terse-sharp: Command for Claude Code

🚫 HARD GATE — findings are measured to the character, never impressionistic, and never leak another project's content.

$ARGUMENTS — a number of weeks to scan. Absent or unparseable → 1 week. Nothing else takes input from the user; do not ask, do not confirm the window.

The six goals this command exists to serve. Every phase feeds at least one; every row logged in M9 is tagged with the one it serves; and M11 proves all six were mined or states what was checked and why that goal came back clean. They are not ranked by taste — they are ranked by what a failure costs:

[accuracy], [correctness] and [speed] are co-primary, and all three outrank [cost]. An intervention costs a whole turn at the corpus turn p50, plus the tokens of the re-issued work, plus the context already spent on the wrong path — so a payload row of equal token size is worth less than an intervention row. And a round trip costs the corpus model gap (tool_result → next tool_use, measured p50 6 097 ms on 35 967 gaps in the 1-week window of 2026-08-26) plus the tool's own p50 — so a row that deletes one call per task is worth more than a row that shortens one response, because the wall clock the user waits through is dominated by how many times the loop turns, not by how wide each answer is. Rank accordingly in M9.

The four [speed] metrics, and they are the run's headline numbers. Every run reports all four in M11 against the previous run, and a run that reports none of them has not measured [speed] and says so as a degraded run:

Ten gates that outrank everything else in this command:

1. A finding without a number is not a finding. "Reads feel wasteful" is banned. "36 Read calls in one session, 214 KB of tool results, ~53 500 tokens, against 10 search_text calls" is a finding. Every row carries a count, a byte/token/millisecond figure and the corpus it came from.
2. Nothing is too small to count — but a trim is only real once a tokenizer says so. This command's reason to exist is the long tail: a two-space column separator, a repeated (truncated=false), an absolute path, one redundant word in a tool [Description] re-sent on every request of every session. A multi-space run costs exactly 1 token wherever it appears, so a 4-column table pays 3 tokens per row for alignment nobody reads — over 40 000 rows that is 120 000 tokens, and it outranks most new tools. "Too small to log" is a banned phrase here; small findings are aggregated into one row per family, never dropped (M9.4). The mirror rule is equally binding: a trim that measures zero is not a finding either. Blank lines, trailing whitespace and abbreviation each measured ≤0.1% and 19-of-20 exactly zero on real payloads (M3.2). Proposing one of those is the same defect as failing to notice the padding — both spend the reader's attention on a number that is not there.
3. Never copy content out of another project's transcript. These sessions are from a private employer codebase (cTraderDev, ctd-worktrees-*, cTraderAutomateApi, …) and IMPROVEMENTS.md is pushed to a public repository. A row may name: tool names, call counts, byte sizes, token estimates, millisecond timings, error codes, and a project's directory slug. A row may never contain: source code, file paths inside those repos, type or member names, ticket text, user prompts, or a quoted tool result. When in doubt, state the shape (a 40 KB whole-file read of a test file) not the value. The measurement script enforces a mechanical version of this rule and you enforce the rest by hand in M10.2.
4. Character economy is a claim about tokens, so state both. Chars are what the script can count; tokens are what is paid. The measured ratio over real TerseSharp responses is 4.18 chars/token, which is why section A estimates at chars // 4 — and why an estimate is all it is. Prefer trims that delete whole tokens (a separator, a padded column, a repeated line prefix, a whole field) over trims that shorten one: deleting characters from inside a word usually saves nothing at all, because a common word is already a single token.
5. Every claim names the instrument that verified it. Four are admissible, in this order: (a) the corpus — a count from the single-pass scan over the real transcripts; (b) a tokenizer experiment — an encode-before/encode-after on a real payload, in both o200k_base and cl100k_base, labelled directional because neither is Claude's; (c) a primary source — a vendor document or a paper, fetched and quoted, never a summary of a summary; (d) the workspace — analyze, search_regex, get_symbol_source, build or run_tests over TerseSharp's own source, which is the only instrument that can confirm a pattern the transcripts show being written rather than being read. Where an outside claim contradicts the corpus, the corpus wins and the row says so. A claim that survived none of the four is written UNVERIFIED in the report and is not allowed into IMPROVEMENTS.md at all. This gate exists because a previous run of this command shipped a trim ledger whose three headline classes were each worth ~0.1%, sourced from plausible reasoning that nobody encoded.
6. The agent's own output is corpus too, and a run that skipped it is incomplete. Section A sees what the agent read; section C sees what it wrote (tool_use.input on the edit family) and the turns the user spent steering it (user records carrying text rather than a tool_result). All three sections come out of one pass now, so there is no longer a phase that can be quietly dropped when the run is long — a section that prints zeros is a defect in the script, not an absence of data, and M11 says which it was.
7. A pattern counted in emitted text is HEURISTIC until instrument (d) confirms it. The emitted-code section is a regex over the characters the agent sent, with no compilation behind it. It ranks candidates; it never proves one. A [modern] or [perf] row states which analyze / search_regex / get_symbol_source call confirmed the class on real source, or it is not written. And a class this repository deliberately leaves off — an .editorconfig severity, the app-wide conventions clause in CLAUDE.md — is not a finding; check before logging, because proposing that the codebase fight its own configuration is the same defect as a number nobody encoded.
8. A call deleted beats a response shortened, and the run must say which it did. The prime directive is "save tokens, increase speed" — conjunction, not disjunction — and M4.1's baseline says the second half is where the hours are. A format trim that saves 2 000 tokens and changes no call count saves zero seconds. So every [speed] row states its saving in hours or milliseconds per window, and names which of the five levers it pulls: