From 1d11862942df915dce243bb3a5c6a0b661197adb Mon Sep 17 00:00:00 2001 From: Fahmi Harun <34875577+kuker24@users.noreply.github.com> Date: Mon, 14 Sep 2026 17:38:45 +0700 Subject: [PATCH] chore: execute catalog stocktake and drop legacy host name PR #18 installed the verdict protocol but never ran it. This pass runs it across all 62 items and records the result in docs/stocktake-1.8.3.md: one verdict per item, 57 KEEP, 2 COMPRESS, 3 UPDATE, 0 MERGE, 0 RETIRE. MERGE is zero on measured evidence, not sentiment. The highest lexical description overlap in the catalog is matt-code-review ~ interrogate at 0.359 Jaccard, below the 50% warn line already enforced in tests. Every near pair is a fenced sibling. RETIRE is zero because no no-skill baseline was run. A baseline cannot be produced from a session that already has the catalog loaded, so five candidates (research, prototype, grilling, wait-what, ask-matt) are recorded as blocked on the eval-harness A/B gate rather than retired on argument alone. GrokBuild is the predecessor product name and survived in five bodies while appearing nowhere in README or docs. Removed where it named the host: the ask-matt description, which loads on every session start, a chrome-devtools-axi heading, adhd, and one visual-studio line claiming GrokBuild owns the tools. Left intact where it names an image/video tool contract rather than the host; renaming that has a wider blast radius and is recorded as a follow-up. Two currency contradictions against rules/00-routing.md also fixed: chrome-devtools-axi offered a $HOME/.grok/bin fallback while the rule says not to depend on ~/.grok at runtime, and grill-with-docs offered /implement while the rule says no such user skill exists. Compressed two bodies whose length carried no extra judgment. grill-with-docs 120 -> 77: the entire body was duplicated behind an orphan overlay marker dating to the 1.0.0 bootstrap and referenced nowhere in the engine. The halves differed, so the surviving text is a reconciled superset, not a truncation. codebase-design 115 -> 97: two ASCII diagrams restated the adjacent sentence. to-tickets, adhd, emil-design-eng, and full-performance-audit were left alone because their length is numeric gates and sequencing judgment. chrome-devtools-axi also shrank 75 -> 66, but as a consequence of deduplicating its contract block. It is counted once, under UPDATE. README documents the optional stitch MCP that shipped in 1.8.2 but was never listed beside serena and exa. No new skills, no new MCP, no new dependencies. VERSION stays 1.8.2 and no 1.8.3 heading is opened; entries remain under Unreleased. --- CHANGELOG.md | 6 ++ README.md | 1 + docs/stocktake-1.8.3.md | 136 ++++++++++++++++++++++++++++ skills/adhd/SKILL.md | 6 +- skills/ask-matt/SKILL.md | 2 +- skills/chrome-devtools-axi/SKILL.md | 11 +-- skills/codebase-design/SKILL.md | 24 +---- skills/grill-with-docs/SKILL.md | 45 +-------- skills/visual-studio/SKILL.md | 6 +- 9 files changed, 155 insertions(+), 82 deletions(-) create mode 100644 docs/stocktake-1.8.3.md diff --git a/CHANGELOG.md b/CHANGELOG.md index 2046fb8..48181eb 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -10,6 +10,12 @@ Catalog hygiene patch. No new skills, no new MCP, no foreign harness runtime. To - `tests/test_skills.py`: reconcile `vendor/skill-policy.json` against the actual `skills/`, `manual-skills/`, and `commands/` trees instead of trusting declared counts alone. Add frontmatter conformance (`name`, `description`, `compatibility: opencode`, `license`), model/manual name collision detection, a lexical description-overlap gate (warn at 50%, fail at 75% Jaccard), and a `~/.claude` runtime path assertion. Network-free and key-free. - `docs/warehouse-inventory.md`: `DEFER` no longer promotes to `NEW` without an existence pass and a no-skill