B
BarzinL/barzins-cc-improvements
Token-efficiency, better outputs, and more autonomy in Claude Code.
★ 2·updated 2026-11-19
View on GitHub ↗Link checked 2026-11-17 · HTTP 200 · opens the vendor's own site
- Star count
- 2 ★ · BarzinL/barzins-cc-improvements · read from the GitHub search API · (checked 2026-11-19 20:00:48)
- Source
- read directly from its GitHub page · type: open-source project
- Harvest history
- first seen 2026-11-16 · last seen 2026-11-20 · seen 5× · listed by 1 sources
- Score
- score 103 · 2 ★ · 1 sources · repo
FAQ (answers come from our records)
Who is it for?
We filed it under these tags: Open Source. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
repository
BarzinL/barzins-cc-improvements;first catalogued 2026-11-16;last seen 2026-11-20。There is no official certification or quality rating here — we record sources and dates only.README excerpt (English, untranslated) · Read the full README on GitHub →
# Barzin's Claude Code Improvements
Excerpted from the repository README (abridged)
Skills, hooks, and standing rules that make Claude Code verify things instead of guessing.
Each one was written after a specific failure in real production work with Opus 4.8, Opus 5
or Opus 5.5, and each says which failure. The Contents table names the model each one was
written against.
## What's new
Added in the Opus 5.5 era. Older additions and changes are in CHANGELOG.md.
| Date | Directory | What it does |
| 2026-11-15 | hook-trial/ | Measures four hooks (claim gate, plan-reconciliation, placeholder-disclosure, task-grounding) on your own work. An optional trial file switches all four off in a random half of sessions, chosen from the session id; compare_arms.py then compares the two halves from the transcripts: unsupported completion claims that reached the user (judged by the claim gate itself), hook firings, output tokens per turn and follow-up messages, each with a bootstrap interval over sessions. Off by default; no trial has been run yet. |
| 2026-11-14 | task-grounding/ | A UserPromptSubmit hook that recognises performance, visual, scientific, multi-step and reader work from the prompt's words and adds one short directive per category, once per session: measure before and after, render and look, read sources before coding, write a plan file, and for a course, guide or long report, state the reading and the reader and show one piece before building the rest (added 2026-11-15). Keyword-matched with narrow patterns, so it misses some prompts and still misfires on others; not yet run live or measured. |
| 2026-11-14 | reference-store/ | A format for a project's compiled domain reference (docs/reference/), in which every claim carries its source, the date it was read, and whether it was read at the source, taken from a secondary source, or inferred; and a checker that refuses untagged claims and finds the one entry a task needs. |
| 2026-11-14 | theory-formation-discipline/verification-claim-gate/ | The claim gate, revised for Opus 5.5, now ties each claim to the check that would support it: a test claim needs a test command that did not fail, and a whole-file claim is compared with the reads. Reads count across the session and include cat, head, tail and sed -n, parsed per pipeline; whole-repository claims are flagged; "read through " and "read all four " are checked against the reads; speed and memory claims need a timing command. Adds replay.py to run a gate offline over your own transcripts. |
| 2026-11-14 | plan-reconciliation/ | A Stop hook: when the session has a checklist at /plan.md, a final report that leaves open items unmentioned is paused once, listing them to be reported as done, deferred or dropped. |
| 2026-11-14 | placeholder-disclosure/ | A Stop hook: placeholder code written this turn (todo!(), NotImplementedError, pass-only bodies, TODO/FIXME comments) that the final message does not disclose is listed by file and line, once. Includes a draft "blocked" CLAUDE.md rule, not installed. |
| 2026-11-14 | pasted-text-guard/ | Finds invisible and bidirectional-control characters in text while keeping the joiners Persian, Arabic, Indic scripts and emoji need. A library and command line, not a hook: Claude Code 2.1.277+ already strips these before a hook would run. The README records that evidence and that pasted text goes unmarked when editor telemetry is off. |
Each one was written after a specific failure in real production work with Opus 4.8, Opus 5
or Opus 5.5, and each says which failure. The Contents table names the model each one was
written against.
## What's new
Added in the Opus 5.5 era. Older additions and changes are in CHANGELOG.md.
| Date | Directory | What it does |
| 2026-11-15 | hook-trial/ | Measures four hooks (claim gate, plan-reconciliation, placeholder-disclosure, task-grounding) on your own work. An optional trial file switches all four off in a random half of sessions, chosen from the session id; compare_arms.py then compares the two halves from the transcripts: unsupported completion claims that reached the user (judged by the claim gate itself), hook firings, output tokens per turn and follow-up messages, each with a bootstrap interval over sessions. Off by default; no trial has been run yet. |
| 2026-11-14 | task-grounding/ | A UserPromptSubmit hook that recognises performance, visual, scientific, multi-step and reader work from the prompt's words and adds one short directive per category, once per session: measure before and after, render and look, read sources before coding, write a plan file, and for a course, guide or long report, state the reading and the reader and show one piece before building the rest (added 2026-11-15). Keyword-matched with narrow patterns, so it misses some prompts and still misfires on others; not yet run live or measured. |
| 2026-11-14 | reference-store/ | A format for a project's compiled domain reference (docs/reference/), in which every claim carries its source, the date it was read, and whether it was read at the source, taken from a secondary source, or inferred; and a checker that refuses untagged claims and finds the one entry a task needs. |
| 2026-11-14 | theory-formation-discipline/verification-claim-gate/ | The claim gate, revised for Opus 5.5, now ties each claim to the check that would support it: a test claim needs a test command that did not fail, and a whole-file claim is compared with the reads. Reads count across the session and include cat, head, tail and sed -n, parsed per pipeline; whole-repository claims are flagged; "read through " and "read all four " are checked against the reads; speed and memory claims need a timing command. Adds replay.py to run a gate offline over your own transcripts. |
| 2026-11-14 | plan-reconciliation/ | A Stop hook: when the session has a checklist at /plan.md, a final report that leaves open items unmentioned is paused once, listing them to be reported as done, deferred or dropped. |
| 2026-11-14 | placeholder-disclosure/ | A Stop hook: placeholder code written this turn (todo!(), NotImplementedError, pass-only bodies, TODO/FIXME comments) that the final message does not disclose is listed by file and line, once. Includes a draft "blocked" CLAUDE.md rule, not installed. |
| 2026-11-14 | pasted-text-guard/ | Finds invisible and bidirectional-control characters in text while keeping the joiners Persian, Arabic, Indic scripts and emoji need. A library and command line, not a hook: Claude Code 2.1.277+ already strips these before a hook would run. The README records that evidence and that pasted text goes unmarked when editor telemetry is off. |
Listed badge (optional)
Copy this line into the README if you want the listing mark. Optional, and it does not affect being listed.
[](https://anclum.com/en/s/barzinl-barzins-cc-improvements/)
- Source
- github · https://github.com/BarzinL/barzins-cc-improvements
- Browse by field
- Open source
- Traction
- 2 ★
- added
- today
More entries
Filed under
Every tag opens the full list for that filter. Counts are the live catalogue figures.