技能 2026-09-25 04:33 UTC
Senior-thinking-skills
由提交人 songjiun10-collab 提交,主要更新类型为:新增功能与能力扩展。提交说明:Add model-aware calibration table based on official Claude model docs (#2)。带来了新的能力与可配置项,扩展了使用场景。
Commit提交信息原文
Add model-aware calibration table based on official Claude model docs (#2)
* Add model-aware calibration table, sourced from Anthropic's model docs
Ties the router's track/format logic to the actual model tier running
it (context window, thinking mode) instead of treating every model as
interchangeable, and cross-references it from delegate-to-subagents
for picking a subagent worker's model.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
* Ground model calibration in Anthropic's model-specific prompting guidance
Adds a "Known per-model tendencies" table sourced from the official
model-specific-guidance section (over-engineering on Opus 4.5/4.6,
over-eager subagent spawning and over-verification on Opus 5, quiet
long-horizon loops on Fable 5.1) and ties each to the bundle skill
that already exists to counter it, so the router treats that guidance
as correcting a documented behavior rather than a hypothetical.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
* Add per-model calibration notes to all 29 skills
Every skill now carries a "Model notes" section covering Haiku 4.5,
Sonnet 5, Opus 5, Opus 5.5, and Fable 5.1, grounded in Anthropic's
model-specific prompting docs (context window, thinking mode, effort
defaults, and documented tendencies like Opus's over-verification and
over-delegation, or Fable 5.1's quieter long-running loops) rather
than treating every model as interchangeable.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
* Split all 29 skills into per-model files (Opus 5.5 / Fable 5.1 / Sonnet 5)
Each skill folder now carries three complete, standalone rewrites —
opus-5-5.md, fable-5-1.md, sonnet-5.md — instead of one shared body
with a bolted-on model-notes appendix. Each file re-tunes delivery
(effort defaults, over-verification, subagent delegation instinct,
verbosity, quiet long-horizon loops) for its specific model while
keeping the underlying discipline/framework intact. SKILL.md in each
folder is now a thin router: frontmatter unchanged (so skill-loading
triggers are unaffected), body replaced with a model -> file table.
Non-model-dependent infrastructure (hook wiring in
senior-engineer-mindset and delegate-to-subagents, the persistent-memory
CLI reference) stays in SKILL.md rather than being duplicated 3x.
Grounded in Anthropic's model-specific prompting docs (context window,
adaptive vs always-on thinking, effort defaults, and documented
tendencies like Opus's over-verification/over-delegation or Fable
5.1's quieter agentic loops).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
* Address Codex review: unlisted-model fallback, dangling refs, false claims
- Every SKILL.md router now names sonnet-5.md as the general-purpose
fallback for a model outside the three tuned tiers (Haiku, an older
Opus/Sonnet/Fable generation, or a non-Claude agent installing this
bundle), instead of leaving no guidance at all (P1).
- Replaced dangling references to senior-engineer-mindset's deleted
"Model-aware calibration" table (in delegate-to-subagents' three
files) with the actual tier-selection guidance inline.
- fresh-context-review/opus-5-5.md no longer claims self-review from
the same context is "no less thorough" than a fresh one — that
directly contradicted adversarial-review's own point about
same-context review confirming rather than disproving. Nontrivial
work now routes to a fresh-context subagent or session.
- bite-sized-plan/fable-5-1.md and plan-on-disk/fable-5-1.md no longer
thin the *persisted* plan down to goal/constraints only — that
undermines the exact resumability (after /clear, a crash, or
handoff) these skills exist for. Concrete steps stay in the written
plan; only live, in-session execution may deviate from them.
- All 22 opus-5-5.md files that said "set/raise effort explicitly"
now clarify that effort is a caller-configured request parameter,
not something adjustable mid-turn — the instruction means flagging
it explicitly and setting it for the next request, not silently
applying it to the current one.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
* Add GPT-6 Astra/Sol/Luna files to all 29 skills
Each skill folder now also carries gpt-6-astra.md, gpt-6-sol.md, and
gpt-6-luna.md alongside the existing Claude model files, and each
SKILL.md router table gains the three corresponding rows plus a
tightened fallback line (GPT-6 is now covered, not "something else").
Grounded in OpenAI's own migration guidance ("Rethinking skills and
prompts for GPT-6 Astra") and the Sol/Luna launch specs: Astra is more
tentative about declaring work done and asks more clarifying questions
than prior models (the opposite of Claude Opus's over-eagerness),
under-delegates to subagents, verifies its own work automatically,
has a documented "slop word" blocklist, and needs less prohibitive
guardrail language given stronger alignment. Sol is the cost-efficient
agentic workhorse tier with comparatively weak guardrail/warning-
circumvention resistance, so autonomous Sol verdicts near safety
boundaries are called out as worth an independent check rather than
trusted outright. Luna is scoped throughout to narrow, well-specified
legs, with judgment-heavy or structural work explicitly routed to a
stronger tier. Reasoning effort (none-max) does not reliably improve
results past a workload-dependent point on Sol/Luna, so effort is
framed as something to tune empirically rather than default to max.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
* Address second Codex review round: routers, effort, controller safeguards
- root-cause-discipline, search-first, record-the-why, and
security-review SKILL.md files were never trimmed to routers in the
original 3-model split — they kept the full ~130-line procedure AND
pointed to 6 model files containing the same content again, defeating
the split's point. Trimmed all four to the same compact-router format
used by the other 25 skills.
- Fixed 4 GPT-6 files (adversarial-review astra/sol, bite-sized-plan
sol, chestertons-fence sol) that repeated the same "escalate effort
mid-turn" error already fixed on the Opus files: effort is fixed by
the caller when a request starts, not adjustable from inside a
running turn. Each now points the escalation at the next request
instead.
- delegate-to-subagents/gpt-6-sol.md claimed a status file was the only
way to get progress visibility, ignoring the live agent-addressing
channel this same file describes elsewhere — now conditional on
whether a live channel exists in the environment.
- design-for-the-next-reader/gpt-6-astra.md's "bias toward inferring
intent over asking" was unscoped and would have overridden the
hard-to-reverse check-in gate for public APIs/data formats — scoped
to reversible ambiguity only.
- delegate-to-subagents/gpt-6-luna.md allowed Luna to act as controller
for a narrow fan-out but gave it none of the controller safeguards
(overlapping edits, permission scoping, baseline, independent
verification) — added the checklist for that allowed case.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F78URLSbRTjfwVYAtpPmRh
---------
Co-authored-by: Claude <noreply@anthropic.com>