baseline. Capability drift is grounds for `COMPRESS`/`RETIRE`, never a twin skill. Wave 1–3 counts are unchanged. - Document NVIDIA SkillEvaluator as `FOREIGN_ON_DEMAND`: never installed, absent is not a doctor failure, and a malformed MCP entry still fails closed. +- Execute the stocktake protocol across all 62 catalog items and record it in `docs/stocktake-1.8.3.md`: one verdict per item, 57 `KEEP`, 2 `COMPRESS`, 3 `UPDATE`, 0 `MERGE`, 0 `RETIRE`. `MERGE` is zero on measured evidence, the highest description overlap being 0.359 Jaccard, below the 50% warn line. No item is retired without a no-skill baseline; five candidates are recorded as blocked on the A/B gate. +- `README.md`: document optional `stitch` MCP alongside `serena` and `exa`. Enabled only by `opencode-bf stitch enable`, keys referenced as `{env:STITCH_API_KEY}` or omitted with `--oauth`, not an owned core server and not a production UI implementer, absent is not a doctor failure. +- Retire the predecessor `GrokBuild` host-product name from four skill bodies (`ask-matt` description, `chrome-devtools-axi` heading, `adhd`, `visual-studio`). `ask-matt` matters most: its description loads on every session. Tool-family references that name an image/video contract rather than the host are left intact. +- `chrome-devtools-axi`: drop a `$HOME/.grok/bin` fallback that contradicted `rules/00-routing.md` ("Do not depend on `~/.grok` at runtime") and reconcile two duplicated invocation blocks. Port `9223`, the `HEADED`/`AUTO_CONNECT` prohibition, and the never-fall-back-to-Google-Chrome fence are unchanged. +- Compress two bodies whose length carried no extra judgment: `grill-with-docs` 120 → 77 (the entire body was duplicated behind an orphan overlay marker dating to the `1.0.0` bootstrap; the halves differed, so the surviving text is a reconciled superset) and `codebase-design` 115 → 97 (two ASCII diagrams restating adjacent prose). Numeric gates, fences, and greppable terms survive verbatim. `to-tickets`, `adhd`, `emil-design-eng`, and `full-performance-audit` were left alone because their length is load-bearing. +- `grill-with-docs`: stop offering `/implement`, which `rules/00-routing.md` states does not exist as a user skill. ## 1.8.2 — 2026-09-10 diff --git a/README.md b/README.md index c98e149..5e7f62e 100644 --- a/README.md +++ b/README.md @@ -126,6 +126,7 @@ Optional: - `serena` — host binary may exist; MCP is **not** registered unless you run `opencode-bf serena enable` - `exa` — `FOREIGN_ON_DEMAND`; installer never adds, removes, or overwrites it +- `stitch` — `opencode-bf stitch enable` registers Google Stitch as a remote comp/mock source. Not an owned core server and not a production UI implementer: hand screens to `found-this-design` or `impeccable` before shipping. Keys are never written into config, only referenced as `{env:STITCH_API_KEY}`, or omitted with `--oauth`. `opencode-bf stitch disable` removes only that server key. Absent is not a `doctor` failure; a malformed entry fails closed. NVIDIA SkillEvaluator is `FOREIGN_ON_DEMAND` in the same sense: a maintainer may run it externally for embedding-based overlap scoring or live catalog evaluation. The installer never adds it, `doctor` does not fail when it is absent, and a malformed MCP entry for it fails closed like any other schema violation. diff --git a/docs/stocktake-1.8.3.md b/docs/stocktake-1.8.3.md new file mode 100644 index 0000000..0a595e9 --- /dev/null +++ b/docs/stocktake-1.8.3.md @@ -0,0 +1,136 @@ +# Catalog Stocktake — 1.8.3 + +Executed per [skills/skill-stocktake/SKILL.md](../skills/skill-stocktake/SKILL.md) against `main` at `4b1fd7f2` (tree `bc3d12c6`). + +Measured tree: **46 model-invoked** (`skills/*/SKILL.md`) + **16 manual** (`manual-skills/*/SKILL.md`, each with a matching `commands/.md`) = **62**. Matches `vendor/skill-policy.json` and `vendor/skill-allowlist.txt`. + +License column is authoritative from `vendor/license-audit.json`, not frontmatter. A missing frontmatter `license` key is not a gap: 28 of 62 deliberately defer to the audit file. + +`RETIRE` requires either a failed existence pass or a no-skill baseline. No baseline was run in this pass, so **no item is retired here**. Items whose retirement would depend on a baseline are listed under [Blocked on A/B](#blocked-on-ab) and carry `KEEP` in the table, per the evidence rule. + +## Verdicts + +| Skill | Kind | Verdict | Evidence | Handoff | +|---|---|---|---|---| +| academic | model | KEEP | 45 lines, 5 refs; owns scholarly/IMRaD lane fenced off `research` and `smartdoc` | — | +| adhd | model | KEEP | Host-brand leak fixed this pass; remaining bulk is calibration numerics + upstream MIT attribution | — | +| agent-architecture-audit | model | KEEP | 59 lines, NOTICE present; agent-loop lane distinct from `diagnosing-bugs` | — | +| api-design | model | KEEP | 37 lines; REST contract lane fenced off `contract-first` | — | +| ask-matt | model | UPDATE (applied) | Description said "GrokBuild skill or flow"; loaded every session | `writing-for-agents` | +| automation-audit-ops | model | KEEP | 44 lines; live-automation inventory, no sibling owns it | — | +| browser-act | model | KEEP | 48 lines; explicit-request-only fence vs `playwright-qa` | — | +| chrome-devtools-axi | model | UPDATE (applied) | Duplicate contract block + `$HOME/.grok` path contradicting `rules/00-routing.md:131` | `writing-for-agents` | +| click-path-audit | model | KEEP | 39 lines; handler/shared-store lane fenced off `playwright-qa` | — | +| code-tour | model | KEEP | 35 lines; produces `.tour` artifacts nothing else emits | — | +| codebase-design | model | COMPRESS (applied) | 115 → 97; two ASCII box diagrams restated adjacent prose | `writing-for-agents` | +| contract-first | model | KEEP | 41 lines; multi-consumer contract lane | — | +| cost-aware-llm-pipeline | model | KEEP | 48 lines; token/model-tier lane fenced off `full-performance-audit` | — | +| diagnosing-bugs | model | KEEP | 139 lines but carries the red-capable loop criterion; ships `scripts/` | — | +| diagram-design | model | KEEP | 46 lines, 3 refs; editorial diagram lane fenced off `impeccable` | — | +| domain-modeling | model | KEEP | 75 lines; glossary/ADR primitive composed by `grill-with-docs` | — | +| emil-design-eng | model | KEEP | 676 lines, largest in catalog; easing/timing numerics are the payload. COMPRESS off-limits by instruction | — | +| eval-harness | model | KEEP | 50 lines, 2 refs incl. `skill-utility.md` added in #18 | — | +| found-this-design | model | KEEP | 127 lines, 4 scripts; Design Bank entry point | — | +| full-audit-keamanan | model | KEEP | 69 lines; security fence, never COMPRESS | — | +| full-performance-audit | model | KEEP | 232 lines; LCP/INP/CLS thresholds are the payload. COMPRESS off-limits by instruction | — | +| gh-axi | model | KEEP | 65 lines; GitHub CLI surface | — | +| grill-with-docs | model | COMPRESS (applied) | 120 → 77; whole body duplicated behind orphan overlay marker | `writing-for-agents` | +| grilling | model | KEEP | 23 lines; non-default primitive, explicitly named-only | — | +| humanizer | model | KEEP | 64 lines, NOTICE present; paired with manual `/unslop` | — | +| hyperframes | model | KEEP | 48 lines, Apache-2.0 with NOTICE; HTML→MP4 lane | — | +| impeccable | model | KEEP | 89-line body against 38 refs + 44 scripts — progressive disclosure working as designed | — | +| install-anti-slop | model | KEEP | 56 lines; pinned upstream commit `e8c4880` verified in body | — | +| matt-code-review | model | KEEP | 88 lines; two-axis review, opt-in only | — | +| mongodb-ops | model | KEEP | 45 lines; vendor lane | — | +| playwright-qa | model | KEEP | 54 lines, Apache-2.0 with NOTICE; primary browser QA | — | +| prompt-optimizer | model | KEEP | 50 lines; advisory-only, does not mutate skills | — | +| prototype | model | KEEP | 27 lines; throwaway-evidence lane | — | +| research | model | KEEP | 15 lines, smallest body; already minimal | — | +| scroll-craft | model | KEEP | 101 lines, 8 refs; scroll-story lane fenced off `scroll-world` | — | +| scroll-world | model | KEEP | 129 lines; Hard rules section is fences, not ceremony | — | +| skill-stocktake | model | KEEP | 82 lines; this audit's own protocol | — | +| smartbook-ingest | model | KEEP | 33 lines; persistent-knowledge lane | — | +| smartdoc | model | KEEP | 62 lines, 4 refs; per-job document lane | — | +| supabase-ops | model | KEEP | 47 lines; vendor lane | — | +| tdd | model | KEEP | 39 lines; red-green lane | — | +| to-spec | model | KEEP | 80 lines; synthesis step before `to-tickets` | — | +| to-tickets | model | KEEP | 126 lines; extra length is expand–contract blast-radius judgment, not ceremony | — | +| vercel-ops | model | KEEP | 44 lines; vendor lane | — | +| visual-studio | model | UPDATE (applied) | Body claimed "GrokBuild already owns the tools" — host-product claim | `writing-for-agents` | +| writing-for-agents | model | KEEP | 82 lines; authoring lane, receives COMPRESS handoffs | — | +| architect | manual | KEEP | 82 lines, 3 refs; slash-only DAG bake-off | — | +| arena | manual | KEEP | 74 lines; multi-model bake-off, slash-only | — | +| blast-radius | manual | KEEP | 46 lines; impact analysis, slash-only | — | +| create-verification-skill | manual | KEEP | 44 lines; paired with `maintain-verification-skill` | — | +| decision-log | manual | KEEP | 66 lines, ships a script | — | +| figure-it-out | manual | KEEP | 53 lines; slash-only exploration | — | +| improve-codebase-architecture | manual | KEEP | 71 lines; slash-only | — | +| interrogate | manual | KEEP | 73 lines, 4 refs; highest overlap pair at 0.359, still under warn | — | +| maintain-verification-skill | manual | KEEP | 39 lines; verification profile upkeep | — | +| matt-implement | manual | KEEP | 15 lines; ticket-loop entry, never auto-started | — | +| reflect | manual | KEEP | 61 lines, 4 refs | — | +| technical-writing | manual | KEEP | 130 lines, largest manual; prose structure lane | — | +| unslop | manual | KEEP | 19 lines; manual twin of `humanizer` by design | — | +| wait-what | manual | KEEP | 9 lines, smallest in catalog | — | +| why | manual | KEEP | 79 lines, 4 refs; repo rationale, fenced off `research` | — | +| wizard | manual | KEEP | 47 lines; slash-only | — | + +## Tallies + +| Verdict | Count | +|---|---| +| KEEP | 57 | +| COMPRESS (applied) | 2 | +| UPDATE (applied) | 3 | +| MERGE | 0 | +| RETIRE | 0 | +| **Total** | **62** | + +`MERGE` is zero on evidence, not on sentiment. The highest lexical description overlap in the catalog is `matt-code-review ~ interrogate` at 0.359 Jaccard, below the 0.50 warn line in `tests/test_skills.py`. Top pairs: `humanizer ~ unslop` 0.357 (intentional model/manual twin), `supabase-ops ~ vercel-ops` 0.349, `grill-with-docs ~ grilling` 0.341 (documented composition), `scroll-craft ~ scroll-world` 0.327, `to-spec ~ to-tickets` 0.326. Each is a fenced sibling, not a duplicate. + +## Applied this pass + +Two COMPRESS, not five. The cap allowed five from a pool of six; two were skipped because their extra lines are load-bearing rather than ceremony, and the remaining pool members were already minimal. + +A third file shrank this pass, but under a different verdict. `chrome-devtools-axi` carries `UPDATE`: the line reduction was a consequence of removing a duplicated contract block, not a prose-trimming decision. It is counted once, as `UPDATE`. + +- **grill-with-docs** 120 → 77. The entire body appeared