komilovsg/f-mode

F-mode — filter-first mode

Principle: raw data never enters the context; only the result does. Move the logic to the data (a shell pipeline, a short script, a subagent) and bring back a compact summary. Every token of a log dump you read is a token you can't spend on reasoning — and long noisy context makes reasoning worse, not just more expensive.

1. Filter and aggregate before reading

Instead of	Do
cat app.log	`grep -c ERROR app.log; grep -E 'ERROR
reading a 5 MB JSON	jq '.items | length', jq -r '.items[] | [.id,.status] | @tsv' | sort | uniq -c
SELECT * FROM orders	SELECT status, count(*) FROM orders GROUP BY 1
git log	git log --oneline -20, git log --format=%s origin/main..HEAD
scrolling a file	grep -n 'def handle' file.py, then read with offset/limit

2. Filtering must never hide failure

A truncated view that looks green is worse than no view.

• Keep the exit code: set -o pipefail, or cmd > out.txt 2>&1; echo "exit=$?"; tail -20 out.txt. Piping straight into tail loses the command's exit code.
• Grep failure markers in addition to tail: grep -cE 'FAIL|ERROR|Traceback|panic' out.txt.
• Print counts with denominators: passed 41/42, never just "OK".

pytest -q --tb=short > /tmp/t.out 2>&1; echo "exit=$?"
grep -cE 'FAILED|ERROR' /tmp/t.out; tail -15 /tmp/t.out

3. Many similar checks → one script, one table

Ten endpoints, twenty files, thirty rows to validate: write one script that loops and prints a compact table (name | status | detail) plus totals. Not ten tool calls with ten full outputs.

for u in / /api/health /api/items; do
  printf "%-14s %s\n" "$u" "$(curl -s -o /dev/null -w '%{http_code}' "localhost:8000$u")"
done

4. Check the summary for meaning, not shape

Valid-looking output can still be wrong. Compare against an expected value or a sanity total: counts add up to the input size, a known row is present, the number of files matches ls | wc -l. If a summary can't be cross-checked, say so.

5. Tests

Quiet mode, short tracebacks, tail of the failure section:

• pytest -q --tb=short
• vitest run --reporter=dot, jest --silent
• go test ./... 2>&1 | grep -E '^(FAIL|ok|---)'
• cargo test -q

6. Browser checks

Evaluate a script that returns compact JSON ({title, errors: n, buttons: n, overflowX: bool}) instead of full DOM snapshots. Take a snapshot only when the DOM structure itself is the question. Screenshots for visual checks, not for reading text.

7. Wide investigation → subagent

Big searches, log forensics, database dumps: delegate to a subagent that returns conclusions with file:line references. Its result is still a claim — verify the key points before acting on them.

8. Secrets and personal data

Never print secrets or real PII in summaries or diagnostics. Mask (postgres://***) or count them (emails found: 12). Mask connection strings and keys inside caught exception messages too.

9. Reading files

• grep -n first, then read only the relevant range with offset/limit.
• Source code you are about to edit is read fully — partial reads cause broken edits.
• Don't re-read a file right after editing it to "verify"; the edit tool already errors on failure.

10. Stop rule

The same fix failed twice → stop. Find the root cause from the exact error line, or ask. No third blind retry.

11. Context hygiene

• Clear context between unrelated tasks; compact long sessions.
• Disable plugins, skills and MCP servers you don't use in this project: each one costs prompt tokens in every request.
• Judge savings by real usage numbers (e.g. /cost), not by claimed percentages.