The REACH DSL

A .reach file is a domain-specific language. The domain is intent — what to reach into, over what timeframe, with what analysis goal. The language doesn't know about COM Interop, NuGet packages, or SQL client libraries. Those are implementation. The DSL knows about sources, qualifiers, and outputs.

You write the vocabulary. Claude compiles the implementation.

---

Syntax Shape

reach <source>
  <qualifier>  <value>
  <qualifier>  <value>
  <action>     <value>
  output       <format>

Multiple sources compose with +:

reach <source-a> + <source-b> + <source-c>
  since     <timeframe>
  analyze   <intent>
  output    <format>

Everything lowercase. Indentation two spaces. Quoted strings for values with spaces.

---

Source Vocabulary

Outlook

reach outlook.inbox
  where     subject contains "[PROJECT]"
  since     start-of-year
  group by  month
  output    cadence

reach outlook.calendar
  since     10-weeks-ago
  analyze   density
  group by  week
  output    summary

reach outlook.sent
  where     subject contains "[PROJECT]"
  since     start-of-year
  group by  month
  output    cadence

reach outlook.folder "Deleted Items"
  count     all
  output    summary

Git

reach git.commits
  since     10-weeks-ago
  analyze   after-hours
  benchmark 15%
  output    report

reach git.log
  since     start-of-month
  format    author + date + message
  output    table

Screen

reach screen
  save as   snapshot

reach screen.window "Chrome"
  save as   snapshot

reach screen.region 0 0 1920 1080
  save as   snapshot

Browser

reach browser.page "https://staging.internal/deployments"
  wait for  ".deployment-status"
  extract   text
  output    summary

reach browser.auth "https://internal-tool/login"
  field     "#username" "myuser"
  field     "#password" "mypass"
  submit    "button[type=submit]"
  then      browser.page "https://internal-tool/dashboard"
  extract   text
  output    summary

Database

reach db.query
  env       staging
  sql       "SELECT TOP 100 * FROM Equipment WHERE Status = 'Active'"
  output    json

reach db.table Equipment
  env       staging
  sample    20
  output    summary

Files

reach file.read "attachments/report.docx"
  extract   text
  output    plain

reach file.merge "attachments/*.txt"
  separator "---"
  output    "merged_context.txt"

reach file.watch "C:\reach\attachments"
  on new    signal "file-arrived"
  output    log

UI Automation

reach ui.window "OUTLOOK"
  find      name "Inbox"
  read      items
  output    list

HTTP

reach http.get "https://internal-api/deployment/status"
  output    json

---

Qualifier Vocabulary

Qualifier	Applies to	Meaning
where	outlook., db.	Filter condition
since	outlook., git., calendar	Time range start
until	outlook., git., calendar	Time range end
group by	any	Group results by dimension
analyze	any	Named analysis intent
benchmark	git.commits, any metric	Target value for comparison
wait for	browser.*	CSS selector to wait for
extract	browser., file.	What to pull — text, table, links
find	ui.*	Control to locate by name or type
env	db.*	Named environment (dev/staging/prod)
sample	db.*	Number of rows to sample
save as	screen.*	Output filename
on new	file.watch	Event trigger
merge	file.*	Combine multiple file outputs
count	any	Count items instead of listing them

---

Output Vocabulary

Output	What you get
summary	Claude narrates what it found
cadence	Frequency pattern over time — counts by period
report	Structured findings with interpretation
json	Raw JSON to stdout
table	Formatted table
list	Line-by-line items
plain	Plain text extraction
snapshot	Image file — Claude reads inline
log	Append to run log
signal	Write signal file for another reach

---

Analyze Vocabulary

Named analysis intents Claude understands:

Value	What Claude does
cadence	Groups by time period, identifies patterns, gaps, spikes
after-hours	Splits activity by time of day, calculates % outside core hours
density	Measures concentration — meetings per week, emails per day
rpm	Reads combined signal across sources, calls green/yellow/red/all-red
velocity	Rate of output over time
sentiment	Tone pattern across communications
recovery	What it would take to normalize current state
summary	Plain language synthesis