twice, split by an orphan `` marker at line 57, present since bootstrap `a62eeb6` and referenced nowhere in `lib/`, `tests/`, `docs/`, `vendor/`, `templates/`, `rules/`, or `install.sh`. The halves were not identical, so this is a reconciled superset: kept the frontier-batching clause and architecture-DAG fence from the first, plus `Sources`, the ADR shape block, and the `After` section from the second. Also fixed a contradiction where the closing line offered `/implement` while `rules/00-routing.md:20` states no such user skill exists. +- **codebase-design** 115 → 97. Two ASCII box diagrams restated the adjacent sentence. Deletion test, internal/external seam distinction, test-surface rule, and the two-adapter rule survive verbatim. +Under `UPDATE`, not counted as COMPRESS: + +- **chrome-devtools-axi** 75 → 66. Carried two near-identical invocation blocks, one headed `OpenCode`, one `GrokBuild`. Renaming the second made the duplication a literal repeated heading, so the blocks were reconciled into one. Port `9223`, the `HEADED`/`AUTO_CONNECT` prohibition, and the never-fall-back-to-Google-Chrome fence are unchanged. + +Skipped from the allowed pool: **to-tickets** (126 lines, but the extra prose is the expand–contract blast-radius sequencing judgment) and **adhd** (216 lines, but the remainder is calibration numerics plus upstream MIT attribution to `UditAkhourii/adhd`). **emil-design-eng** and **full-performance-audit** were off-limits by instruction, and both hold numeric gates that confirm the call. + +## Currency hits + +`GrokBuild` is the predecessor product name. It survived in five bodies while appearing nowhere in `README.md` or `docs/`. Treated as UPDATE evidence per instruction, not a blind rename: + +- `ask-matt` frontmatter description — fixed. This string loads at every session start, so a stale host name is paid on every turn. +- `chrome-devtools-axi` section heading — fixed, along with a `$HOME/.grok/bin` fallback that contradicted `rules/00-routing.md:131` ("Do not depend on `~/.grok` at runtime"). +- `adhd` — fixed. Dropped `Claude Code` and `GrokBuild` as if either were this runtime; the batch-vs-interactive distinction is preserved. +- `visual-studio` — fixed one line only. "GrokBuild already owns the tools" is a host-product claim. The `GrokBuild image_gen` tool-family references in the description and Hard rules were left alone; they name a tool contract, not the host. +- `scroll-world` — **no edit**. All three mentions name the image/video tool family. Follow-up note only. + +Remaining after this pass: 2 files (`scroll-world`, `visual-studio`) with 5 tool-family references total, plus a repo-relative `.grok/scroll-world//` scratch path in `scroll-world:57` that is not `~/.grok` and so is outside the `00-routing.md` prohibition. Renaming the tool contract is a separate decision with a wider blast radius. + +## Blocked on A/B + +No no-skill baseline was run, so these carry `KEEP` rather than a retirement verdict. Each is a candidate whose value claim is plausible but unmeasured, and each needs the `eval-harness` protocol in `references/skill-utility.md` before any retirement is defensible. + +| Candidate | Why it needs a baseline | Missing evidence | +|---|---|---| +| `research` | 15 lines that mostly say "prefer primary sources, use Context7". A current model may do this unprompted. | Run A on a library-facts question without the skill; compare citation discipline. | +| `prototype` | 27 lines describing throwaway evidence gathering, a behavior models default to. | Run A on a single design question; check whether scratch isolation degrades. | +| `grilling` | Non-default primitive already composed by `grill-with-docs`; standalone traffic is unknown. | Whether direct invocation ever beats the composing skill. | +| `wait-what` | 9 lines, smallest