| name | prompt-warrior |
| description | Use when the user wants to analyze their communication style with AI from Claude Code logs — asks for a "prompt warrior" profile, an "AI profile", "wrapped", a psychological portrait of how they prompt, statistics of their sessions/messages, or personalized prompt-engineering recommendations based on their chat history. Works in any language (RU/EN lexicons built in). |
Prompt Warrior
Builds a gamified analytical portrait of the user's AI collaboration style from their
local Claude Code JSONL logs: real metrics on a FIXED scale (comparable between people)
plus fun statuses — a title, a level, achievements with rarity — and evidence-based
prompt-engineering recommendations. It is NOT an RPG character sheet:
no classes, no stat blocks; the analytics are the product, the statuses are the garnish.
The scale is a contract. All formulas and thresholds live in scripts/analyze.py
(SCALE v1) and references/scale.md. NEVER adjust formulas, thresholds, rank mappings,
or achievement conditions per user — comparability is the whole point. If a metric looks
unflattering, it stays. Report numbers only from profile.json, never invent or soften them.
Workflow
-
Run the analyzer (Python 3, stdlib only, read-only — it never modifies logs):
python scripts/analyze.py -o profile.json # all time, all projects
python scripts/analyze.py --days 7 -o profile.json # last week
python scripts/analyze.py --days 30 -o profile.json # last month
python scripts/analyze.py --date-from 2026-06-01 --date-to 2026-06-30 -o profile.json
python scripts/analyze.py --project <dir-name> -o profile.json # one project
THE RANGE IS THE USER'S CHOICE: if they named a period («за неделю», "this month",
dates), pass it; if they didn't, default to all time and mention that a range is
available. Never pick a range for the user silently.
Default logs location: ~/.claude/projects. Use --projects-dir to override.
Try python3 if python is missing.
RUNNING IN A DIFFERENT HARNESS (OpenCode, Codex CLI, Gemini CLI, Copilot)?
See references/harnesses.md — known log locations and formats, and how to plug a
new source into SOURCES. If your paths differ, orient by analogy as described there.
-
Read profile.json.
{"error": "not_enough_data"} → tell the user at least 30 of their own messages are
needed; stop."low_confidence": true (< 100 messages) → include a visible "small sample" caveat.
-
Generate the local avatar (optional, offline) — a deterministic monster from
the user's title via the robohash package (assembles bundled PNG layers locally;
zero network):
python scripts/avatar.py --from-profile profile.json -o avatar.png
If the package is missing, offer ONCE: «поставить pip install robohash для
локальных аватаров-монстров?» — install on yes, skip on no. Without the avatar the
card simply renders without an image — NEVER substitute placeholders, letters or
icons in its place. The monster evolves when the title changes — mention it.
3а. Write the Warrior's Chronicle («Хроника воина») — 2-4 paragraphs (~100-180
words) that YOU, the model, author from profile.json: a characterization /
mini-biography of the user in the grimoire-chronicle voice with light humor. Rules:
facts strictly from the profile (no inventions), NO numbers (numbers live in the
card sections — each number appears exactly once), neutral to metric values (do not
shame, do not flatter), 1-2 accents max, written in the card's language. This
is the one place the model speaks freely — make it feel written, not generated.
-
Build and SHOW the card — output degrades gracefully, in this order:
a. You do NOT write HTML/CSS/SVG. Write content.json — ONLY your authored parts,
in the card language:
{"chronicle": ["<p1>", "<p2>", "<p3>"],
"strengths": ["..."],
"weaknesses": [{"w": "...", "a": "...", "ev": "...", "url": "https://..."}]}
(chronicle = step 3а; strengths ≤3 and weakness→fix pairs ≤4 picked by the rules
in references/widget.md). Then run the canonical renderer:
python scripts/render.py profile.json --content content.json --lang ru -o ai-profile.html
(--lang en for English — labels are symmetric, data fields are bilingual.)
It fills the whole card: numbers exactly once, bars, gauges, icons, favicon,
seal, achievements accordion, avatar if avatar.png exists. Tell the user the
path. NEVER Read .svg asset files into context (widget.md, rule 2д).
b. OPEN it for the user: if a preview panel/tool is available, open it there;
otherwise open in the default browser ( on Windows, on macOS,
on Linux). Do not make the user open the file by hand.
c. If an inline widget tool is available, ALSO render an inline widget mirroring
the same sections and numbers.
d. If NONE of the above is available (headless/console-only environment), render
the card as TEXT in the console following the text-fallback spec at the end of
— same sections, same numbers, bars as unicode blocks.
