prompt-profile 0.1.5 → 0.2.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +173 -360
- package/dist/bin.js +13868 -10069
- package/dist/prompts/map.md +69 -5
- package/dist/prompts/reduce.md +213 -10
- package/dist/prompts/review.md +112 -0
- package/dist/prompts/translate.md +17 -1
- package/package.json +3 -2
package/README.md
CHANGED
|
@@ -1,417 +1,230 @@
|
|
|
1
1
|
# prompt-profile
|
|
2
2
|
|
|
3
|
-
|
|
4
|
-
with
|
|
5
|
-
Reading, cleaning, redaction and the report happen locally; the selected, redacted prompts are sent
|
|
6
|
-
through that CLI to its model service and use your subscription quota (no API key).
|
|
7
|
-
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
[Sample HTML report](https://erictsai83.github.io/prompt-profile/sample-report.html)
|
|
11
|
-
· [Sample profile JSON](https://erictsai83.github.io/prompt-profile/sample-profile.json) (all sample data is hand-written and fictional;
|
|
12
|
-
every example is rendered by the current code on each change to `main`).
|
|
13
|
-
|
|
14
|
-
### Example reports
|
|
15
|
-
|
|
16
|
-
The same fictional profile, rendered two ways:
|
|
17
|
-
|
|
18
|
-
- [Full report](https://erictsai83.github.io/prompt-profile/demo/full/report.html): what you see on your own machine, with prompt
|
|
19
|
-
quotes, suggested rewrites, project names, exact dates and fun facts.
|
|
20
|
-
- [Shareable report](https://erictsai83.github.io/prompt-profile/demo/without-personal/report.html): the same profile rendered with
|
|
21
|
-
`--exclude personal`: no quotes, rewrites or fun facts, projects shown as "Project A, B…",
|
|
22
|
-
dates by month. This is what you would post publicly.
|
|
23
|
-
|
|
24
|
-
The card never contains prompt quotes. Project names may appear in normal mode;
|
|
25
|
-
`--exclude projects` (also included in `--exclude personal`) anonymizes known project names
|
|
26
|
-
in both the report and card, including generated identity text. The full report
|
|
27
|
-
in every look (`--look`; render only, no new analysis):
|
|
28
|
-
|
|
29
|
-
<table>
|
|
30
|
-
<tr>
|
|
31
|
-
<td width="50%" align="center">
|
|
32
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/full/report.html"><img src="https://erictsai83.github.io/prompt-profile/demo/full/card.png" width="420" alt="Classic card"></a><br>
|
|
33
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/full/report.html">Classic</a> (default)
|
|
34
|
-
</td>
|
|
35
|
-
<td width="50%" align="center">
|
|
36
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/signal/report.html"><img src="https://erictsai83.github.io/prompt-profile/demo/signal/card.png" width="420" alt="Signal card"></a><br>
|
|
37
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/signal/report.html">Signal</a>
|
|
38
|
-
</td>
|
|
39
|
-
</tr>
|
|
40
|
-
<tr>
|
|
41
|
-
<td width="50%" align="center">
|
|
42
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/afterhours/report.html"><img src="https://erictsai83.github.io/prompt-profile/demo/afterhours/card.png" width="420" alt="Afterhours card"></a><br>
|
|
43
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/afterhours/report.html">Afterhours</a>
|
|
44
|
-
</td>
|
|
45
|
-
<td width="50%" align="center">
|
|
46
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/studio/report.html"><img src="https://erictsai83.github.io/prompt-profile/demo/studio/card.png" width="420" alt="Studio card"></a><br>
|
|
47
|
-
<a href="https://erictsai83.github.io/prompt-profile/demo/studio/report.html">Studio</a>
|
|
48
|
-
</td>
|
|
49
|
-
</tr>
|
|
50
|
-
</table>
|
|
3
|
+
Turn your local Claude Code / Codex prompt history into an evidence-backed report of how you
|
|
4
|
+
collaborate with AI. Uses your logged-in `claude` or `codex` CLI: no API key needed.
|
|
51
5
|
|
|
52
|
-
|
|
6
|
+
**Analysis sends selected, redacted prompts to your CLI's model service and uses your
|
|
7
|
+
subscription quota.** Reports and analysis caches stay on your computer.
|
|
53
8
|
|
|
54
|
-
|
|
55
|
-
and log in once (`claude`, or `codex login`). prompt-profile uses it to analyse your prompts;
|
|
56
|
-
**no API key needed**.
|
|
57
|
-
2. Run it (Node.js ≥ 24):
|
|
9
|
+
|
|
58
10
|
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
11
|
+
[Full report](https://erictsai83.github.io/prompt-profile/sample-report.html)
|
|
12
|
+
· [Shareable report](https://erictsai83.github.io/prompt-profile/demo/without-personal/report.html)
|
|
13
|
+
· [Template gallery](https://erictsai83.github.io/prompt-profile/demo/templates/index.html)
|
|
62
14
|
|
|
63
|
-
|
|
15
|
+
All examples use fictional data.