in catalog. Genuine trigger, but may be covered by ordinary clarification. | Whether the slash command changes behavior at all. | +| `ask-matt` | Router-selection helper whose job partly overlaps `AGENTS.md` routing itself. Table verdict is `UPDATE` (applied this pass); the utility question is separate and still open. | Whether routing accuracy drops without it. | + +A baseline cannot be produced from inside a single session that already has the catalog loaded. Each run needs a fresh session with the skill absent, which is maintainer work. + +## Follow-ups + +1. Run the A/B gate on the five candidates above before proposing any RETIRE. +2. Decide whether the `GrokBuild image_gen` tool-family name should be renamed catalog-wide, or documented as the intended tool contract. +3. `impeccable` frontmatter declares `Apache 2.0` while `vendor/license-audit.json` records `Apache-2.0`; the test normalizes this, but aligning the string would remove the special case. +4. 28 of 62 skills omit a frontmatter `license` key by design. If that ever becomes confusing, document it in the skill authoring guide rather than adding keys. diff --git a/skills/adhd/SKILL.md b/skills/adhd/SKILL.md index d4830ea..af80495 100644 --- a/skills/adhd/SKILL.md +++ b/skills/adhd/SKILL.md @@ -197,15 +197,15 @@ points where the cost of the obvious answer is high. ## Companion library and CLI There is a Node/TS implementation that does the same loop with structured -JSON parsing, score weighting, and a CLI. Use it when running outside -Claude Code or in batch. +JSON parsing, score weighting, and a CLI. Use it for batch runs outside an +interactive session. npm install -g adhd-agent adhd "your problem here" Code, paper, evals, and contributing guide at https://github.com/UditAkhourii/adhd. The skill above gives you the same -loop inside GrokBuild with no extra install required. +loop in-session with no extra install required. ## Source spec diff --git a/skills/ask-matt/SKILL.md b/skills/ask-matt/SKILL.md index f77d0eb..cef184e 100644 --- a/skills/ask-matt/SKILL.md +++ b/skills/ask-matt/SKILL.md @@ -1,6 +1,6 @@ --- name: ask-matt -description: Pick the GrokBuild skill or flow that fits. Use when the user asks which skill to use, which workflow, "alur apa", "ask matt", or is choosing between planning, implement, review, and design paths. +description: Pick the OpenCodeBestFriend skill or flow that fits. Use when the user asks which skill to use, which workflow, "alur apa", "ask matt", or is choosing between planning, implement, review, and design paths. compatibility: opencode --- diff --git a/skills/chrome-devtools-axi/SKILL.md b/skills/chrome-devtools-axi/SKILL.md index 764a8bd..138b6f1 100644 --- a/skills/chrome-devtools-axi/SKILL.md +++ b/skills/chrome-devtools-axi/SKILL.md @@ -7,18 +7,9 @@ license: MIT ## OpenCode browser contract -Follow the browser engine rules. Invocation only: - -1. Start background Chromium: `opencode-chromium-cdp start` (or `$HOME/.local/bin/opencode-chromium-cdp start`). -2. Every command: `CHROME_DEVTOOLS_AXI_BROWSER_URL=http://127.0.0.1:9223 npx -y chrome-devtools-axi `. -3. Never set `CHROME_DEVTOOLS_AXI_HEADED=1` or `CHROME_DEVTOOLS_AXI_AUTO_CONNECT=1` unless the user asks to see a window. -4. If the helper prints `NOT_CONFIGURED`, stop. Do not fall back to Google Chrome. - -## GrokBuild browser contract - Follow the browser engine rules in `00-routing.md`. Invocation only: -1. Start background Chromium: `opencode-chromium-cdp start` (or `$HOME/.grok/bin/opencode-chromium-cdp start`). +1. Start background Chromium: `opencode-chromium-cdp start` (or `$HOME/.local/bin/opencode-chromium-cdp start`). 2. Every command: `CHROME_DEVTOOLS_AXI_BROWSER_URL=http://127.0.0.1:9223 npx -y chrome-devtools-axi `. 