signal

SIGNAL

Less noise, same signal.

Token-efficient answers (templates · symbols · checkpoints) and one-shot git workflows — Agent Skills for Claude Code, Cursor, Gemini CLI, Codex, and more.

Install · Before / After · Tiers · Commands · Benchmark · Evidence · Hosts · Layout

---

Repo	github.com/mattbaconz/signal
Chat	Discord
Protocol	signal/SKILL.md · Changelog

Token savings (typical ranges, not guarantees): tier design targets are ~65% (SIGNAL-1), ~80% (SIGNAL-2), and ~90%+ on long sessions when SIGNAL-3 checkpoints dominate output cost vs verbose replies. Replacing a long transcript with one checkpoint line reached ~94% fewer tokens than verbatim history in one representative run (~18×; rough estimate, ~4 chars/token). Single-turn Gemini runs with matched project GEMINI.md show lower tokens.total and much shorter replies vs verbose control — see Evidence and Benchmark snapshot. Per-turn wins vary by task; hosts can add large fixed overhead (e.g. session reset). See docs/token-metrics.md (prompt vs output vs history) and the table below.

How it fits together

Where tokens go (three buckets to measure separately):

flowchart LR
  subgraph input [Prompt]
    P1[System_and_rules]
    P2["GEMINI.md / CLAUDE.md"]
    P3[Skills_and_user]
  end
  subgraph gen [Generation]
    G1[Assistant_reply]
  end
  subgraph hist [History]
    H1[Prior_turns]
    H2[CKPT_atom]
  end
  input --> gen
  hist --> gen

Tiers (activate the signal skill):

flowchart LR
  s1["/signal SIGNAL-1"]
  s2["/signal2 + BOOT aliases delta"]
  s3["/signal3 + auto CKPT"]
  s1 --> s2 --> s3

Host integrations in this repo: skills (npx skills add), Gemini extension + Claude plugin, repo-local rules (.cursor, .windsurf, .clinerules, Copilot), optional Claude hooks and Codex SessionStart.

Benchmark snapshot

Single-turn Gemini CLI runs (gemini-3.1-pro-preview, same prompt.txt); reproduce. API 429 can require retry.

Pair	What it shows	Highlights
EqualContext (-Pair EqualContext)	Matched project GEMINI.md size; fair prompt comparison	~flat prompt (+0.3%); tokens.total max −1.9% vs verbose; ~74% fewer reply characters
Default	No GEMINI.md vs SIGNAL project file	~88% fewer reply chars; higher prompt/total on SIGNAL arm (input tax) — see docs/token-metrics.md

Full tables: Evidence · JSON: results_chess_equal_compare.json, results_chess_compare.json.

---

Before / after