|
|
64
16
|
|
|
65
|
-
|
|
17
|
+
## Quick start
|
|
66
18
|
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
In range: Claude Code 1,204 · Codex 311 prompts
|
|
70
|
-
1,515 prompts in 182 sessions after cleaning → 1,515 selected in 24 chunks
|
|
71
|
-
No sampling: every chunk is analysed
|
|
72
|
-
Read, cleaned and redacted on this computer; the report and analysis cache are written only here.
|
|
73
|
-
● Sent off this computer: the selected, redacted prompts go through your logged-in Claude Code CLI (claude -p) to its model service
|
|
74
|
-
No API key: uses your subscription quota, base estimate 25 calls (24 chunks + 1 final profile)
|
|
75
|
-
Intermediate merges and retries (up to 2 per call) can add calls.
|
|
76
|
-
No hard limit: add --max-calls <n> to stop after n CLI calls.
|
|
77
|
-
Free first: --dry-run writes what would be sent without calling any LLM; `prompt-profile stats` shows statistics.
|
|
78
|
-
```
|
|
19
|
+
You need Node.js ≥ 24 and either [Claude Code](https://claude.com/claude-code) or
|
|
20
|
+
[Codex](https://github.com/openai/codex) installed and logged in.
|
|
79
21
|
|
|
80
|
-
|
|
22
|
+
```bash
|
|
23
|
+
npx prompt-profile@latest
|
|
24
|
+
```
|
|
81
25
|
|
|
82
|
-
|
|
83
|
-
|
|
26
|
+
The CLI shows the selected history and estimated calls, then asks for confirmation before
|
|
27
|
+
sending anything. By default it analyses the last two months and opens the finished report
|
|
28
|
+
in your browser. Files are saved under `./prompt-profile-report/`.
|
|
84
29
|
|
|
85
|
-
|
|
30
|
+
To inspect your history without calling an LLM (no login required):
|
|
86
31
|
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
32
|
+
```bash
|
|
33
|
+
npx prompt-profile@latest stats # local statistics only
|
|
34
|
+
npx prompt-profile@latest --dry-run # export the redacted chunks that would be sent
|
|
35
|
+
```
|
|
91
36
|
|
|
92
|
-
|
|
37
|
+
Dry-run exports contain your prompt text: keep them private and delete them when done.
|
|
38
|
+
The call estimate is not an upper bound; use `--max-calls` to cap CLI invocations, including
|
|
39
|
+
merges, evidence review and retries.
|
|
93
40
|
|
|
94
41
|
## Install
|
|
95
42
|
|
|
43
|
+
To install or update a global command:
|
|
44
|
+
|
|
96
45
|
```bash
|
|
97
|
-
|
|
98
|
-
npm i -g prompt-profile # or install the `prompt-profile` command
|
|
46
|
+
npm i -g prompt-profile@latest
|
|
99
47
|
prompt-profile --help
|
|
100
|
-
npm i -g prompt-profile@latest # update the installed command
|
|
101
48
|
```
|
|
102
49
|
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
50
|
+
Without Node.js, download a standalone executable for macOS, Linux or Windows from
|
|
51
|
+
[GitHub Releases](https://github.com/EricTsai83/prompt-profile/releases).
|
|
52
|
+
See the [standalone installation guide](https://github.com/EricTsai83/prompt-profile/blob/main/docs/user-guide.md#standalone-installation)
|
|
53
|
+
for extraction and macOS setup.
|
|
107
54
|
|
|
108
|
-
|
|
109
|
-
tar -xzf prompt-profile-<version>-darwin-arm64.tar.gz
|
|
110
|
-
./prompt-profile-<version>-darwin-arm64/prompt-profile
|
|
111
|
-
# macOS, if the archive was downloaded with a browser (the binary is not notarised):
|
|
112
|
-
xattr -d com.apple.quarantine ./prompt-profile-<version>-darwin-arm64/prompt-profile
|
|
113
|
-
```
|
|
55
|
+
## What you get
|
|
114
56
|
|
|
115
|
-
|
|
116
|
-
|
|
57
|
+
- **HTML report**: an offline portrait of your collaboration preferences, supported observations
|
|
58
|
+
and prompt suggestions, plus activity charts, workflow techniques and source evidence.
|
|
59
|
+
- **PNG share cards**: a 1200 × 630 card per language, without prompt quotes.
|
|
60
|
+
- **`profile.json`**: the data used to render the report again without another analysis.
|
|
61
|
+
- **Local history**: successful analyses are kept in a report library; open the output folder's
|
|
62
|
+
`index.html` to browse them.
|
|
117
63
|
|
|
118
|
-
|
|
64
|
+
Reports support English and Traditional Chinese; quotes stay in their original language.
|
|
65
|
+
The default is English, or both languages on a Chinese system locale. Each extra language
|
|
66
|
+
normally uses one translation call, with one retry for an incomplete or failed translation.
|
|
67
|
+
A Chinese-only request stops if translation still fails; it does not substitute an English report. To export a PDF, use your browser's Print → Save as PDF.
