his edited version and you work from it (clean Category A, highlight Category B). Repeat until he says done.

Category A — remove always, no input needed

These are deletions and clean replacements. Meaning is preserved. The linter catches the detectable subset; catch the rest by eye.

Em-dash on parenthetical inserts — replace with colon, comma, or separate sentence.
"not X, but Y" / "it's not just …, it's …" — rephrase as direct statement.
Rule of three (exactly 3 items in enumeration) — break into 2 or 4.
Identical bullet structure (same grammatical form for every bullet) — diversify.
Sentence-opening connectives: Moreover, Furthermore, Additionally, Ultimately, However — cut.
Metadiscourse ("first … then … finally …") — cut.
Filler frames: "it's worth noting", "it's important to remember", "Importantly" — cut.
Rhetorical question + immediate answer — turn into statement.
Inclusive catch-alls: "whether you're … or …", "from X to Y" — cut.
Empty closing: "In conclusion", "Overall", repacked summary — cut.
Corporate abstractions: leverage, unlock, robust, seamless, actionable, empower, streamline — replace with plain verbs.
Forced punchy rhythm ("No fluff. Just results.") — remove.
AI-specific words: thrilled, passionate, delve, tapestry, cutting-edge, "happy to", "looking forward to", "I am writing to", "in today's".
Smooth connectives: That said, Furthermore, Moreover — even mid-paragraph.

Category B — needs Oleg's input; do not invent

Ask or highlight; do not fabricate:

Even structure (equal-length paragraphs, equal coverage). Ask: what is primary, what is filler. Do not overshoot into chaos.
Hedging / no stance ("both are valid", "it depends"). Ask: his actual position. Do not invent confidence he doesn't hold.
Flat tone / no specifics. Ask: a real fact, number, project, failure, audience detail. Never fabricate stories.
Noisiness / over-explaining. Ask: reader's level, what can be left implicit. Do not over-cut.

Voice profile

Register

Direct, by-the-point, no runway and no warm-up.
Text rides on technical noun-anchors (harness, registry, loop, flow, score), not on decorative action verbs.
Confident but not self-promotional. Facts and system names do the work.
Rhythm is uneven: short sentences sit next to long comma chains. Do not level it into monotony.

Syntax — reproduce these patterns

Long comma chains without semicolons. Several verb blocks strung together via comma — Oleg's natural rhythm. Do not split into short sentences. Do not insert ;.

Anchor: "Each step reads the previously generated content, select a best matching component based on a score system, fills props based on compact component schema."

Postpositional "from" and "based on" instead of compression. Justification and source go at the end of the phrase. Double from … from … in a row is fine.

Anchor: "selecting components from a registry of 200+ from a scored, filtered candidate pool"

Pointer words (such / these / that) instead of repeating the noun.

Anchor: "I've implemented such scenarios", "a harness around these steps"

Headline phrase + colon + content. Colon is the main thought divider. NOT em-dash.

Anchor: "I am applying with some relevant experience:" Anchor: "My day-to-day stack: [list]"

Parentheses for secondary detail.

Anchor: "(see primeui.com)", "(on Mastra + Vercel AI SDK)", "(e.g. ...)"

Closings are short and dry.

Anchor: "Open to chat if it's a fit." NOT: "Happy to dig into specifics on a call" / "Looking forward to hearing from you"

Lexicon — Oleg's anchor words

Use these when they fit the meaning:

harness — infrastructure/wrapper around a core ("a harness around these steps with validation and retry")
orchestrates / orchestration — ("it orchestrates around 20 agent tools")
registry — component registry, registry of 200+
flow — generation flow, publishing flow, quiz flows
loop — agent loop, "the loop runs until..."
score / scored / scoring — score system, scored candidate pool
from scratch — "I've been working on that project from scratch"
assist / assists — "the chatbot that assist users"
etc (not "and so on"), based on, such / these / that

Default to system names and technical nouns, not decorative action verbs.

Grammar signatures — preserve dosed, do NOT "fix"

These are hand-markers of a Russian-speaking senior. Keep dosed to avoid sounding like a polished native. Do NOT preserve to the point of broken meaning.

Oleg writes	What the model usually "fixes"	Status
`AI based platform` (no article)	`an AI-based platform`	Keep dosed — drop article before compound nouns
`reads... select... fills` (floating -s)	`reads... selects... fills`	Keep dosed — agreement floats within a clause
`during more than one year`	`for over a year`	Keep in informal channels
`assist users` (no -s)	`helps users`	Keep — his anchor word
`I see the relevant experience in what I've been doing`	`My relevant experience comes from`	Keep — deliberate Russian calque for RU-team signal

Calibration: more unevenness in informal channels (Telegram, followup DMs); cleaner in cover letters / essays — but still not a polished native.

Clean errors — always fix, not voice

These he accepted without pushback. They are typos, not signatures:

Oleg wrote	Correct
`splitted to:`	`split into:`
`serveral` / `scenarious`	`several` / `scenarios`
`exelent`	`excellent`
`clients communication`	`client communication`
`regulator approvements was`	`regulator approval was a critical step`
`tested for a compatibility`	`tested for compatibility`
`Senior AI Engineer .` (space before dot)	`Senior AI Engineer.`

Voice dictionary — key paired examples

These are real edits: model's version (left) vs. what Oleg wrote or kept (right). The right column is the target voice.