3. Never set `CHROME_DEVTOOLS_AXI_HEADED=1` or `CHROME_DEVTOOLS_AXI_AUTO_CONNECT=1` unless the user asks to see a window. 4. If the helper prints `NOT_CONFIGURED`, stop. Do not fall back to Google Chrome. diff --git a/skills/codebase-design/SKILL.md b/skills/codebase-design/SKILL.md index 0b77df8..3a264a6 100644 --- a/skills/codebase-design/SKILL.md +++ b/skills/codebase-design/SKILL.md @@ -30,27 +30,9 @@ Use these terms exactly — don't substitute "component," "service," "API," or " ## Deep vs shallow -**Deep module** = small interface + lots of implementation: - -``` -┌─────────────────────┐ -│ Small Interface │ ← Few methods, simple params -├─────────────────────┤ -│ │ -│ Deep Implementation│ ← Complex logic hidden -│ │ -└─────────────────────┘ -``` - -**Shallow module** = large interface + little implementation (avoid): - -``` -┌─────────────────────────────────┐ -│ Large Interface │ ← Many methods, complex params -├─────────────────────────────────┤ -│ Thin Implementation │ ← Just passes through -└─────────────────────────────────┘ -``` +**Deep module** = small interface + lots of implementation. Few methods, simple params, complex logic hidden behind them. + +**Shallow module** = large interface + little implementation. Many methods, complex params, and a body that mostly passes through. Avoid it. When designing an interface, ask: diff --git a/skills/grill-with-docs/SKILL.md b/skills/grill-with-docs/SKILL.md index ae8fe35..dc5e76c 100644 --- a/skills/grill-with-docs/SKILL.md +++ b/skills/grill-with-docs/SKILL.md @@ -52,49 +52,6 @@ Leave the repo with: ## Open questions -## Glossary -``` - - -# Grill with docs - -Run the interview in this session. Do not delegate to missing Matt interview primitives. - -## Goal - -Leave the repo with: - -- `CONTEXT.md` — problem, decisions, open questions, glossary -- ADRs under `docs/adr/` (or `adr/` if that already exists) for hard-to-reverse choices - -## Rules - -- You gather facts. The user makes decisions. -- One question at a time when the answer branches. Batch only factual checks. -- Use the project's words. When a term is overloaded, resolve it and write it into the glossary. -- Do not implement code in this skill. -- If Codebase Memory has no project for cwd, skip it and use repo files. -- Stop when you can implement or write `/to-spec` without inventing decisions. - -## Loop - -1. Read `CONTEXT.md`, existing ADRs, and enough of the repo to speak the domain. -2. State the frontier: what you believe, what is undecided, what would change the design. -3. Ask the next question that most reduces that frontier. -4. After each answered decision, update `CONTEXT.md`. If the decision is hard to reverse, write an ADR. -5. Repeat until the stop condition. - -## CONTEXT.md shape - -```markdown -# - -## Problem - -## Decisions - -## Open questions - ## Glossary ## Sources @@ -117,4 +74,4 @@ Date: ## After -Tell the user the paths you wrote. If the work is multi-session or they asked for tickets, offer `/to-spec` then `/to-tickets`. Otherwise they can implement in this session or type `/implement`. +Tell the user the paths you wrote. If the work is multi-session or they asked for tickets, offer `/to-spec` then `/to-tickets`. Otherwise implementation stays in this session; there is no user `/implement` skill. diff --git a/skills/visual-studio/SKILL.md b/skills/visual-studio/SKILL.md index 165b5a1..1100bce 100644 --- a/skills/visual-studio/SKILL.md +++ b/skills/visual-studio/SKILL.md @@ -7,9 +7,9 @@ license: MIT # Visual Studio -Production director for photoreal stills and short videos. GrokBuild already -owns the tools. This skill owns the production order, identity, modes, and -VFX method. +Production director for photoreal stills and short videos. The image and +video tools already own generation. This skill owns the production order, +identity, modes, and VFX method. ## Load first