|
|
119
68
|
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
- **No API key needed**: the analysis runs your own logged-in `claude -p` or `codex exec` in the
|
|
124
|
-
background and **uses your subscription quota**. The base call estimate is shown
|
|
125
|
-
before confirmation. Intermediate merges and up to two retries per call can add calls;
|
|
126
|
-
use `--max-calls` to set a hard limit on CLI invocations, including retries.
|
|
69
|
+
Observations cite verified prompt quotes, but quote matching and model review do not prove an
|
|
70
|
+
interpretation is correct. Small histories may show “Early observations” instead of a role.
|
|
71
|
+
The [radar](docs/ability-scores.md) assesses six AI collaboration abilities from verified quotes in these conversations. Scores describe supported practice, not psychological traits or population ranks; “not assessed” is not a low score.
|
|
127
72
|
|
|
128
|
-
|
|
73
|
+
New analyses also retain [reusable instruction provenance](docs/instruction-provenance.md): what a skill is documented to do, who selected it for the task, and which historical requirements were supplied.
|
|
129
74
|
|
|
130
|
-
|
|
131
|
-
For a static PDF, open `report.html`, select a language, then use your browser’s Print → Save
|
|
132
|
-
as PDF. Printing includes both report views in that language. The offline HTML stays interactive;
|
|
133
|
-
both formats use the same profile without another analysis call. Files in `output/pdf/` are
|
|
134
|
-
fictional development review exports, not automatic CLI outputs.
|
|
135
|
-
|
|
136
|
-
Written to `./prompt-profile-report/` (change it with `--out`):
|
|
137
|
-
|
|
138
|
-
- `report.html`: your private, single-file report that works offline, with two pages:
|
|
139
|
-
- **Who they are** (opens first): archetype, up to three observations with medium or high
|
|
140
|
-
confidence each with one of your own prompts, observed collaboration preferences, the portrait story with its quotes,
|
|
141
|
-
and fun facts
|
|
142
|
-
- **Facts & evidence**: every measurement with its source and denominator: four measured
|
|
143
|
-
habits with their definitions, tool and prompt-wording shares, a daily calendar and a
|
|
144
|
-
weekday × hour heatmap (hover or tap for the count and main projects), project chapters,
|
|
145
|
-
all traits with their quotes and how widely they are supported, the evidence for and
|
|
146
|
-
against each preference, the five dimensions and how it was measured
|
|
147
|
-
- **how you work with AI** (facts page): a capability map (plan, delegate, extend, verify,
|
|
148
|
-
ship) with the skills you used and what each does, subagent types, MCP servers, models,
|
|
149
|
-
pull requests, parallel sessions and prompt-writing habits; counted from tool metadata and
|
|
150
|
-
the skill descriptions recorded in your transcripts, never tool inputs, outputs or skill
|
|
151
|
-
arguments
|
|
152
|
-
- **how you steer the AI** (facts page): in how many of the analysed conversations with
|
|
153
|
-
follow-ups you redirected the AI's approach, and separately how often you asked it to fix a
|
|
154
|
-
broken result, with examples; one-shot requests are not counted against you
|
|
155
|
-
- **how to improve your prompts** (facts page): 2–4 suggestions, each with a real prompt of
|
|
156
|
-
yours and a suggested rewrite, plus signals such as very short
|
|
157
|
-
follow-ups, and what changed since your last report
|
|
158
|
-
- **usage during this period** (facts page only): Claude Code and Codex side by side with
|
|
159
|
-
the plan each runs on, tokens by kind, the share of each tool, months and models, and the
|
|
160
|
-
**API list-price equivalent** in USD. Detailed breakdowns stay here; the card can show a
|
|
161
|
-
compact token total and API estimate. The "Who they are" page does not show usage
|
|
162
|
-
- a language switch when there is more than one language
|
|
163
|
-
|
|
164
|
-
It shows everything by default, quotes and project folder names included. To share it, leave
|
|
165
|
-
parts out with `--exclude` (below).
|
|
166
|
-
|
|
167
|
-
- `card.png`: a 1200 × 630 share card in the default language, plus `card.<lang>.png` for each
|
|
168
|
-
other language (**omits evidence quotes**). Includes the introduction, up to two supported
|
|
169
|
-
behaviors, prompt/conversation counts, token usage and an API estimate in USD. Missing and
|
|
170
|
-
partial pricing are labeled; `--exclude usage` removes usage from HTML and PNG.