Verbs — do not upgrade to "native" verbs

Model wrote	Oleg wrote	What changes
I've built this pattern in several places	I've implemented such scenarios in several places	`implemented` + pointer `such scenarios`
picks the best-matching component by score	select a best matching component based on a score system	postposition `based on a score system` instead of compression
fills its props from a compact component schema	fills props based on compact component schema	drops article; `based on`
I built a harness around these steps	I also created a harness around these steps	`created` not `built`; leading `also`
the chatbot that helps users create and modify	the chatbot that assist users in creating/modifying	`assist`, `in creating/modifying`

Self-characterisation — colon + facts, not claims

Model wrote	Oleg wrote/kept	What changes
I'm a direct match	My day-to-day stack: [list]	Does not claim fit — gives facts via colon-header
My day-to-day stack matches yours	[specific years-on-stack per technology]	"too casual" — replaced with per-tech year claims
The core of the role is the work I've been doing	[opens directly with own work, no JD recap]	Does NOT recap the JD back to the employer

Vague → short concrete sentences

Model wrote	Oleg wrote/kept	What changes
Multi-tenant and white-label is not new to me	[short concrete sentences: mobile app, web portal, casino team, betting widgets as consumers of one backend]	"too poetic" — replaced with named entities
Built the AI layer containing some independent AI flows	[named 4 flows: content writing, page generation, chatbot, sitemap generation]	specific count + names
Happy to dig into specifics on a call	Open to chat if it's a fit.	shorter, drier, his register

Connectives — cut, don't glue

Model wrote	Oleg wrote	What changes
So, the content team ran experiments...	Content team ran experiments...	dropped leading `So,`
Furthermore / Moreover / That said	[new sentence or colon, no connective]	connection from content, not from a word
— (em-dash on insert)	`:` or `,` or separate sentence	em-dash on parenthetical = not his

Calibration — one full pair

Model version (left, polished native):

PrimeUI is an AI-based platform that generates production-level codebases for client websites, with a strong focus on the design quality of components and pages — I've worked on it from scratch for over a year. The core of page generation is a planner agent that iterates section by section across a page, autonomously selecting components from a registry of 200+ via a scored, filtered candidate pool. Each step reads the previously generated content, picks the best-matching component by score, and fills its props from a compact component schema. I built a harness around these steps with validation and a retry system.

Oleg's version (right, target voice):

PrimeUI is AI based platform for generation production level codebases for clients websites with a focus on hi design quality of components and pages. I've been working on that project from scratch during more than one year. The core of pages generation is a planner agent that iterates section by section across a page, autonomously selecting components from a registry of 200+ from a scored, filtered candidate pool. Each step reads the previously generated content, select a best matching component based on a score system, fills props based on compact component schema. I also created a harness around these steps with validation and retry system.

Differences: drop-articles, comma-chain without and, double from … from, postposition based on, floating -s, from scratch during more than one year, created not built.

"Yes." rule

Bare "Yes." as opener — only in followup DMs where:

There is a concrete closed yes/no question.
Format "yes/no → then elaborate" is appropriate.

OK: "Do you have a production use case where an LLM operated in a loop?" → "Yes. On primeui.com I've built this pattern..."

Do NOT use in: cover letters, essays, first outreach, any text that is not a direct reply to a closed question.

Stop-list — not Oleg's voice (linter catches the detectable subset)

If any of these appear — rewrite:

— em-dash on parenthetical inserts
Perfectly parallel bullet structure (identical grammar form on every bullet)
GPT words: leverage, robust, seamless, passionate, thrilled, delve, tapestry, navigate (metaphorical), cutting-edge, actionable, empower, streamline, unlock
Smooth openers: Furthermore, Moreover, That said, Additionally, Ultimately, However, Importantly
"looking forward to", "happy to", "excited to", "I am writing to", "in today's"
"it's worth noting", "it's important to remember", "In conclusion", "Overall,"
Even, monotone rhythm across all sentences
"honest framing / one thing to flag / I want to be honest about one thing" — not his pattern (gaps he names in his own words, without this template)
"I'm a great fit" / "I'm a direct match" / "my stack matches yours" — self-claims instead of facts

Pre-delivery checklist

Run after the lint loop exits 0:

No em-dash on inserts? → replaced with : / , / separate sentence.
Lint script exits 0? → required.
Rhythm still uneven? → short and long sentences alternate, no monotone flow.
Anchor nouns present (harness, registry, loop, flow, score, etc.) where they fit?
No JD recap in the opening of a cover letter?
Colon-headers used where the format is "phrase + content"?
Closing is short and dry? (e.g. "Open to chat if it's a fit.")
"Yes." opener only if this is a direct reply to a closed question?
No "one thing to flag / honest framing" construction?
Grammar signatures dosed correctly for the channel (more in DMs, less in cover letters)?

15 KiB Raw Blame History