Built with bassclef
What bassclef actually did on this site, counted from files. Every number here comes from a chronicle, a state marker, or a git-log commit.
What bassclef actually did on this site, counted from files. Every number here comes from a real file — luminaries in chronicles, skills in git log, gates as markers on disk. Bassclef vocabulary used on this page is glossed in the FAQ at the bottom.Last refreshed .
Headline metrics 30-day window
The top row answers “is the discipline actually running?” Every count comes from a file in this repo — no vanity numbers.
Temperance per PR
85.0%
68 of 80 PRs merged this month had a temperance marker on their branch.
Chronicles written
61
One narrative log per substantive session; 97 lifetime.
Luminaries dispatched
41 distinct
alan-cooper, linus-torvalds, michael-feathers top the frequency.
Skills dispatched
41 distinct
/luminary most frequent.
Markers written
339
Across 21 gate types; audit trail per commit.
Data freshness
fresh
Generated 10/2/2026, 6:40:29 PM.
Cost signal from Claude Code transcripts · 2026-09-02 onwards
The cache-reuse ratio below. Raw token counts are withheld — see the attribution note.
Approximate turns
2,742
Sum of chronicle turns field across 17 chronicles this window. 44 more chronicles missing the field — real count is higher.
Cache reuse ratio
98%
cache read / (fresh input + cache write + cache read). This ratio reflects Claude Code's automatic caching, not bassclef specifically. Green above 75%; amber 50-75%; red below 50%.
Models used
Claude Opus 4.7 (1M context)
From chronicle frontmatter model field where present, else session-config default.
Attribution note. The cache-reuse ratio above reflects Claude Code's automatic caching of the system prompt, tool definitions, and rolling conversation history. Whether bassclef changes this ratio versus a vanilla harness — not measured yet. Raw token counts are withheld from this page until #240 provides the vanilla-baseline comparison that would let us attribute honestly.
In this window, 98% of input rode cache. What that means for bassclef specifically — not measured yet.
Luminaries dispatched 30d · top 10
Which named practitioners the framework consulted this month — Cooper for marketing surfaces, Beck for infrastructure, Norman for interaction. Which lens leads signals what kind of work shipped.
| Slug | Dispatches | Trend |
|---|---|---|
| alan-cooper | 22 | |
| linus-torvalds | 19 | |
| michael-feathers | 16 | |
| edward-tufte | 9 | |
| michael-nygard | 9 | |
| donald-norman | 8 | |
| karl-popper | 8 | |
| david-ogilvy | 7 | |
| sophia-prater | 6 | |
| don-norman | 5 |
Skills dispatched 30d · top 15
Which named workflows the framework fires. Dispatch counts show what earns its keep vs. what lives only in the README.
| Skill | Dispatches |
|---|---|
| /luminary | 25 |
| /launch | 23 |
| /longrun | 22 |
| /riff | 21 |
| /extract-intent | 18 |
| /temperance | 14 |
| /promote | 9 |
| /build | 8 |
| /ux-migration | 8 |
| /kiss | 7 |
| /visual-qa | 7 |
| /session-end | 6 |
| /sprint | 6 |
| /ticket | 6 |
| /onboard-repo | 5 |
Ranking counts commit-body mentions. Many mid-session skill dispatches never appear in a commit and don't get counted — the ranking is a floor, not a ceiling. See #357.
Negative space what the framework did not do
Numbers only mean something when we show what we didn't do too. Every count above says the discipline ran; the counts below say where it didn't. Both matter.
What the framework did NOT do — the counter-signal that separates honest data from vanity metrics.
28 of the last 61 chronicles missing the authoring_luminaries frontmatter field (46%). Discipline holds unevenly; the framework surfaces its own gaps.
12 PRs merged this month WITHOUT a temperance marker. Discipline is not perfect; honest signal beats vanity metric.
FAQ — bassclef vocabulary on this page
Terms specific to bassclef that appear in the metrics above. Plain-English glosses so no hover-tooltip friction breaks the reading flow of the numbers.
- /temperance
- Pause-and-think gate — "am I building the right thing?" before edits land.
- /kiss
- Rewrite prose in grade-10 plain English; strip bassclef-internal jargon.
- /luminary
- Consult a named practitioner's checklist — Beck for TDD, Norman for interaction, Cooper for goal-directed design.
- /diagnose
- Root-cause a bug before writing the fix — Is/Is Not + Five Whys + Hypothesis.
- /verify
- Post-build gate — check the change matches acceptance before committing.
- /pre-mortem
- Klein-style pre-mortem — assume the ship failed; write every reason why.
- /pattern-review
- Review code through world-class pattern lens — Gang of Four, Fowler, SOLID.
- /roadmap-reconcile
- Reconcile canvas + roadmap state against shipping reality — flag drift.
- /promote
- Propose a promotion — new skill, rule, or fix — up to bassclef substrate.
- /sprint
- Orient — show what's in flight and what's next this session.
- /whereami
- Point-in-time project state snapshot — phase, active work, gates.
- /decompose
- Pre-build responsibility audit — GRASP assignment + pattern selection.
- /architect-review
- Full architectural review — C4, ADR fitness, security audit.
- /riff-prototypes
- Build 2-3 clickable prototype variants driven by different luminary lenses.
- /session-log
- Write the narrative log for the session — what shipped, what's open.
- /build
- Take a buildable plan and ship it — dispatches builder per step + opens PRs.
- /launch
- Turn an idea into a buildable plan — spec + decomposition + migration.
- /longrun
- Run a long autonomous session that paces itself — prep + checkpoints + closeout.
- Skills
- Named markdown workflows the framework fires — /temperance (pause + think), /kiss (rewrite in plain English), and more.
- chronicle frontmatter
- Per-session narrative log with YAML frontmatter — dates, luminaries dispatched, PRs shipped, promotes filed.
- state markers
- Small marker files written by each gate — temperance, luminary, diagnose, pre-mortem, etc.
- Fresh input
- Net-new prompt tokens the API bills at the full input rate. Each session pays this on turn one, then rides the cache for turns after.
- Cache write
- Cache-creation tokens — the substrate loaded into the API cache on the first turn of each session. Higher unit cost than fresh input.
- Cache read
- Cache-hit tokens on subsequent turns. About 10% of the cache-write rate. This is where bassclef spends most of its input tokens.
- Output tokens
- Tokens the model generated as its response. Billed at the highest rate.
- Cache reuse ratio
- Fraction of total input that came from cache reads. Formula: cache_read / (fresh_input + cache_write + cache_read). High = cache-front-loaded discipline works.
Reach out
Question, bug report, or something bassclef should surface? Email [email protected]. It comes straight to Sunjay.
For source code + issues, the substrate repo at sunj-labs/bassclef is private today and opens publicly at v1.0. If you already have a GitHub account and want to file, the repo accepts issues from signed-in users. Otherwise, email works for everyone.