|
|
171
|
-
`--card-hide stats` hides the card’s measured summary
|
|
172
|
-
- `profile.json`: the raw profile data (re-create the report with `prompt-profile render`)
|
|
173
|
-
|
|
174
|
-
### Leaving parts out
|
|
175
|
-
|
|
176
|
-
`--exclude` (on the main command and on `render`) takes a comma list:
|
|
177
|
-
|
|
178
|
-
| Part | Left out |
|
|
179
|
-
| -------------- | ---------------------------------------------------------------------------------------------------- |
|
|
180
|
-
| `quotes` | Evidence quotes and the before / after rewrites |
|
|
181
|
-
| `projects` | Project names, shown as "Project A, B…", also inside the generated text and on the card |
|
|
182
|
-
| `dates` | Exact dates: no daily calendar, the data period by month |
|
|
183
|
-
| `coaching` | "Improve your prompts" and "since your last report" |
|
|
184
|
-
| `funfacts` | Fun facts |
|
|
185
|
-
| `custom-names` | Names and descriptions of your own skills, custom agent types and private MCP servers (only counted) |
|
|
186
|
-
| `usage` | Plans, token counts and API list-price equivalents |
|
|
187
|
-
| `personal` | All of the above |
|
|
75
|
+
## Usage
|
|
188
76
|
|
|
189
77
|
```bash
|
|
190
|
-
prompt-profile
|
|
191
|
-
|
|
78
|
+
prompt-profile --since 6m # 1m / 2m (default) / 3m / 6m / 1y / all
|
|
79
|
+
prompt-profile --engine codex # default: auto, Claude Code first
|
|
80
|
+
prompt-profile --model <name> # choose the analysis model
|
|
81
|
+
prompt-profile --max-calls 60 # cap CLI calls, including retries
|
|
82
|
+
prompt-profile --engine claude --max-cost-usd 1.50 # local stop threshold; may overshoot
|
|
83
|
+
prompt-profile --lang zh-TW,en # first language is the default
|
|
84
|
+
prompt-profile --voice first # I/我; second (you/你) is the default
|
|
85
|
+
prompt-profile --out <dir> # choose the output folder
|
|
86
|
+
|
|
87
|
+
prompt-profile render profile.json --look prism # change the look without LLM calls
|
|
88
|
+
prompt-profile cache clear # delete analysis cache entries
|
|
89
|
+
prompt-profile history --executions # list saved execution records (free)
|
|
192
90
|
```
|
|
193
91
|
|
|
194
|
-
|
|
92
|
+
Seven looks are available: classic (default), signal, afterhours, studio, prism, slate and folio.
|
|
93
|
+
Rendering a new look keeps the original report. Use `--card-hide stats` to hide the card's
|
|
94
|
+
measured summary, or `--avatar` to choose its illustration.
|
|
195
95
|
|
|
196
|
-
|
|
197
|
-
|
|
198
|
-
|
|
96
|
+
See the [user guide](https://github.com/EricTsai83/prompt-profile/blob/main/docs/user-guide.md)
|
|
97
|
+
for all documented options, language behavior, history and recovery. If a run stops, follow the
|
|
98
|
+
printed retry command: completed chunks and merges are cached; later steps may use quota again.
|
|
199
99
|
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
203
|
-
|
|
204
|
-
```
|
|
100
|
+
Each analysis saves calls, retries, cache reuse, token coverage and cost provenance, including
|
|
101
|
+
failed executions. The report shows generation usage separately from historical activity.
|
|
102
|
+
See [execution accounting](docs/execution-accounting.md) for recovery, unknown measurements
|
|
103
|
+
and cost-threshold limits.
|
|
205
104
|
|
|
206
|
-
|
|
207
|
-
| ------------- | ------------------------------------------------------------------------------------------------------- | ------------------------ | -------------- | ------------------------ |
|
|
208
|
-
| Secrets | private keys, API tokens, JWTs, bearer tokens, connection strings, `password=`… , random-looking tokens | masked | masked | nothing written (exit 1) |
|
|
209
|
-
| Personal data | email, home path, username, internal URL, phone, Taiwan ID, credit card | masked | kept, listed | nothing written (exit 1) |
|
|
105
|
+
## How much a run uses
|
|
210
106
|
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
107
|
+
A run reads everything locally, then analyses a **sample** of the selected history: about
|
|
108
|
+
20–26 chunks of redacted prompts (roughly 700k characters) once the history is larger than
|
|
109
|
+
that. Time, tokens and cost therefore stop growing with history size; `--full` analyses every
|
|
110
|
+
chunk and scales with the history instead. `--dry-run` prints your own estimate for free.
|
|
214
111
|
|
|
215
|
-
|
|
112
|
+
### Measured runs
|
|
216
113
|
|
|
217
|
-
|
|
218
|
-
|
|
219
|
-
prompt-profile --since 3m # 1m / 3m / 6m / 1y / all (also <n>d, <n>w)
|
|
220
|
-
prompt-profile --engine codex # which CLI analyses your prompts (default: auto, Claude Code first)
|
|
221
|
-
prompt-profile --model <name> # model passed to claude / codex
|
|
222
|
-
prompt-profile --effort medium # reasoning effort: low / medium / high / xhigh / max (default: the CLI's)
|
|
223
|
-
prompt-profile --out <dir> # output folder
|
|
224
|
-
prompt-profile --max-calls 40 # cap CLI invocations, including merges and retries
|
|
225
|
-
prompt-profile --full # no sampling, analyse everything (slower, uses more quota)
|
|
226
|
-
prompt-profile --no-cache # do not read the analysis cache
|
|
227
|
-
prompt-profile --yes # skip the confirmation (required in non-interactive sessions)
|
|
228
|
-
prompt-profile --dry-run # free: the summary, plus what would be sent written into --out
|
|
229
|
-
prompt-profile --verbose # full errors and LLM call timings
|
|
230
|
-
prompt-profile --lang zh-TW,en # report languages, the first is the default (en, zh-TW)
|
|
231
|
-
prompt-profile --look signal # classic (default) / signal / afterhours / studio / prism / slate / folio
|
|
232
|
-
prompt-profile --card-hide stats,graphic # leave traits / stats / graphic out of card.png
|
|
233
|
-
prompt-profile --name Eric # title the report "Eric’s prompt profile" (not stored in profile.json)
|
|
234
|
-
|
|
235
|
-
prompt-profile stats # statistics only, no LLM (free and fast)
|
|
236
|
-
prompt-profile render profile.json # re-create the report from an existing profile.json
|
|
237
|
-
prompt-profile cache clear # delete the analysis cache
|
|
238
|
-
|
|
239
|
-
prompt-profile render profile.json --look studio --card-hide stats
|
|
240
|
-
# a new look, written to variants/studio-no-stats/
|
|
241
|
-
```
|
|
114
|
+
One real history, default settings (last 2 months, sampled, report in Traditional Chinese
|
|
115
|
+
and English), measured on 2026-11-01 with both engines on the same input:
|
|
242
116
|
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
|
|
246
|
-
|
|
247
|
-
as the report.
