A 9-step pre-mortem skill that interrogates you before you write a single line of code.
Inspired by 滕雅辛 (Téng Yǎxīn / Yōusù), 《置身钉内》 (2026, ~75k characters) — a post-mortem of DingTalk's ONE project, a strategic-level AI office product that hit 3M DAU at launch and was abandoned within a year. Every one of the 9 interrogations in this skill corresponds 1:1 to a specific structural failure documented in that essay.
Original article (Chinese): https://www.163.com/dy/article/KUR5AI960556J2HN.html
This skill is not an "idea generator." It's the missing middle layer between brainstorming and execution — the moment when you're excited and your answers are still soft, and someone needs to push back.
When you say "I want to build X" / "let's make a tool that…", the skill walks you through one question at a time (not a dump), each with multiple-choice options grounded in a known failure mode:
| # | Interrogation | The ONE failure it prevents |
|---|---|---|
| 1 | Origin intent — pick ONE primary motive from 5 | ONE juggled 4 motives → "greed is one of the seven sins" |
| 2 | User targeting — who is it NOT for? (≥3 explicit exclusions required) | "We set off with a box of Schrödinger's users" |
| 3 | Scenario — minute-level specific, not "daily tool" | Card-bloat: from 10/day to 50+/day after every team plugged in |
| 4 | Value — quantify what users currently spend, then state your improvement multiplier | Morning dashboard: "just a re-skinned to-do list, novelty died in a week" |
| 5 | Moat — will it still be a moat in 2 years? | Stack Overflow's monthly questions halved within a year of ChatGPT |
| 6 | Bilateral game — who must yield when the two sides conflict? | "Boss pays, employee uses. Boss wants control, employee wants freedom. 100% mutually exclusive." |
| 7 | MVP boundary — can you cut it to one action? | "Couldn't walk yet, forced into a 100m sprint" |
| 8 | Deadline trap — is an external fixed deadline reverse-driving everything? | "August 25 launch event was immovable. Every spec, every change, made way for the keynote." |
| 9 | Stop-loss line — write the trigger NOW, with a specific number, date, and person | "Bad data: suppressed. User pushback: ignored. Retention drop: doubled down on ops. Backlash: PR cleanup." |
Outputs a frozen baseline.md you can slap yourself with later when the project drifts.
Most projects don't die in execution — they die at the moment of "OK let's start building", because nobody is forcing the founder to answer the questions that would have killed the project.
Existing skills cover the surrounding layers:
- Idea generation → diverging on what's possible
- Implementation plans → executing a fixed direction
- Spike → validating one technical assumption
But the "direction is fuzzy, decision is imminent" zone — where 80% of doomed projects pass through — has no skill guarding it. This is that skill.
This skill is runtime-agnostic. It contains no tool-specific calls. Tested with:
- ✅ Hermes Agent (hermes-agent.nousresearch.com)
- ✅ Claude Code (claude.com/product/claude-code)
- 🟡 Any agent that can read a markdown skill spec and conduct a conversational protocol
Place inside the project you're starting:
mkdir -p .claude/skills
git clone https://github.com/execbbbr/project-kickoff-interrogation \
.claude/skills/project-kickoff-interrogationThen in Claude Code:
> Use the project-kickoff-interrogation skill. I want to build [your idea].
Use across all your projects:
mkdir -p ~/.claude/skills
git clone https://github.com/execbbbr/project-kickoff-interrogation \
~/.claude/skills/project-kickoff-interrogationmkdir -p ~/.hermes/skills/product-thinking
git clone https://github.com/execbbbr/project-kickoff-interrogation \
~/.hermes/skills/product-thinking/project-kickoff-interrogationHermes will auto-discover the skill on the next session. Trigger:
> I want to build [your idea].
…and Hermes will load this skill automatically based on the description.
Just clone anywhere and point your agent at SKILL.md. The protocol is plain English (with Chinese choice labels) and works in any chat.
git clone https://github.com/execbbbr/project-kickoff-interrogation
# Then paste SKILL.md content into your agent's system prompt or rules file..
├── SKILL.md # The 9-question protocol (18k chars)
├── references/
│ └── one-case-study.md # Full 8-volume failure map with original quotes (15k chars)
├── templates/
│ └── baseline-template.md # Output structure for the frozen baseline (6k chars)
├── README.md # This file
└── LICENSE # MIT
The skill front-matter declares it as a Hermes-compatible skill, but the body is pure markdown protocol — any agent can execute it.
The skill instructs the running agent to:
- Ask one question at a time — do not dump all 9 at once.
- Give the user multiple-choice options with brief intros — never blank-slate prompts.
- Reject mushy answers ("both", "depends", "kind of") — push back until exactly one choice is given.
- Flag red flags out loud — if the user picks "MVP = build everything" or "we have a hard deadline", the agent must explicitly say "this is a ONE-pattern red flag — do you acknowledge?" before continuing.
- Force #9 to have a concrete number, date, and human name. "We'll all watch out for it" = nobody will pull the brake.
- Save
baseline.mdto disk when the 9 questions are done. No file = the exercise was theater.
Full protocol with all the choices, follow-ups, and pitfalls is in SKILL.md.
When the agent finishes, it produces a baseline.md like:
# my-tool — Frozen Baseline
⚠️ This is a Frozen Baseline.
Frozen: 2026-06-07 by Claude Code via project-kickoff-interrogation v1.0.0
## 1. Origin Intent
Primary: B. Efficiency improvement for an already-solved problem
Side benefits: C (serving a specific group — connector developers)
Distraction watch: none
## 2. User Targeting
Shape: A. Single-side (dev pays attention = dev benefits)
NOT for:
- non-technical PMs
- end customers of the connector
- executives looking at dashboards
…
## 9. Stop-Loss Line
Primary metric: weekly active connector devs
Threshold: 3 months post-launch (2026-09-07), if <40% of target team uses it weekly → pivot or stop
Brake-puller: [your name]Open-ended kickoff prompts produce open-ended kickoff answers: "users want to be productive." The 4–6 mutually exclusive options in each question force a real commitment. The ONE post-mortem shows this is the entire game — being precise about who you're NOT serving and what you'll NOT build is harder and more important than being precise about the positive.
The source material is Chinese, and many key concepts (发心, 已读未读, 事找人, 反身性) lose precision in translation. The interrogation protocol is bilingual: English narration with Chinese option labels where the source warrants it. Both Hermes and Claude Code handle this without issue. If you want a fully-English fork, see "Contributing" below.
Issues and PRs welcome. Particularly useful contributions:
- Other case studies. This skill happens to be grounded in ONE because the post-mortem is unusually honest and structured. A
references/companion for another famous-failure post-mortem (e.g. Quibi, Google+, Windows Phone, Stadia) would slot in cleanly. - Fully-English version. A
SKILL-en.mdalongside the current one. - Agent-specific quirks. If you find a runtime (Cursor, Cline, Open Interpreter, etc.) that mishandles the protocol, file an issue with the transcript.
MIT. See LICENSE.
- 滕雅辛 (Téng Yǎxīn / Yōusù) — author of 《置身钉内》. Without her 75k-character honest record of a strategic-level project's slow-motion collapse, this skill would just be vibes. Every numbered question is a debt owed to a paragraph she wrote.
- DingTalk ONE project team (2025–2026) — for being the empirical data.
- Built and packaged in Hermes Agent, tested across Hermes + Claude Code.