Distilly: Turning a Person's Work and Voice into a Reusable Agent Skill

What Distilly Actually Solves

Distilly addresses a specific loss: when a colleague quits, a mentor graduates, or a teammate transfers, they take their unspoken playbook with them. The README frames the problem in emotional terms, but the technical core is concrete. It distills source material, such as messages, documents, interviews, and public sources, into a Person Profile that captures observable experience, decision patterns, expression, and ways of working. That profile is then packaged as an Agent Skill, which supported hosts can install and invoke. The project explicitly does not claim to clone the person behind the profile. This is a boundary worth noting: the output is a functional model of behavior, not a replica of consciousness. The intended users are teams losing institutional knowledge, individuals preserving relationship patterns, and fans or researchers wanting to ask a public figure's style questions. The README calls it a person-modeling layer for agents, which is a fair description of its position in the stack.

Three Person Families, One Workflow

The project divides its use cases into three families, each with its own collection strategy and profile structure. The colleague family covers coworkers, mentors, teammates, and up or downstream partners. It builds a Work Skill plus a Persona from material-derived technical standards, workflows, expression, and workplace behavior. It also supports collection from Lark, DingTalk, and Slack. The relationship family covers exes, partners, parents, friends, and close family. It organizes expression patterns, emotional triggers, conflict patterns, and repair patterns into a reusable Persona Skill. The celebrity family covers public figures, creators, and fictional characters. It ships with a six-dimension research toolchain described as subtitles to transcript cleanup, research merge, and quality check. The README names the dimensions but does not enumerate them, which is a gap. The key point is that one creator Skill, named distilly, builds all three families. That unification is the project's central design claim: a single mechanism, multiple person types.

How the Profile Becomes an Agent Skill

The data flow in the README is linear: source material plus your description goes into Distilly, which outputs a source-grounded Person Profile. That profile is then packaged as an Agent Skill. The current release packages each profile as a Skill so supported hosts can install and invoke it. The canonical creator Skill is named distilly, and the README instructs installing it in a distilly directory. This packaging choice is significant. It means the output is not a standalone program but a Skill definition, likely in the form of a SKILL.md file, that agent hosts read natively. The README lists eight hosts with native local Skill discovery: Claude Code, Hermes, OpenClaw, Codex, DeepSeek Harness, Pi, Grok Build, and OpenCode. Grok Bot is listed separately as a saved-Skill workflow preview, with a caveat that its official docs do not describe direct local SKILL.md imports. That caveat is a useful sign of honesty: the project distinguishes between native support and a preview workflow.

Getting It Running: What the README Actually Says

The README is sparse on concrete installation commands. It does not provide a pip install line, a git clone URL, or a configuration file example in the visible portion. The only explicit instruction is to install the creator Skill named distilly in a distilly directory. The repository's default branch is dot-skill, which suggests the project expects users to clone the repo and place the Skill directory somewhere their agent host scans. The release v0.01 is labeled demo, and the last push date is 2026-03-30. For a project at this stage, the absence of a step-by-step quickstart is a real limitation. A user familiar with Agent Skills might infer the pattern: clone the repo, copy the distilly Skill into the host's skills directory, and invoke it with source material. But the README does not confirm that flow. If you need exact commands, you will have to read the repository files directly, which the README does not link to in the visible text.

The Celebrity Toolchain: Promised Depth, Undisclosed Details

The celebrity family is the most ambitious part of the project. It claims a six-dimension research toolchain that processes subtitles, cleans transcripts, merges research, and runs quality checks. The goal is to organize observable decisions, expression, and mental models of public figures or fictional characters. This is a different challenge from the colleague family. A colleague's Slack history is direct evidence, but a celebrity's public output is curated and mediated. The toolchain's subtitle-to-transcript pipeline suggests a focus on video interviews, which is a sensible source for capturing a person's speaking style and reasoning. However, the README does not name the six dimensions, nor does it explain how the quality check works. This is a gap between promise and documentation. For an engineer evaluating the project, the celebrity path is the least verifiable. You cannot confirm what the toolchain does without inspecting the code, and the README gives no file paths to start from.

Limitations and Failure Modes

The most obvious limitation is maturity. The release is v0.01 demo, which means the project is at a proof-of-concept stage. The README's tone is promotional, with milestone announcements that include star counts and community gallery statistics, but those do not substitute for functional documentation. A second limitation is the source-grounded constraint. The README stresses that profiles are built from observable material, not from speculation. That is a strength for accuracy, but it also means the tool is useless without a substantial corpus. If you have only a few emails from a departing colleague, the resulting profile will be thin. The project does not claim to generate personality from nothing. A third failure mode is host compatibility. The README lists eight hosts, but it also notes that Grok Bot does not support direct local SKILL.md imports. That caveat implies that not every listed host works the same way. If your agent host is not in the list, you have no documented path. The project's renaming from Colleague Skill to Distilly also creates search and continuity friction, which the README acknowledges.

Alternatives and the Difference in Approach

The closest alternative is a custom prompt or system message that encodes a person's style manually. That approach is simpler and requires no tooling, but it is static and not source-grounded. Distilly's difference is that it builds the profile from material through a defined pipeline, which should produce more consistent and evidence-based output than hand-written prompts. Another alternative is a general-purpose agent memory system, such as a vector database that stores a person's messages and retrieves them at query time. That approach is retrieval-based: it pulls relevant snippets on demand. Distilly's approach is distillation-based: it compresses the material into a structured profile once, then reuses that profile. The trade-off is clear. Retrieval preserves fidelity but requires the source corpus to be available at runtime. Distillation is portable, because the profile is a self-contained Skill, but it loses nuance in the compression. For a colleague who has left and whose Slack history may be deleted, a distilled profile is more durable. For a living person whose style evolves, retrieval would stay current. The README does not compare itself to either approach, but the design choices point to distillation as the intended advantage.

Maintenance, License, and Upgrade Cost

The project is licensed under MIT, which is permissive and allows commercial use, modification, and redistribution with attribution. That is a low-friction license for adoption. The maintenance cost is harder to assess. The project has a roadmap, a technical report on arXiv, and a community gallery, which suggests active development. But the last push is from 2026-03-30, and the release is still v0.01. The README announces a 2026.08.13 milestone and a 2026.08.24 update, but those dates are after the last push, which is inconsistent with the repository metadata. That inconsistency is a red flag for an evaluator: the README may be ahead of the actual code. The upgrade cost is tied to the Skill format. If the project changes its packaging, for example, from a SKILL.md file to a different format, existing profiles may need re-exporting. The README does not describe a migration path. For a project that is renaming itself and expanding its scope, that is a risk. The community gallery, which lets any skill drive traffic to a GitHub repo, is a distribution feature, but it also means the project depends on community contributions for breadth.

Editorial conclusion

Adopt Distilly if you need to preserve a departing colleague's working style, a loved one's communication patterns, or a public figure's observable decision logic in a form your agent host can load. Skip it if you expect a finished product: the demo is v0.01, the README is heavy on vision and light on verified commands, and the celebrity toolchain's six dimensions are named but not detailed. Before committing, verify three things: that your agent host supports native local Skill discovery as claimed, that the source material you have matches the collection strategies for your chosen person family, and that the MIT license and the project's rapid renaming (Colleague Skill to Distilly) do not conflict with your own naming or distribution plans. The project is a promising sketch with a clear architecture, but its current release is a proof of concept, not a production tool.

Official sources

1. Official README
2. Project repository
3. Release notes