|
|
248
|
-
|
|
249
|
-
Languages: without `--lang`, a Chinese system locale gives a Traditional Chinese report with an
|
|
250
|
-
English switch (`zh-TW,en`); any other locale gives English. The profile is written in English
|
|
251
|
-
and then translated, one extra CLI call per extra language; quotes stay as you wrote them.
|
|
252
|
-
|
|
253
|
-
Exit codes: `0` success · `1` error · `2` usage / needs `--yes` · `3` no data · `4` no usable LLM
|
|
254
|
-
CLI · `5` rate limited · `130` interrupted (a re-run resumes from the cache).
|
|
255
|
-
|
|
256
|
-
## Small histories and stopped runs
|
|
257
|
-
|
|
258
|
-
**Small histories.** A role, featured observations and prompt coaching each need verified quotes
|
|
259
|
-
from at least two sessions (the same rule that sets trait confidence and preference axes). When
|
|
260
|
-
nothing in a profile reaches that, the report is titled "Early observations": it states how many
|
|
261
|
-
prompts and sessions it is based on, keeps the observations, five dimensions and measurements,
|
|
262
|
-
and leaves out the generated role and the "Improve your prompts" section. The card does the same.
|
|
263
|
-
The terminal says so after the run; a longer `--since` range or more history adds evidence.
|
|
264
|
-
|
|
265
|
-
**If a run stops.** Progress is shown by phase (chunks, merges, the final profile, verification,
|
|
266
|
-
translations, writing files); there is no time estimate. When a run stops early, the error says
|
|
267
|
-
what was saved and prints the exact command to run again:
|
|
268
|
-
|
|
269
|
-
| What happened | What is saved | What a re-run does |
|
|
270
|
-
| --------------------------------------------------------------------- | ------------------------------------------------------- | ----------------------------------------------------------------------------- |
|
|
271
|
-
| Ctrl-C, rate limit, timeout, malformed output, too many failed chunks | Finished chunk results in the analysis cache; no report | Reuses cached chunks and merges; the final profile and translations run again |
|
|
272
|
-
| `--max-calls` reached | Same; cached results do not use the budget again | Same; raise `--max-calls` if it stops again |
|
|
273
|
-
| The final profile step failed | Chunk results | Repeats only the final profile and translation calls |
|
|
274
|
-
| An intermediate merge failed (large histories) | Chunk results and finished merges | Repeats the unfinished merges, then the final profile and translations |
|
|
275
|
-
| A translation failed | The report, without that language | The printed command adds it (final profile + translation) |
|
|
276
|
-
| A card was not created (e.g. the Chinese font download failed) | The report and `profile.json` | `prompt-profile render <profile.json>` retries the cards, no analysis |
|
|
277
|
-
|
|
278
|
-
The re-run command keeps your flags and leaves out `--no-cache`, so it reads the cache. When
|
|
279
|
-
prompt-profile was started with npx, the printed commands start with `npx prompt-profile@<version>`
|
|
280
|
-
(the same version, so the cache matches); otherwise with the installed `prompt-profile`.
|
|
281
|
-
A CLI's own error message can quote what it was sent, so it is shown only with `--verbose`.
|
|
282
|
-
|
|
283
|
-
## Privacy
|
|
284
|
-
|
|
285
|
-
**Files read** (read-only):
|
|
286
|
-
|
|
287
|
-
| Source | Paths (override with an environment variable) |
|
|
288
|
-
| ----------- | ------------------------------------------------------------------------------------------------------------------ |
|
|
289
|
-
| Claude Code | `~/.claude/projects/*/*.jsonl`, `~/.claude/history.jsonl` (`CLAUDE_CONFIG_DIR`) |
|
|
290
|
-
| Codex | `~/.codex/sessions/**/rollout-*.jsonl` (and zstd-compressed `.jsonl.zst`), `~/.codex/history.jsonl` (`CODEX_HOME`) |
|
|
291
|
-
|
|
292
|
-
**Usage and plan** (the "Usage during this period" section): token counts and model names from
|
|
293
|
-
the same logs (Claude Code subagent transcripts in `projects/*/*/subagents/` included), the plan
|
|
294
|
-
from `claude auth status --json` (`subscriptionType` only), an undocumented rate-limit hint from
|
|
295
|
-
`~/.claude.json` (`oauthAccount.userRateLimitTier` only), and the ChatGPT plan type from
|
|
296
|
-
`~/.codex/auth.json`, decoded locally from the id token's payload. No token, API key, email,
|
|
297
|
-
account or organization id is stored, cached, logged or sent; a missing or unreadable file gives
|
|
298
|
-
"unknown", never an error. Usage, plans and amounts stay on your machine: they are not sent to
|
|
299
|
-
the LLM. Amounts are computed from an embedded list-price table (no download) and are an **API
|
|
300
|
-
list-price equivalent**: what the usage would cost on the public API, not what you paid. Neither
|
|
301
|
-
provider offers a way to read a subscription's usage as a bill, so this is how every local usage
|
|
302
|
-
tool works (Claude Code's own session cost does the same); responses in fast mode (Claude
|
|
303
|
-
`speed: "fast"`, Codex priority) use the fast-mode list prices, and a model or tier without a
|
|
304
|
-
listed price stays "unknown" rather than being estimated.
|
|
305
|
-
|
|
306
|
-
**Prompt selection and preprocessing**: adapters select user prompts, excluding recognized system
|
|
307
|
-
messages, tool output and sub-agent conversations, then:
|
|
308
|
-
|
|
309
|
-
- pasted code / logs are collapsed into `[pasted code: N lines]`
|
|
310
|
-
- API keys, tokens, passwords, private keys, connection strings, emails, phone numbers, national
|
|
311
|
-
IDs, credit card numbers, internal URLs, UUIDs / hashes, home-folder paths and usernames that
|
|
312
|
-
match the supported rules are redacted **before** anything is sent (and before long
|
|
313
|
-
prompts are shortened); this is best-effort, not a guarantee that every secret is found
|
|
314
|
-
- projects are reduced to their folder name, never the full path
|
|
315
|
-
- `prompt-profile --dry-run` previews selected map chunks; generated observations, statistics
|
|
316
|
-
used by the final step, system instructions and retry messages are not a complete network dump
|
|
317
|
-
|
|
318
|
-
Selected, cleaned and redacted prompt text is sent **off your machine** through your logged-in
|
|
319
|
-
CLI to its model service, along with project basenames. Intermediate observations, supporting
|
|
320
|
-
quotes and prompt statistics are sent during merge/final analysis. Text that does not match a
|
|
321
|
-
redaction rule can remain unchanged, including names and confidential prose.
|
|
322
|
-
|
|
323
|
-
Reports, `profile.json`, dry-run exports and analysis caches are written locally and are not
|
|
324
|
-
uploaded by prompt-profile. They can contain personal information. Created folders use `0700`
|
|
325
|
-
and files use `0600`; existing folder permissions are not changed. `cache clear` removes analysis
|
|
326
|
-
entries only, and reports deletion errors rather than claiming success. Delete reports and dry-run
|
|
327
|
-
exports separately when no longer needed.
|
|
328
|
-
|
|
329
|
-
Claude runs with safe mode, tools disabled, strict MCP configuration and disabled hooks. Codex
|
|
330
|
-
runs with `--ignore-user-config`, disabled hooks/notifications, a read-only sandbox and an ephemeral
|
|
331
|
-
session. Codex still uses your existing login, but user-configured providers and model defaults
|
|
332
|
-
are ignored; specify the model with `--model`. Both use fresh temporary working folders.
|
|
333
|
-
Administrator policy, environment settings and CLI/provider diagnostics can still apply. These
|
|
334
|
-
controls are tested on specific versions, and do not promise zero CLI/provider logs or traffic.
|
|
335
|
-
See [runner notes](src/llm/NOTES.md) and the [security policy](SECURITY.md).
|
|
336
|
-
|
|
337
|
-
prompt-profile has no telemetry. The report makes no network requests and its fonts ship with
|
|
338
|
-
the package. The one exception is the Chinese share card: on first use it downloads Noto Sans TC
|
|
339
|
-
(about 11 MB, from a pinned GitHub commit), checks its SHA-256 and caches it; no prompt data is
|
|
340
|
-
sent. The borrowed CLI manages its own network requests and provider data retention.
|
|
117
|
+
| Input | |
|
|
118
|
+
| ------------------ | --------------------------------------------------------------------- |
|
|
119
|
+
| History in range | 4,374 prompts in 613 conversations (Claude Code and Codex) |
|
|
120
|
+
| Analysed (sampled) | 1,499 prompts in 252 conversations: 21 of 68 chunks, ~665k characters |
|
|
341
121
|
|
|
342
|
-
|
|
122
|
+
| | Claude Code | Codex |
|
|
123
|
+
| ------------------------- | ---------------------------------- | ------------------------------ |
|
|
124
|
+
| Model | `claude-opus-5-5` (default effort) | `gpt-6.1-sol` (default effort) |
|
|
125
|
+
| CLI calls | 48 (no retries) | 48 (no retries) |
|
|
126
|
+
| Wall time | 25 min | 20.5 min |
|
|
127
|
+
| Tokens in / out | 1.46M / 0.38M | 1.73M / 0.29M |
|
|
128
|
+
| API list-price equivalent | **US$13.8** (reported by the CLI) | **≈ US$6.1** (estimated¹) |
|
|
343
129
|
|
|
344
|
-
|
|
345
|
-
|
|
346
|
-
|
|
347
|
-
invocations rather than tokens, provider requests or money.
