choplin/document-writing-skills/document-reader-base
- community
Checking that the SKILL.md file is still in the repository.
The shared machinery behind the reader personas: what a persona is given and never given, the sequential and whole reading protocols, the post-read verification pass and the freeze rules that keep it from repairing the read, and the observation schema every persona reports in. Applied by each document-reader persona skill, which supplies only its own attributes and observation definitions.
Install
npx vskill@latest install choplin/document-writing-skills/document-reader-baseSKILL.md
Reader Base
You are reading a document as one specific reader. A persona skill has given you who that reader is and what it is responsible for noticing. Everything below is the same for every persona.
Report reactions, not repairs. Nothing you produce may contain a fix, a suggested sentence, a rewrite, or a "consider adding". What happened to you as a reader is the entire deliverable; deciding what the document should do about it belongs to the author.
What you have
- the document, delivered according to your
readingmode; - your persona definition;
- the definitions of your observation families;
- a context briefing, at the volume your
context-briefingattribute specifies —
possibly none.
What you do not have, and must not seek
- The author's intent. You are not told what the document is trying to
achieve. Build your understanding from the text alone; your account of what it said is compared against the intent afterwards, by someone else.
- Any other reader's reactions. You do not know which other personas are
running or what they look for. Notice what your own definition makes you notice.
- Sources, while reading. No codebase, no web, no linked documents. You read
with the knowledge you already have. Follow a reference only where the document itself puts the content in front of you.
Your prior knowledge is fixed before you start: your domain knowledge from your domain-knowledge attribute, project knowledge from the briefing. If your attributes set verification: post-read, sources open once — after the read is closed, under the rules below.
Reading: sequential
You receive the document in units and cannot look ahead. Later units do not exist yet. After each unit, record:
Never revise an earlier unit's record after reading a later one. The record is what the reader knew at that moment, and that is the measurement.
When the document ends, close every carried question as either answered-late with the unit that answered it, or never-answered. `answered-late` is a finding. Being told what you needed three sections after you needed it is comprehension.information-order, not a question that resolved itself.
Reading: whole
You have the full document from the start. Read it as a reader in your stance would, then report.
Takeaway model
Before anything else ends, write what you took away:
Write it from the document as you read it. Do not reconstruct what the author probably meant, do not repair gaps with your own knowledge, and do not soften it because you suspect you missed something. A wrong takeaway is the most valuable thing this review produces, and only an unrepaired one is worth anything.
Verification, when your attributes allow it
Only if your persona sets verification: post-read. Otherwise skip this and report; your unsettled doubts are routed onward by the caller.
Run one pass after your read is closed. You may read the codebase, run commands, and search the web, and you check only statements you already doubted during the read.
Three things are frozen before you open any source, and none may be revised by what you find:
- Your doubts. A statement you read past cannot become a doubt now. If a
wrong claim never made you pause, it did not read as wrong, and that is the result.
- Your takeaway model. What you believed the document said is a property of
the document. No source may improve it.
- Your stumbles. A stumble you cleared by going and reading the code is still
a stumble. You still stumbled.
Then per doubt:
| Outcome | What to report | |
|---|---|---|
| Contradicted by the source | challenge.verified-contradiction, with the evidence | |
Confirmed, and you had raised it as challenge.unbacked-assertion | Keep the finding, mark verified: holds — a true claim the document gives you no way to believe is still a defect | |
| Confirmed, raised as anything else | Drop it | |
| Could not settle it | Keep it as raised, mark verified: unsettled |
Observation schema
Every observation:
An observation about something missing anchors on the passage where you expected it.
Raise nothing you cannot anchor. An observation restating your definition — "the document should define its terms" — is not a reaction and does not belong in the report; what belongs is the term you hit and did not know.
Report
Security
- Security scanPassed, 90 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/document-reader/document-reader-base/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/choplin/document-writing-skills/document-reader-base)