Agent Skills: Context Efficiency: Filter at the Source

Context Efficiency: Filter at the Source

Every token of noise in context is a token the model spends navigating instead of reasoning. Spend a few tokens on a precise query; save many on the backend.

Decision tree

Does the tool support structured output? (--json, -o json, --format json)
├─ Yes → select fields with jg when available; if `command -v jg` fails, use tool-native selectors, python3 -c, or jq instead of retrying
└─ No → pre-filter with path scopes, grep, head, API params, or SQL LIMIT

Structured-output patterns

# Select one field or path
tool -o json | jg 'precise.selector'


# Filter to the few fields needed
tool -o json | jg 'items matching condition -> id,name,status'


# Transform when selection is not enough
tool -o json | python3 -c 'import json,sys; print(json.load(sys.stdin)[0]["field"])'

jq/python fallback patterns

Use these only when jg cannot express the needed transformation.

# python3 -c patterns
# Count by state
tool -o json | python3 -c "
import json, sys; d = json.load(sys.stdin)
from collections import Counter; print(Counter(x['state'] for x in d))
"

# Find matching entity
tool -o json | python3 -c "
import json, sys; d = json.load(sys.stdin)
print(next(x['entity_id'] for x in d if 'couch' in x['attributes'].get('friendly_name','').lower()))
"

Bounded command output

Before running commands that can dump a tree, log, search result, build output, or API collection, add a path scope, --limit, head, structured selector, or SQL LIMIT. Do not raise output-token caps to compensate for an unbounded command. If output truncates, rerun a narrower command instead of scanning the dump.

Oversized tool results

If a read, MCP call, or CLI returns "output too large" or saves overflow to a file, do not retry the same broad request. Recover with a narrower query:

1. Re-run the source tool with tighter fields, date ranges, IDs, channel/project filters, or lower limits.
2. If the tool saved an overflow artifact, search or read only the relevant ranges from that saved file.
3. Prefer summaries or metadata endpoints before transcripts, full logs, Slack exports, calendar event lists, or issue dumps.

# ❌ broad transcript/log dump
tool get-transcript --id "$id"

# ✅ metadata or filtered content first
tool get-notes --id "$id"
tool search --query '"exact error" after:2026-01-01' --limit 5

Tool-specific examples

Home Assistant (hass-cli)

hass-cli -o json state list 'light.*' | jg 'entity_id,state,attributes.friendly_name'
hass-cli -o json area list | jg 'area_id,name'
hass-cli -o json device list | jg 'devices in kitchen -> id,name,area_id'

GitHub CLI

gh pr checks 42 --json name,state | jg 'failed check names'
gh issue list --json number,title,labels --limit 30 | jg 'issues with bug label'
gh api repos/:owner/:repo/pulls --jq '.[].head.ref'

Nix

nix eval .#packages --json | jg 'package names'

Database

-- Always: WHERE + LIMIT over SELECT *
SELECT entity_id, state FROM states WHERE domain = 'light' ORDER BY last_changed DESC LIMIT 20;

Anti-patterns

# ❌ dumps hundreds of entities to find one
hass-cli state list

# ✅ returns exactly what you need
hass-cli -o json state list 'light.*' | jg 'entity_id,state,attributes.friendly_name matching desk'

# ❌ loads full PR list into context
gh pr list

# ✅ targeted
gh pr list --json number,title --limit 30 | jg 'titles matching fix'