|
|
348
|
-
Chunks analysed before are read from the cache on the next run.
|
|
130
|
+
¹ Codex does not report which model answered, so prompt-profile leaves its cost unpriced; the
|
|
131
|
+
estimate applies `gpt-6.1-sol` list prices to the measured tokens and leaves out cache writes,
|
|
132
|
+
which the Codex log does not report.
|
|
349
133
|
|
|
350
|
-
|
|
134
|
+
With a subscription these calls use your plan's quota; the dollar figures are what the same
|
|
135
|
+
tokens would cost on the API, not a bill. Another model, `--effort` level or history mix will
|
|
136
|
+
change them.
|
|
351
137
|
|
|
352
|
-
|
|
353
|
-
program checks that each quote really exists in them. Invalid and duplicate quotes are removed;
|
|
354
|
-
multiple excerpts of one prompt do not count as independent support, and high-confidence traits
|
|
355
|
-
require support from multiple sessions. Quote matching does not prove the interpretation is correct.
|
|
356
|
-
Dimensions without verified quotes are marked as insufficiently supported.
|
|
138
|
+
The measured costs above include 21 label calls that new runs no longer make. New costs have not yet been measured.
|
|
357
139
|
|
|
358
|
-
|
|
359
|
-
notifications sent by the tools themselves, long templates repeated across sessions, and slash
|
|
360
|
-
commands of sessions that contain nothing else (scripts and tools) are excluded. Each run shows
|
|
361
|
-
how many were excluded.
|
|
140
|
+
Historical breakdown (before ability scoring; the label stage below has now been removed):
|
|
362
141
|
|
|
363
|
-
|
|
364
|
-
|
|
365
|
-
|
|
142
|
+
| Stage | Calls | What it does | Share of cost (Claude / Codex) |
|
|
143
|
+
| -------------- | -------------------- | -------------------------------------------------------- | ------------------------------ |
|
|
144
|
+
| map | 1 per chunk | reads the redacted prompts and writes cited observations | 52% / 57% |
|
|
145
|
+
| merge | 0–3 | condenses observations when there are too many | 10% / 9% |
|
|
146
|
+
| draft + review | 2 | writes the profile, then checks it against the evidence | 13% / 10% |
|
|
147
|
+
| label | 1 per chunk | counts how often each trait shows (the trait radar) | 24% / 24% |
|
|
148
|
+
| translation | 1 per extra language | translates the generated text | 1% / 1% |
|
|
366
149
|
|
|
367
|
-
|
|
150
|
+
### Scale reference
|
|
368
151
|
|
|
369
|
-
|
|
152
|
+
Current base call estimates for the same history at different `--since` ranges (from `--dry-run`, so
|
|
153
|
+
merges and retries come on top):
|
|
370
154
|
|
|
371
|
-
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
|
|
375
|
-
|
|
155
|
+
| Range | Prompts after cleaning | Conversations | Chunks analysed / total | Base calls |
|
|
156
|
+
| ------------------ | ---------------------- | ------------- | ----------------------- | ---------- |
|
|
157
|
+
| 2 days | 392 | 29 | 10 / 10 (no sampling) | 13 |
|
|
158
|
+
| 1 week | 1,278 | 195 | 16 / 34 | 19 |
|
|
159
|
+
| 2 weeks | 1,690 | 241 | 18 / 40 | 21 |
|
|
160
|
+
| 1 month | 3,132 | 448 | 23 / 60 | 26 |
|
|
161
|
+
| 2 months (default) | 4,038 | 611 | 21 / 68 | 24 |
|
|
162
|
+
| 6 months | 7,605 | 1,629 | 26 / 111 | 29 |
|
|
163
|
+
| All (7 months) | 7,884 | 1,686 | 25 / 115 | 28 |
|
|
376
164
|
|
|
377
|
-
|
|
165
|
+
As a rule of thumb, a sampled run costs about one call per analysed chunk plus a handful, and
|
|
166
|
+
time and cost follow the number of analysed chunks: a history of a few hundred prompts takes
|
|
167
|
+
about half of the measured run above, and anything past ~1,300 prompts costs about the same.
|
|
168
|
+
`--full` on the 6-month history would analyse 111 chunks, about five times the measured run.
|
|
378
169
|
|
|
379
|
-
|
|
170
|
+
## Sharing and privacy
|
|
380
171
|
|
|
381
|
-
|
|
172
|
+
**Reports include quotes, project names and exact dates by default.** Cards omit quotes, but
|
|
173
|
+
can still include project names and usage. Create a version with personal sections left out:
|
|
382
174
|
|
|
383
|
-
|
|
384
|
-
|
|
385
|
-
|
|
386
|
-
Evidence is expanded by default. Language and view switches work offline without JavaScript;
|
|
387
|
-
printing includes both views.
|
|
388
|
-
Quotes are verified against redacted prompts. Only the user's side of an interaction is
|
|
389
|
-
available; requests do not prove successful outcomes. Statistics and contextual observations
|
|
390
|
-
remain available alongside the narrative in their respective views. Older profiles remain renderable.
|
|
175
|
+
```bash
|
|
176
|
+
prompt-profile render prompt-profile-report/profile.json --exclude personal
|
|
177
|
+
```
|
|
391
178
|
|
|
392
|
-
|
|
393
|
-
|
|
394
|
-
|
|
395
|
-
|
|
396
|
-
|
|
397
|
-
|
|
179
|
+
For a profile in the report library, use the current profile path printed by the CLI.
|
|
180
|
+
The full report is kept; review the exported files before sharing them. Individual exclusions
|
|
181
|
+
such as `quotes`, `projects` and `usage` are also available.
|
|
182
|
+
|
|
183
|
+
- Source histories are read only. Recognized system messages and tool output are excluded;
|
|
184
|
+
pasted code and logs are collapsed before analysis.
|
|
185
|
+
- Supported secrets and personal information are redacted before sending, and report text is
|
|
186
|
+
checked again before writing. These rules can miss confidential prose or unrecognized secrets.
|
|
187
|
+
- Selected prompts, project folder names, intermediate observations and prompt statistics go to
|
|
188
|
+
the model service through your CLI. Token usage, plans and API price estimates stay local;
|
|
189
|
+
price estimates are not your subscription bill.
|
|
190
|
+
- Reports, dry-run exports and caches are local files that can contain personal information.
|
|
191
|
+
Clearing the cache does not delete reports or dry-run exports.
|
|
192
|
+
- prompt-profile has no telemetry. Reports work offline. The Chinese PNG card downloads a
|
|
193
|
+
verified font on first use; the borrowed CLI manages its own network traffic and retention.
|
|
194
|
+
|
|
195
|
+
See [privacy details and output checks](https://github.com/EricTsai83/prompt-profile/blob/main/docs/user-guide.md#privacy)
|
|
196
|
+
for source paths, permissions, CLI isolation and redaction controls.
|
|
398
197
|
|
|
399
|
-
|
|
400
|
-
analysis. `pnpm dev:report` renders an entirely fictional portrait for local inspection.
|
|
198
|
+
## FAQ
|
|
401
199
|
|
|
402
|
-
|
|
200
|
+
**Can I use an API key?** No; analysis uses your logged-in Claude Code or Codex CLI.
|
|
403
201
|
|
|
404
|
-
|
|
405
|
-
|
|
406
|
-
- **Folio** gives the portrait and evidence a quieter, connected reading rhythm.
|
|
202
|
+
**Why are some prompts excluded?** Short acknowledgements, resent duplicates, recognized tool
|
|
203
|
+
notifications, repeated templates and command-only sessions are filtered. Each run shows counts.
|
|
407
204
|
|
|
408
|
-
|
|
409
|
-
|
|
205
|
+
**Can it analyse other languages?** Yes. Report text supports English and Traditional Chinese;
|
|
206
|
+
source prompts can be in any language.
|
|
410
207
|
|
|
411
|
-
|
|
412
|
-
|
|
413
|
-
|
|
414
|
-
|
|
415
|
-
|
|
208
|
+
## Development
|
|
209
|
+
|
|
210
|
+
See [CONTRIBUTING.md](CONTRIBUTING.md) for setup, checks and releases, and
|
|
211
|
+
[AGENTS.md](AGENTS.md) for architecture and repository rules.
|
|
212
|
+
|
|
213
|
+
## License
|
|
214
|
+
|
|
215
|
+
MIT. Not affiliated with or endorsed by Anthropic or OpenAI.
|
|
216
|
+
|
|
217
|
+
### Narration voice
|
|
218
|
+
|
|
219
|
+
`--voice first|second` controls generated descriptive prose in every report language
|
|
220
|
+
and card. It is independent of `--look`. Second person (`you` / `你`) is the default; first
|
|
221
|
+
person uses `I` / `我`.
|
|
222
|
+
Quotes retain their original wording, and suggested sample prompts keep their intended speaker.
|
|
416
223
|
|
|
417
|
-
The
|
|
224
|
+
The selected voice is stored as `profile.voice` and shown in the report. First-person text is
|
|
225
|
+
an AI-authored draft; review it before presenting its claims as your own. Rerun the analysis
|
|
226
|
+
with another `--voice` to generate a coherent version; source-evidence map caches can be reused.
|
|
227
|
+
Changing voice does not add analysis stages or subscription calls beyond the normal run.
|
|
228
|
+
`render` preserves stored prose and voice without calling an LLM; it does not offer a voice
|
|
229
|
+
override or replace pronouns. Legacy profiles without voice metadata retain their original text
|
|
230
|
+
and do not receive an inferred voice label. Interface instructions still address the reader.
|