---

Timeframe Vocabulary

Value	Meaning
today	Current day
this-week	Monday to now
last-week	Previous full week
start-of-month	First day of current month
start-of-year	January 1
10-weeks-ago	N weeks before today
30-days-ago	N days before today
"2026-01-01"	Explicit date

---

Composition — The Most Powerful Operator

+ between sources is the most powerful operator in the language. Each source reaches independently; Claude holds all results in context and synthesizes across them.

reach outlook.inbox + outlook.calendar + git.commits
  since     10-weeks-ago
  analyze   rpm
  benchmark after-hours 15%
  output    report

This is the sprint cadence review in one block. Claude reaches into all three sources, reads the combined signal, calls the RPM state, flags the after-hours against the benchmark, and delivers a narrative report.

---

The Pause Layer — Prompts and Gates

REACH has three modes. The pause layer is what separates reads from writes:

READ    →  runs freely, no prompt
ACT     →  low risk, may run freely
PAUSE   →  write gate — Windows UI surfaces, human approves

prompt.confirm

reach outlook.draft
  to        "[email protected]"
  subject   "Sprint Summary — {date}"
  body      from git.commits since today
  then      prompt.confirm
    title   "Send this email?"
    on yes  outlook.send
    on no   outlook.save-draft

prompt.preview

reach outlook.draft
  to      "[email protected]"
  subject "Daily Summary — {date}"
  body    from git.commits + outlook.inbox since today analyze summary
  then    prompt.preview
    title   "Email draft ready"
    actions send | edit | save-draft | discard
    on send       outlook.send
    on save-draft outlook.save

prompt.form

reach timesystem.timesheet
  then  prompt.form
    title  "Review timesheet entries"
    fields
      request-id   text      "Request ID"
      start-time   time      "Start"
      end-time     time      "End"
      description  textarea  "Description"
    on submit  timesystem.post
    on cancel  outlook.save-draft

The prompt layer is the conscience of the automation. Reads run freely. Writes pause. no-prompt is an explicit override — never a default.

---

Named Reaches — Reusable Patterns

# sprint-review.reach

name        sprint-cadence-review
description Read sprint health across email, calendar, and commits

reach outlook.inbox + outlook.calendar + git.commits
  since     10-weeks-ago
  analyze   rpm
  benchmark after-hours 15%
  output    report

Run it:

dotnet reach sprint-review.reach

Share it. Fork it. Someone changes the project filter and runs it against their data. The pattern travels as a file.

---

The Compilation Model

You write          →  sprint-review.reach
Claude reads DSL   →  understands intent per source block
Claude generates   →  typed C# per source (NuGet chosen per block)
dotnet run         →  executes each reach
prompt surfaces    →  if write action — Windows UI appears, human decides
results flow back  →  Claude synthesizes across all sources
artifact written   →  sprint-review.reach-artifact

You never see the NuGet package names, the COM Interop calls, the Playwright instantiation, or the WinForms boilerplate. Just intent. Just results. Just the right pause at the right moment.

---

Design Principles

Reads like intent, not code — Anyone can read a .reach file and understand what it does. No training required.

Sources are the vocabulary, not the implementation — outlook.inbox is a concept. How Claude reaches it is Claude's decision, not yours.

Composition is first-class — + between sources is the most powerful operator. The sprint review, the timesheet draft, the deployment check — all compositions of simple sources.

Artifacts outlast sessions — Every meaningful reach produces a .reach-artifact. The finding is permanent even though the script is ephemeral.

Benchmarks make findings actionable — benchmark after-hours 15% turns a number into a verdict. Claude doesn't just report 30% — it calls it against the target.

Writes pause by default — The prompt layer is not optional. It is the design.