Lance Martin

if your skill depends on dynamic content, you can embed !`command` in your SKILL.md to inject shell output directly into the prompt Claude Code runs it when the skill is invoked and swaps the placeholder inline, the model only sees the result!

more details here: #march-18-2026" target="_blank" rel="nofollow noopener">platform.claude.com/docs/en/releas…

a useful trick: Claude API now programmatically lists capabilities for every model (context window, thinking mode, context management support, etc). just ask Claude Code or use the API directly. platform.claude.com/docs/en/api/py…

i think we might have undersold 1M context tbh, the performance is so so good, I really just don't clear the context window much these days

@anvanvan yes! check out the blog. claude.com/blog/1m-contex…

@RLanceMartin Any data on a context rot curve for the Opus 1M?

i've been surprised to see how quickly Claude is getting better at long context retrieval + reasoning. also, we just announced flat pricing across the full context window to 1M tokens. it's a good time to explore long context when building with Claude. x.com/RLanceMartin/s…

see the Claude API features overview: platform.claude.com/docs/en/build-… skill here (will be updating regularly): github.com/anthropics/ski…

a project i've been excited about: make Claude Code better at using the Claude API. just ask Claude Code about Claude API features (e.g., prompt caching, adaptive thinking + effort, tools, etc).

@ajay_sreeram not quite, but this question is helpful b/c we should share some example traces when we talk about PTC (e.g., examples w web search).

@RLanceMartin But doesn’t Claude need to spend even more tokens to write code in same context? (writing code without knowing tool response is even worse) (Or is writing post processing code is handled as separate agent, even here Claude need to prepare enough context to pass to the agent)

@dav_taylo not quite, Claude will write code that calls the search tool and performs any post processing that it sees fit for the task. so it’s more general than any specific hard coded step (re-rank) etc. Claude decides what to do based on the query.

@RLanceMartin Is the web search dynamic filtering effectively just an addl reranking step, though?

Claude Code plugin: github.com/anthropics/cla… repo: github.com/anthropics/ski… details here: claude.com/blog/improving…

check out the updated skill-creator. i esp like built-in support for test generation (e.g., to measure + optimize tricky things like skill trigger rate). available in Claude Code as plugin, Claude.ai, + Cowork.

detail on how the latest web search tool is using code execution to improve performance: x.com/RLanceMartin/s…

claude opus 4.6 + its web search tool is now #1 on @arena's search arena. the web search tool is available in the API, and uses code execution to process intermediate search results (saves tokens / improves performance). docs: platform.claude.com/docs/en/agents… x.com/arena/status/2…

ya, this is an interesting and subtle point: sub-agents and PTC both isolate context from an orchestrator. but sub-agents allocate tasks to LLMs. PTC chains actions in code. sub-agents can use PTC. PTC is useful anytime you have actions that you can easily compose in code.

interesting distinction. haven't tried PTC yet - I've been solving the composition tax with sub-agent delegation instead. built scoring that decides complexity per task, spawns isolated agents in their own context window, only the final result comes back. avoids the round-trip problem differently but I'm curious how PTC handles cases where intermediate results need judgment before the next step. that's where my sub-agents earn their keep.

@Ville_AI two considerations -- 1/ with Claude API just add "cache_control": {"type": "ephemeral"} see example: x.com/RLanceMartin/s… 2/ still, design your prompt w/ caching in mind see tips: x.com/trq212/status/…

@RLanceMartin How do you actually use this?

@Agimon_AI @ashvanth_s1 @Ashwinizer yes, Claude Code uses prompt caching. see here: x.com/trq212/status/…

@ashvanth_s1 @Ashwinizer @RLanceMartin Built in caching keeps prompts simpler and more stable. The consistent prefix thing is the sneaky hard part. Most "manual optimization" falls apart right there.

@kchdmn @lateinteraction @a1zhang ya, this part is the same as the PTC findings: RLMs can write code in a REPL to call tools as functions. RLMs have some additional properties -- prompt itself lives in the REPL external to LLM and LLM can call itself that are not part of PTC specifically, but are interesting!

@RLanceMartin The programmatic tool calling in Claude sounds like RLM @lateinteraction @a1zhang