Repository navigation
Releases
Releases · hayfever/succinct
v1.0.0
First release of succinct: one output style for coding agents. It puts the answer first and the next action last, strips AI writing tells, and writes technical text in plain Simplified Technical English.
The skill
- The skill is a 164-line core,
skills/succinct/SKILL.md, at 2,080 o200k tokens. Two reference files load only on demand:references/patterns.mdhas the full 25-tell catalog, andreferences/plain-english.mdhas the STE detail and replacement lists. The core stands alone when the references do not resolve. - Part 1 has ten shape rules for an ADHD reader: the answer first, numbered steps, one concrete next action last, no tangents, state restated every turn, specific time estimates, visible wins, matter-of-fact errors, lists capped at five items, and no preamble or recap.
- Every response obeys three more rules. A compressed rule must keep its full meaning. No dash is a connector in prose (em dash, en dash, " -- ", " - "). No response ends with a pleasantry or a stock offer. An explicit output contract ("return only X", a word limit, an exact format) overrides the skill habits.
- Part 2 removes the five strongest AI writing tells and never invents a fact. The tell inventory comes from Wikipedia: Signs of AI writing.
- Part 3 summarizes the writing rules of ASD-STE100 Issue 9. It does not reproduce the dictionary.
Harnesses
- Install routes for Claude Code, Codex, OpenCode 2, Pi, and OMP, plus
npx skills addfor any skills CLI agent and an always-on snippet for other agents.INSTALL.mdhas each route. - One activation model in all five harnesses.
/succinctturns the mode on for the current session ($succinctin Codex)./succinct oncreates a.succinct-alwaysflag file and/succinct offdeletes it. Every session starts from the flag. "stop succinct mode" or "normal mode" turns the mode off for the current session only. In Pi and OMP, the footer shows● SUCCINCT ON, andpi --succinctstarts with the mode on. - Claude Code and Codex read the flag in a
SessionStarthook and set it in a prompt hook (hooks/always-on.sh,hooks/toggle.sh, their PowerShell versions, andhooks/codex-hooks.jsonfor Codex). No hook needs Node.
Tooling
scripts/build_adapters.pygenerates the skill metadata version and every manifest version frompackage.json.--checkfails on drift.scripts/validate-package.pychecks frontmatter, versions, paths, rule numbering, and drift.- CI runs the unit tests on Linux, macOS, and Windows, and load checks for Claude Code, Codex, OpenCode, OMP, and Pi with no model request.
release.ymldrafts a release only after the full matrix passes. - Paired evals: 17 cases, 3 trials, a blind judge, and a relative release gate. The shipped core scored 4.75 against a 4.48 baseline with no blocking finding. See
evals/RESULTS.md.
What's Changed
- docs: same-line status placement for OMP by @hayfever in #1
- v1.1.0: persisted /blunt preference by @hayfever in #2
- Rename blunt to succinct (2.0.0) by @hayfever in #4
- Production-readiness fixes by @hayfever in #5
- Fix the eval harness, tighten output rules, and gate against baseline by @hayfever in #7
- Generate harness adapters and gate drift in CI by @hayfever in #8
- Slim the core skill to 2,055 tokens with reference catalogs by @hayfever in #9
- Polish the README and docs by @hayfever in #6
- Put the next action last by @hayfever in #10
New Contributors
Full Changelog: https://github.com/hayfever/succinct/commits/v1.0.0