business-research
Use when collecting evidence for the business value of an AI skill or workflow into a provenance-tagged research package.
- License 许可
- MIT
- 所属仓库
- tony/skills
- 最近更新
- 2026-11-08
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- business-research
- metadata
- { "source": "plugins/business/skills/research/SKILL.md", "argument-hint": "[skill or workflow to measure] [date range, e.g. 2026-04-01..2026-06-30]" }
- description
- Use when collecting evidence for the business value of an AI skill or workflow into a provenance-tagged research package.
- allowed-tools
- [ "Bash", "Read", "Write", "Grep", "AskUserQuestion", "EnterPlanMode", "ExitPlanMode" ]
- disable-model-invocation
- true
02 资源文件 7
- 参考文档agents/openai.yaml292 B
- 脚本references/business.py79.1 KB
- 参考文档references/instruments.md8.3 KB
- 参考文档references/interim-format.md14.2 KB
- 参考文档references/measurement.md6.9 KB
- 参考文档references/prose-rules.md4.5 KB
- 参考文档references/provenance.md4.3 KB
03 SKILL.md 正文
Business Research
Collect the data behind a business-value claim for an AI skill or agentic workflow (productivity, time saved, quality, capacity, delivery outcomes) into the interim run package the report skills render from.
Read these first; they bind every step:
references/interim-format.md: the run package layout, where runs live, and the value units (never money).references/provenance.md: the four tags and the anti-inflation rules.references/instruments.md: per-instrument probes, collection discipline, the pull-request query, and the targets to collect.
Load references/measurement.md only for the counterfactual
ladder and the statistics discipline, which owns the n<5 range rule.
Intervals come from references/business.py stats (absolute path).
User arguments: $ARGUMENTS
Procedure
1. Scope and orchestration plan
Enter plan mode before the first edit. In Claude Code call
EnterPlanMode; in Cursor, Codex, or Gemini use /plan or
Shift+Tab. Where plan mode is unavailable, present the plan as text.
Present the plan (shape below) and ask where to write the run,
defaulting per interim-format.md § Where runs live. Write nothing,
including index.md, until the user approves. Exit plan mode, then
open the run at the approved root with business.py init, per
interim-format.md § Where runs live. Pass --date when the request
names a dated run path; if init refuses, follow the same section.
Non-interactive (no question tool, headless): print the plan with the default root and stop. A request that itself names the root approves the plan; record the root in the run README and continue.
Reply in this shape (reply form per interim-format.md § Finished
run), with no greeting, naming roots as the user would type them:
Plan: review-skill, window 2026-04-01..2026-06-30. Nothing is written yet.
## Measure
- Cycle time and first-review latency per PR; rework signals.
- Adoption and per-task timing only from exports you provide.
## Instruments
- git and gh to probe; trackers and timing logs only from exports you provide.
## Output
- 2026-07-14/business/review-skill under ~/Documents, the default root.
- Questions: confirm the root and the repo paths.
The - Questions: bullet is the last line: no closing paragraph that
restates it. One fact per bullet. Report what you will probe, not
what the current directory happens to hold or what you have not yet
probed.
2. Instrument discovery
Probe each instrument per instruments.md; never assume one exists.
Record every instrument as available or unavailable in the run
README (for gh auth status, only that it succeeded). Data an unavailable instrument would have provided is
unknown, never fabricated and never silently skipped, unless the
user supplies an assumption, which is registered as ESTIMATED with
rationale and owner (provenance.md gives its fields).
3. Collect
For each available instrument, run pinned-window queries per
instruments.md and snapshot raw output into raw/ before deriving
anything. Collect the targets listed there; record the rest as
unknown. Derive pull-request tables with business.py collect-prs,
piping the query into it with --output so reviewer and author logins
never reach raw/, rather than rewriting the query or the filter. Write
every sources.md entry with business.py source <run> <entry> <file> --query <command> --tool-version <version> (absolute path) right after the
command produces <file>: the script writes the output path, the clock
time, and the sha256, so none is typed by hand. Record every collect-prs
warning, including "rows lack" ones, in that entry and in the README.
4. Write the package
Fill the package init started, per interim-format.md: run README,
source manifest with verbatim queries, assumptions register, raw
snapshots, per-topic measurements, and findings.md. Tag every figure
per provenance.md; list every unknown with what data would resolve it.
List every organization, repository, team, and workflow name the run
touched on the README Names: line, and every individual on the
People: line (never on both); the scans depend on both.
Compute medians, quartiles, and intervals with business.py stats
(--csv <file> --column <a-b> for a CSV column); copy its n, rows,
and dropped_blank beside each statistic. Give each MEASURED
figure computed from a raw/ CSV a check: line. Run
business.py gate <run> and fix every problem before reporting.
Write prose per references/prose-rules.md.
Output
Hero success form: ✓ Run written: <date>/business/<workflow>, window <start>..<end>; a stopped run follows interim-format.md §
Rendering a report. Then exactly these sections:
## Scope: what was measured and the pinned window.## Instruments: available vs unavailable, and what each unavailable one leaves unknown.## Collected: per instrument, what the data shows and what landed inraw/andmeasurements/; one fact per bullet, one sentence per bullet, with no clause defending a choice or restating the hero.## Package: the run location, the count of registered assumptions (0 is valid; a measured count is not an assumption), and the count of open unknowns; the unknowns themselves stay infindings.md.
End with an ask-user-choice panel: generate a report (ask which
tier), collect more, or stop.
Panel rules as in interim-format.md § Finished run.
Reply in this shape:
✓ Run written: 2026-07-14/business/review-skill, window 2026-04-01..2026-06-30
## Scope
- Review latency and rework rate for the review skill, 41 PRs.
## Instruments
- Available: git history, gh pull-request data.
- Unavailable: timing logs; per-use verification time stays unknown.
## Collected
- raw/ holds 2 snapshots; measurements/ holds latency and rework.
## Package
- Run at 2026-07-14/business/review-skill, 3 assumptions, 1 open unknown.
Next, you can:
- Generate a report (leadership, org-wide, internal or public case study).
- Collect more data.
- Stop here.
The last list is the non-interactive form; with a question tool it is the panel instead.
Portability notes
ask-user-choice— present the listed options and wait for the user to pick one. Hosts with a structured multiple-choice tool (Claude Code'sAskUserQuestion) should use it; otherwise print a numbered list and wait for a numbered reply. Never proceed on an assumed answer.$ARGUMENTS— the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.- Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.
04 同仓库更多 Skills
serene-bliss
Use when asked to make a CLI --help, docs page, error message, or tooling UX more serene or calmer for developer experience; returns text, no edits.
review
Weave code review: run independent adversarial reviews in parallel, then synthesize findings
refine
Use when asked to refine a draft, iterate on code until cleaner, or polish a README through repeated critique and improvement.
prompt
Weave prompt: compare independent adversarial implementations, then adopt the single best one
plan
Weave planning: compare independent adversarial plans, then synthesize the best one
fix-review
Fix weave review findings: validate, add test coverage, fix, and commit each as atomic changes